{"meta":{"generated":"2026-11-06T00:57:39Z","close":"2026-11-05T07:59:59Z","sources":[{"file":"bench/runs.jsonl","rows":null,"what":"job, time, problem, agent, kind, speedup, passed, partial, shapes, of, source_sha256, snapshot, wall_s","slice":"trajectory"},{"file":"bench/remote/box/runs.jsonl","rows":null,"what":"same","slice":"trajectory"},{"file":"bench/traces/traces.db","rows":null,"what":"gpu_runs(job, t, problem, kind, x, passed, partial, shapes, of, sha); board(t, problem, leader, leader_x, our_rank, our_scored_x)","slice":"trajectory"},{"file":"bench/board_history.jsonl","rows":null,"what":"time, problem, leader, leader_x, our_rank, entries, our_scored_x","slice":"trajectory"},{"file":"bench/FLOOR.md","rows":null,"what":"physical ceiling column","slice":"trajectory"},{"file":"bench/runs/<problem>/<job>.py","rows":null,"what":"judge-run source snapshots","slice":"trajectory"},{"file":"bench/search/lb/*.json, bench/outbox/*overall*.txt, bench/human/reports/*/*.json (boards), bench/manager/overall_rank.jsonl, bench/now.json","rows":null,"what":"overall rank snapshots","slice":"trajectory"},{"file":"bench/tools/success_model.py","rows":null,"what":"definitions only (record, valid, KINDS, 24 h bar, Poisson design)","slice":"trajectory"},{"file":"bench/remote/*/runs.jsonl","rows":null,"what":"GPU runs present only on a host mirror (1749 jobs added after de-duplication by job)","slice":"models"},{"file":"bench/pi/runs.jsonl","rows":null,"what":"pi lane runs: name, model, variant, usd (calibrated bill), minutes, turns, tools, state, started/ended","slice":"models"},{"file":"okc/*/*/desk.json","rows":null,"what":"desk employees: rounds[] with seat, model, start_pt, end_pt (Codex / Kimi Code / Claude seats)","slice":"models"},{"file":"bench/cma/runs.jsonl","rows":null,"what":"CMA sessions: lane, models, usd, input/output tokens, t_start/t_end (event=end)","slice":"models"},{"file":"bench/oai/runs.jsonl","rows":null,"what":"OAI sessions: lane, models, usd, input/output tokens, t_start/t_end (event=end)","slice":"models"},{"file":"bench/claude/state/*.json","rows":null,"what":"Claude Code seat lanes: name -> model, profile (latest state only)","slice":"models"},{"file":"bench/kimi/state/*.json","rows":null,"what":"Kimi Code lanes: name -> model (kimi-code/k3.1-pro)","slice":"models"},{"file":"bench/desks/seats.json","rows":null,"what":"Claude seat lane -> desk employee map (c5-stacker -> dec-stacker, ...)","slice":"models"},{"file":"bench/pi/writer_registry.jsonl","rows":null,"what":"writer registry: sha256, path, writer_brain (sonnet/opus/subscription), lane, time","slice":"models"},{"file":"okc/muse_decode_layer/strike/lanes/attempts.jsonl","rows":null,"what":"strike attempts: who, file, verdict, at","slice":"models"},{"file":"bench/chain/<problem>.jsonl + .queue.jsonl","rows":null,"what":"chain landings (ev=cand: judged_file, evidence.job, by) and submits (origin_file, source)","slice":"models"},{"file":"bench/manager/BRIEF_MODEL_GENERIC.md","rows":null,"what":"latest generic brief-model refit: AUC, raise/lower terms (copied)","slice":"models"},{"file":"bench/rl/dataset/brief_rank_log.jsonl","rows":null,"what":"live brief scores of WORK_QUEUE writer tasks (t, id, score, tercile)","slice":"models"},{"file":"AGENTS.md","rows":null,"what":"","slice":"science"},{"file":"bench/HUNCHES.md","rows":null,"what":"","slice":"science"},{"file":"bench/HUNCH_LEDGER.md","rows":null,"what":"","slice":"science"},{"file":"bench/INTERVIEWS.md","rows":null,"what":"","slice":"science"},{"file":"bench/MISTAKES.md","rows":null,"what":"","slice":"science"},{"file":"bench/PROTOCOL.md","rows":null,"what":"","slice":"science"},{"file":"bench/QUESTIONS.md","rows":null,"what":"","slice":"science"},{"file":"bench/SYSTEMS.md","rows":null,"what":"","slice":"science"},{"file":"bench/SYSTEMS_THINKING.md","rows":null,"what":"","slice":"science"},{"file":"bench/WORKED.md","rows":null,"what":"","slice":"science"},{"file":"bench/agents/FAMILY_LEARNINGS.md","rows":null,"what":"","slice":"science"},{"file":"bench/agents/SPEC_V3.md","rows":null,"what":"","slice":"science"},{"file":"bench/arena/dms.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/arena/feed.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/audit/gates/pdl_order_gate.py","rows":null,"what":"","slice":"science"},{"file":"bench/books/","rows":null,"what":"","slice":"science"},{"file":"bench/books/POSTMORTEM_1029.md","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_attention_decode.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_attention_decode.queue.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_bf16_proj.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_bf16_proj.queue.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_decode_layer.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_decode_layer.queue.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_moe_w4a16.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_moe_w4a16.queue.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_prefill_layer.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/chain/muse_prefill_layer.queue.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/groups/*/log.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/human/questions.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/hunch_log.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/hunches.json","rows":null,"what":"","slice":"science"},{"file":"bench/manager/AB_RESEARCH.md","rows":null,"what":"","slice":"science"},{"file":"bench/manager/COORDINATOR.md","rows":null,"what":"","slice":"science"},{"file":"bench/manager/NEXT_ACTIONS.md","rows":null,"what":"","slice":"science"},{"file":"bench/manager/SLEEP.md","rows":null,"what":"","slice":"science"},{"file":"bench/manager/WORK_QUEUE.json","rows":null,"what":"","slice":"science"},{"file":"bench/manager/paired.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/manager/sleep/dreams.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/manager/sleep/onpolicy/*.json","rows":null,"what":"","slice":"science"},{"file":"bench/mapreduce/reduce.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/modal/box.py","rows":null,"what":"","slice":"science"},{"file":"bench/modal/judge.py","rows":null,"what":"","slice":"science"},{"file":"bench/modal/reaped.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/modal/statecopy.py","rows":null,"what":"","slice":"science"},{"file":"bench/papers/INDEX.md","rows":null,"what":"","slice":"science"},{"file":"bench/papers/txt/","rows":null,"what":"","slice":"science"},{"file":"bench/reports/evolution/2026-10-26-01a17a50/confirm-decode.md","rows":null,"what":"","slice":"science"},{"file":"bench/research/PAPER_AUDIT.md","rows":null,"what":"","slice":"science"},{"file":"bench/research/blogs/","rows":null,"what":"","slice":"science"},{"file":"bench/research/deep/","rows":null,"what":"","slice":"science"},{"file":"bench/research/fork_flush_1101.md","rows":null,"what":"","slice":"science"},{"file":"bench/sci/bridge_*.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/sci/events.jsonl","rows":null,"what":"","slice":"science"},{"file":"bench/tools/alarm.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/basin.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/board.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/boxwatch.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/brain.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/chain.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/clone_inputs.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/codex_island.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/deep_research.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/exa.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/forum.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/gap.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/gpu_fill.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/groups.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/hunch_loop.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/internal_judge.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/isa_budget.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/islands.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/jexact_scoped.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/judge_cal.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/keeps.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/limit_check.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/mailbox.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/manager.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/mapreduce.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/okc_run.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/paired_confirm.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/paired_guard.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/papers.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/pi_agent.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/pi_island.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/provenance.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/push_sync.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/race.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/safe_pull.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/sci.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/sleep.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/spend_guard.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/strike_lanes.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/traces.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/turn_end.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/unblock.py","rows":null,"what":"","slice":"science"},{"file":"bench/tools/win_audit.py","rows":null,"what":"","slice":"science"},{"file":"okc/muse_decode_layer/strike/lanes/BASINS.md","rows":null,"what":"","slice":"science"},{"file":"bench/manager/prices.json","rows":null,"what":"USD per 1M tokens per model (repo's single price table; Anthropic list prices cached 2026-10-06, cacheWrite = 1.25x input (5 min)). Confirmed vs the claude-api reference 2026-11-05: Opus 5.5 $4/$20 (cache read $0.20), Sonnet 5.5 $2/$10 ($0.20), Haiku 5.5 $0.10/$0.50.","slice":"economics"},{"file":"bench/modal/costs.jsonl","rows":7751,"what":"internal GPU cost ledger: box/builder/daytona hours booked up front at list rates, refunded at 'down'; fn rows = serverless GPU calls; per-problem rows = early Modal clone/judge functions. Lines deduped like judge._spent_today (git union merges copied lines).","slice":"economics"},{"file":"git log -p -- bench/manager/COMPUTE.json (+ working tree)","rows":70,"what":"Modal billing report snapshots written by bench/tools/compute_report.py ('modal billing report' per workspace and app: today_usd = current UTC day, month_usd = calendar month). 'at' is Mac local time (PT).","slice":"economics"},{"file":"bench/reports/modal-billing-20261026-29-hourly.json","rows":null,"what":"provider bill, hourly, per app and resource, 2026-10-26 19:00 to 10-28 20:00 UTC (last hour may be partial); read by the 10-28 utilization audit (bench/reports/modal-utilization-audit-20261028/REPORT.md)","slice":"economics"},{"file":"bench/prime/pods.jsonl","rows":null,"what":"Prime Intellect SSH H200 pods (2 pods, 10-23): terminate rows carry hours and usd_est","slice":"economics"},{"file":"bench/prime/guard.log","rows":null,"what":"Prime wallet guard log: wallet $ every 10 min","slice":"economics"},{"file":"bench/pi/state/*.json","rows":27,"what":"pi run states not in runs.jsonl (runs that ended unrecorded)","slice":"economics"},{"file":"~/.claude*/projects/-Users-stefan-Programming-meta*, ~/.claude-profiles/*/projects/-Users-stefan-Programming-meta*","rows":null,"what":"Claude Code session transcripts (subscription seats on this Mac): per assistant message model + usage tokens. Cloud seats (tether / Claude Code cloud sessions) are not on this disk.","slice":"economics"},{"file":"~/.codex*/sessions/2026/10/21+ and 2026/11 (rollout-*.jsonl[.zst])","rows":null,"what":"Codex CLI sessions (ChatGPT-login subscription): token_usage_record events per response; kept when the session cwd is the meta repo. .zst files are read with the zstd CLI.","slice":"economics"},{"file":"bench/manager/USAGE.json","rows":null,"what":"usage_report.py: Claude main/second accounts per Pacific day (all projects): msgs, tokens, rate_limit errors; api_spend per day","slice":"economics"},{"file":"bench/chain/<problem>.jsonl","rows":null,"what":"chain heads: 'cand' events (by != chain-import) = a confirmed change landed on the problem's chain head; 'cand_void' removes one. Chain started 2026-10-30 22:37Z.","slice":"economics"},{"file":"bench/modal/boxuse.jsonl","rows":null,"what":"box sampler (boxwatch): per box every ~5 min (80 s on 10-27): GPU util %, queue length, memory. Mirror gaps exist (10-28 audit: Mac mirror ~5 h behind).","slice":"economics"},{"file":"git log --since=2026-10-21 (HEAD ancestry, author date)","rows":36127,"what":"commit counts; 'sync:', 'keep:', 'push_sync:' and 'Merge' subjects are automation (squash/union syncs of the shared checkout)","slice":"economics"},{"file":"AGENTS.md, bench/modal/GPU_BUDGET.json, bench/manager/alloc.json, bench/manager/anthropic_budget.json, bench/manager/bedrock_budget.json, bench/prime/pods.jsonl, bench/tools/llm_spend.py","rows":null,"what":"owner budget decisions; each quote located by exact substring (src = file:line)","slice":"economics"},{"file":"bench/manager/spend_pause.json","rows":null,"what":"lanes paused for spending GPU with zero results in 48 h","slice":"economics"},{"file":"bench/spend_guard.jsonl","rows":null,"what":"spend_guard.py sweeps: rule hits (burn-alarm = Modal > $60/h in the last hour, box-cap, orphan-builder, argon-budget, ...); most are dry-run reports","slice":"economics"},{"file":"**/LEARNINGS.md","rows":null,"what":"per-folder notebooks: what moved the number, what did not, what was derived","slice":"knowledge"},{"file":"bench/ASSUMPTIONS.md","rows":null,"what":"assumptions tested","slice":"knowledge"},{"file":"bench/KNOBS.md","rows":null,"what":"tuning knobs and ranges","slice":"knowledge"},{"file":"bench/PRIMITIVES.md","rows":null,"what":"reusable kernel building blocks","slice":"knowledge"},{"file":"bench/hw/sources/","rows":null,"what":"hardware documents read","slice":"knowledge"}],"files":[{"file":"data/site.json","bytes":687420,"by":"tools/build_site.py","what":"Everything the pages read, reshaped from the slices below."},{"file":"data/raw/trajectory.json","bytes":578759,"by":"tools/build_trajectory.py","what":"Board rows, scored and internal best-so-far, GPU runs per hour, the scaling fits."},{"file":"data/raw/models.json","bytes":443280,"by":"tools/build_models.py","what":"Lane runs, agent-hours, cost and new bests per model, harness and role; writer attribution; brief model."},{"file":"data/raw/strategies.json","bytes":445139,"by":"tools/build_strategies.py","what":"Every scored-row move and chain landing by kind of change; milestones; mistake classes; candidate answers."},{"file":"data/raw/science.json","bytes":484011,"by":"tools/build_science.py","what":"The idea ledger: funnel, verdicts, calibration, hit rate by source, kill rule, harness components."},{"file":"data/raw/economics.json","bytes":488352,"by":"tools/build_economics.py","what":"Spend per day by provider, unit costs, utilization, waste, budget decisions."},{"file":"data/raw/bench.json","bytes":123783,"by":"tools/build_sample.py","what":"FloorBench families, tasks, gates, red-team cheats, dataset counts, the sample manifest."},{"file":"data/checks.json","bytes":5120,"by":"tools/check_site.py","what":"The latest verifier run."}],"definitions":[{"term":"Speedup, x","text":"SGLang's time divided by ours, per case, combined over the scored loads as the judge does. Higher is faster."},{"term":"Judge-scored row","text":"Our best scored run on the organizers' board. It is the best of several reads of noisy runs, so it typically sits 1% to 2% above the median of reruns of the same bytes (believed, from our mistake log)."},{"term":"Internal run (clone protocol)","text":"A run on our own H200 boxes against a copy of the judge. It decided what to keep. It reads about 2% to 4% above the judge on decode and MoE. The 24-hour record bar pools clone and judge runs; every other number keeps them apart."},{"term":"Physical floor and model ceiling","text":"A case's floor is the time the H200 needs at 4.8 TB/s of memory bandwidth or its power-capped tensor rate, whichever binds, plus launch cost. A problem's ceiling is the speedup if every case ran at its floor. It is a model value from bench/FLOOR.md, not a bound: one rival read above the prefill ceiling."},{"term":"Gap closed","text":"(x - 1) / (ceiling - 1): the share of the distance from SGLang parity to the ceiling."},{"term":"24-hour record (new best)","text":"A passed, all-shape run more than 0.2% faster than the best valid run of the previous 24 hours on that problem, as bench/tools/success_model.py defines it. The bar sits inside run noise, and late in the competition most records re-ran bytes that had already run; site/tools/build_records.py counts both."},{"term":"GPU run","text":"One job with a GPU measurement in the run ledgers: box, Modal, serverless or judge, deduplicated by job id. Lab runs on a subset of shapes are counted but never quoted as a problem speedup."},{"term":"Lane run and lane-hour","text":"A lane run is one agent session on one brief. Lane-hours are summed session time. Claude Code, Codex and Antigravity keep no session ledger, so their sessions are rebuilt from GPU activity (GPU span plus 15 minutes); pi lanes also log idle time. Neither unit compares harnesses exactly."},{"term":"Measured, believed, estimated","text":"Measured: recomputed from a raw ledger. Believed: copied from a note or inferred by a rule. Estimated: a modelled value, such as subscription use at API list prices. Each slice file tags its numbers this way."}],"caveats":["The overall rank over time is a proxy: our decode-board rank, which matched our overall rank in every snapshot that recorded both (all of them outside the #1 window). Board reads show decode #1 for 6.8 hours; there were no reads for the next 11 hours.","Writer credit for new bests is mostly inferred: the writer registry covers few files. Runner credit is measured.","Model and harness comparisons are associations. Roles, phases and briefs differ between lanes; no controlled model comparison ran during the competition.","Idea sources come from a keyword map over the idea text, so source labels are believed, not measured.","The organizers' final placing uses medians of reruns on held-out data. Every row here is provisional.","Agent-written text from the cold-start episodes stays off the public pages (bench/product/RISKS.md section 2: model output ships only to the provider that wrote it). The grader's numbers stay.","Builders count from slightly different universes: 24-hour records are 209 to 224 and GPU runs 58,452 to 60,466 depending on host mirrors and glitch filters. Each figure states its own source.","Fonts load from Google Fonts; everything else is served from this machine."]},"science":{"funnel":[{"stage":"ideas registered","n":14936},{"stage":"ideas with a timed GPU trial","n":7621},{"stage":"ideas with a full-problem hit label","n":340},{"stage":"ideas with a proven (controlled) full hit","n":236},{"stage":"paired confirmations run (candidate files, not ideas)","n":1612},{"stage":"paired CONFIRMED (CI low > 0)","n":130},{"stage":"chain landings by the merge queue (all problems)","n":135}],"funnel_ideas":[{"stage":"Ideas registered","n":14936},{"stage":"Timed on a GPU at least once","n":7621},{"stage":"Hit on some shapes","n":767},{"stage":"Hit on all shapes","n":340},{"stage":"Full hit with a same-box control","n":236},{"stage":"Reached a scored row or a chain landing","n":88}],"verdicts":{"hit":1074,"error":2309,"bad":987,"flat":2897,"miss":6052},"verdict_defs":{"hit":">= +1.0% over parent","flat":"0..+1.0% (labelled 'small gain' since 10-26)","miss":"<= 0%","bad":"<= -2.0%","error":"no timing (build fail, crash, harness bug)"},"trials_live":13319,"trials_voided":575,"partial_trials":6929,"full_trial_hit_rate":{"p":0.0868,"lo":0.0808,"hi":0.0932,"k":486,"n":5597},"daily":{"days":["2026-10-22","2026-10-23","2026-10-24","2026-10-25","2026-10-26","2026-10-27","2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-04","2026-11-05"],"series":{"ideas_added":[0,4279,1133,1498,209,2222,1869,1201,685,572,733,342,156,36,1],"trials":[0,2287,875,1147,291,2441,2011,967,640,701,987,751,188,32,1],"trial_hit":[0,392,42,90,8,80,96,117,51,33,112,20,32,1,0],"trial_flat":[0,438,67,251,43,565,376,262,127,243,248,197,69,11,0],"trial_miss":[0,938,508,490,90,1152,716,402,309,299,550,496,82,19,1],"trial_bad":[0,335,42,87,20,112,77,87,79,66,60,21,0,1,0],"trial_error":[0,184,216,229,130,532,746,99,74,60,17,17,5,0,0],"trials_voided":[0,216,23,36,12,135,83,8,11,13,18,14,1,5,0],"kill_events":[0,472,83,68,16,174,96,35,5,12,43,14,0,0,0],"auto_retired":[0,65,8,5,2,13,24,28,27,28,27,38,4,0,0],"revive_events":[0,2,3,2,0,1,0,0,0,1,0,0,0,0,0],"runs_box":[0,751,752,980,361,11633,15532,7668,10309,3416,5039,714,112,284,92],"runs_modal":[14,382,0,30,1,43,0,0,0,0,0,0,0,0,0],"runs_check":[29,96,25,15,0,74,33,25,34,5,7,1,10,2,0],"runs_scored":[5,47,18,5,0,19,13,11,10,7,9,5,6,20,6],"runs_fn":[0,0,0,0,0,0,0,0,0,0,0,28,430,96,0]}},"by_origin":[{"name":"Lane's own work (origin not stated)","ideas":3187,"tried":1535,"p":0.0658,"lo":0.0561,"hi":0.077,"k":101,"n":1535,"trial_rate":0.0599,"pred":0.004,"meas":0.0006},{"name":"Codex island","ideas":4124,"tried":890,"p":0.0157,"lo":0.0102,"hi":0.0242,"k":14,"n":890,"trial_rate":0.0927,"pred":0.0008,"meas":-0.0},{"name":"No source given","ideas":623,"tried":278,"p":0.1079,"lo":0.081,"hi":0.1424,"k":30,"n":278,"trial_rate":0.1076,"pred":0.0015,"meas":0.0004},{"name":"Measurement / oracle / profiling","ideas":592,"tried":181,"p":0.1105,"lo":0.0778,"hi":0.1547,"k":20,"n":181,"trial_rate":0.1386,"pred":0.003,"meas":0.0001},{"name":"Combination engine","ideas":3113,"tried":157,"p":0.0701,"lo":0.0433,"hi":0.1114,"k":11,"n":157,"trial_rate":0.1336,"pred":0.0025,"meas":0.0019},{"name":"First principles / floor","ideas":941,"tried":198,"p":0.1465,"lo":0.1099,"hi":0.1926,"k":29,"n":198,"trial_rate":0.1789,"pred":0.004,"meas":0.0004},{"name":"Knob sweep","ideas":152,"tried":123,"p":0.0163,"lo":0.0054,"hi":0.0479,"k":2,"n":123,"trial_rate":0.0127,"pred":0.001,"meas":-0.0006},{"name":"Port / amplify a win","ideas":185,"tried":86,"p":0.1047,"lo":0.0619,"hi":0.1715,"k":9,"n":86,"trial_rate":0.1168,"pred":0.0019,"meas":0.001},{"name":"Rival / peer code","ideas":242,"tried":76,"p":0.1316,"lo":0.0803,"hi":0.2082,"k":10,"n":76,"trial_rate":0.1683,"pred":0.002,"meas":0.0003},{"name":"Evolution search","ideas":152,"tried":49,"p":0.0408,"lo":0.0136,"hi":0.1161,"k":2,"n":49,"trial_rate":0.0625,"pred":0.0012,"meas":0.0004}],"conversion_by_origin":[{"origin":"papers (arXiv)","ideas":144,"with_trial":40,"with_measured_trial":37,"with_full_hit":4,"with_proven_full_hit":0,"working_optimizing_shipped_now":7},{"origin":"web / blogs / deep research","ideas":45,"with_trial":19,"with_measured_trial":19,"with_full_hit":1,"with_proven_full_hit":1,"working_optimizing_shipped_now":1},{"origin":"books","ideas":87,"with_trial":25,"with_measured_trial":18,"with_full_hit":1,"with_proven_full_hit":0,"working_optimizing_shipped_now":6},{"origin":"interviews / talks","ideas":18,"with_trial":3,"with_measured_trial":2,"with_full_hit":0,"with_proven_full_hit":0,"working_optimizing_shipped_now":0},{"origin":"nature / fractals","ideas":21,"with_trial":7,"with_measured_trial":6,"with_full_hit":3,"with_proven_full_hit":0,"working_optimizing_shipped_now":6},{"origin":"rival / peer code","ideas":242,"with_trial":143,"with_measured_trial":132,"with_full_hit":10,"with_proven_full_hit":10,"working_optimizing_shipped_now":10}],"calibration":{"first":{"n":3423,"median_predicted":0.0025,"median_measured":0.0001,"median_abs_error":0.0047,"median_measured_over_predicted":0.0285,"share_within_2x":{"k":708,"n":3131,"rate":0.2261,"ci90":[0.2141,0.2387]},"sign_agreement":{"k":1688,"n":3423,"rate":0.4931,"ci90":[0.4791,0.5072]},"share_measured_below_predicted":{"k":2262,"n":3131,"rate":0.7225,"ci90":[0.7091,0.7354]},"share_measured_positive":{"k":1756,"n":3423,"rate":0.513,"ci90":[0.4989,0.527]}},"prob":[{"p_bin":[0,0.2],"mean_p":0.109,"k":2,"n":277,"rate":0.0072,"ci90":[0.0024,0.0216]},{"p_bin":[0.2,0.4],"mean_p":0.2699,"k":21,"n":1030,"rate":0.0204,"ci90":[0.0143,0.029]},{"p_bin":[0.4,0.6],"mean_p":0.4755,"k":111,"n":1704,"rate":0.0651,"ci90":[0.056,0.0757]},{"p_bin":[0.6,0.8],"mean_p":0.6491,"k":137,"n":532,"rate":0.2575,"ci90":[0.2276,0.2899]},{"p_bin":[0.8,1.0],"mean_p":0.8472,"k":69,"n":220,"rate":0.3136,"ci90":[0.2647,0.3671]}],"scatter":[{"id":"pre-sweeper-194641-prank_last_all","problem":"pre","pred":0.0,"meas":0.0027,"origin":"codex island","status":"TRYING"},{"id":"dec-800400051","problem":"dec","pred":0.03,"meas":-0.0015,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-737","problem":"pre","pred":0.01,"meas":0.0041,"origin":"plan files / owner","status":"TRYING"},{"id":"bf16-141","problem":"bf16","pred":0.006,"meas":-0.0365,"origin":"codex island","status":"KILLED"},{"id":"dec-2131","problem":"dec","pred":0.0002,"meas":0.0004,"origin":"codex island","status":"TRYING"},{"id":"attn-414000079","problem":"attn","pred":0.008,"meas":-0.0062,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-637800042","problem":"bf16","pred":0.001,"meas":0.0079,"origin":"knob sweep","status":"TRYING"},{"id":"dec-2365","problem":"dec","pred":0.0003,"meas":-0.0006,"origin":"(empty)","status":"TRYING"},{"id":"dec-777300062","problem":"dec","pred":0.0005,"meas":-0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-548000025","problem":"attn","pred":0.01,"meas":-0.0108,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50562","problem":"attn","pred":0.0005,"meas":0.007,"origin":"codex island","status":"TRYING"},{"id":"dec-637800019","problem":"dec","pred":0.0005,"meas":-0.0028,"origin":"knob sweep","status":"TRYING"},{"id":"pre-sweeper-151021-guard_v4_long","problem":"pre","pred":0.0001,"meas":-0.0017,"origin":"codex island","status":"TRYING"},{"id":"dec-2795","problem":"dec","pred":0.001,"meas":-0.003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-587","problem":"bf16","pred":0.003,"meas":-0.0046,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-470500120","problem":"moe","pred":0.001,"meas":-0.0298,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-810","problem":"attn","pred":0.012,"meas":0.0068,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50209","problem":"moe","pred":0.002,"meas":0.0018,"origin":"combination engine","status":"TRYING"},{"id":"attn-829","problem":"attn","pred":0.006,"meas":-0.0032,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-888100003","problem":"dec","pred":0.008,"meas":0.0109,"origin":"(empty)","status":"WORKING"},{"id":"moe-985","problem":"moe","pred":0.005,"meas":0.0013,"origin":"combination engine","status":"TRYING"},{"id":"dec-717600003","problem":"dec","pred":0.001,"meas":-0.0006,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-673600034","problem":"dec","pred":0.001,"meas":-0.002,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-2065","problem":"dec","pred":0.004,"meas":0.0014,"origin":"codex island","status":"TRYING"},{"id":"dec-548000008","problem":"dec","pred":0.01,"meas":0.0927,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-640800016","problem":"dec","pred":0.0004,"meas":0.0001,"origin":"(empty)","status":"TRYING"},{"id":"dec-3085","problem":"dec","pred":0.0005,"meas":0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000014","problem":"dec","pred":0.03,"meas":-0.0026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-854","problem":"attn","pred":0.004,"meas":-0.0087,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50296","problem":"pre","pred":0.0003,"meas":-0.0042,"origin":"codex island","status":"TRYING"},{"id":"dec-1320","problem":"dec","pred":0.006,"meas":0.004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-429000045","problem":"dec","pred":0.002,"meas":-0.0026,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-314","problem":"moe","pred":0.0,"meas":-0.001,"origin":"codex island","status":"TRYING"},{"id":"dec-429000009","problem":"dec","pred":0.0024,"meas":-0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50360","problem":"pre","pred":0.0002,"meas":0.0029,"origin":"codex island","status":"TRYING"},{"id":"attn-414000059","problem":"attn","pred":0.0,"meas":-0.0494,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-509400010","problem":"attn","pred":0.0,"meas":-0.2952,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-326","problem":"pre","pred":0.005,"meas":0.0005,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-50337","problem":"bf16","pred":0.0001,"meas":0.0078,"origin":"knob sweep","status":"TRYING"},{"id":"bf16-637800007","problem":"bf16","pred":0.002,"meas":0.0014,"origin":"knob sweep","status":"TRYING"},{"id":"pre-50317","problem":"pre","pred":0.0001,"meas":-0.0072,"origin":"codex island","status":"TRYING"},{"id":"moe-414000063","problem":"moe","pred":0.002,"meas":-0.0011,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-15300013","problem":"moe","pred":0.01,"meas":0.0603,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1312","problem":"attn","pred":0.003,"meas":-0.0007,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-463200021","problem":"attn","pred":0.01,"meas":0.0113,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-637800020","problem":"dec","pred":0.0005,"meas":0.0031,"origin":"knob sweep","status":"TRYING"},{"id":"bf16-637800009","problem":"bf16","pred":0.002,"meas":0.0024,"origin":"knob sweep","status":"TRYING"},{"id":"attn-391600010","problem":"attn","pred":0.03,"meas":-0.0068,"origin":"knob sweep","status":"TRYING"},{"id":"dec-2150","problem":"dec","pred":0.0045,"meas":0.0051,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-273","problem":"moe","pred":0.004,"meas":-0.0016,"origin":"codex island","status":"KILLED"},{"id":"dec-888100039","problem":"dec","pred":0.03,"meas":-0.0011,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-861200007","problem":"attn","pred":0.0,"meas":0.5467,"origin":"(empty)","status":"WORKING"},{"id":"bf16-042","problem":"bf16","pred":0.015,"meas":-0.0943,"origin":"codex island","status":"TRYING"},{"id":"bf16-440","problem":"bf16","pred":0.008,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-679700039","problem":"dec","pred":0.0002,"meas":-0.0001,"origin":"(empty)","status":"TRYING"},{"id":"dec-2349","problem":"dec","pred":0.006,"meas":0.007,"origin":"(empty)","status":"TRYING"},{"id":"dec-2921","problem":"dec","pred":0.03,"meas":-0.0393,"origin":"first principles / floor","status":"KILLED"},{"id":"dec-779600003","problem":"dec","pred":0.0,"meas":0.0048,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-765700004","problem":"moe","pred":0.0,"meas":0.0021,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-765700009","problem":"attn","pred":0.03,"meas":-0.0067,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-001","problem":"bf16","pred":0.012,"meas":0.0035,"origin":"nature / fractals","status":"SHIPPED"},{"id":"moe-414000066","problem":"moe","pred":0.002,"meas":0.0002,"origin":"(empty)","status":"TRYING"},{"id":"dec-470500136","problem":"dec","pred":0.0022,"meas":-0.0041,"origin":"(empty)","status":"TRYING"},{"id":"pre-50368","problem":"pre","pred":0.0001,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"dec-2020","problem":"dec","pred":0.0015,"meas":0.0029,"origin":"codex island","status":"TRYING"},{"id":"dec-429000047","problem":"dec","pred":0.0003,"meas":0.0003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50337","problem":"pre","pred":0.001,"meas":-0.0014,"origin":"codex island","status":"TRYING"},{"id":"dec-517600053","problem":"dec","pred":0.0005,"meas":-0.0138,"origin":"port / amplify a win","status":"TRYING"},{"id":"moe-277","problem":"moe","pred":0.04,"meas":0.0144,"origin":"measurement / oracle / profiling","status":"WORKING"},{"id":"dec-1742","problem":"dec","pred":0.012,"meas":0.0018,"origin":"codex island","status":"TRYING"},{"id":"dec-2723","problem":"dec","pred":0.0025,"meas":0.004,"origin":"combination engine","status":"TRYING"},{"id":"attn-1347","problem":"attn","pred":0.05,"meas":-0.0044,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-50467","problem":"attn","pred":0.012,"meas":0.0072,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-673600061","problem":"dec","pred":0.0,"meas":0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1104","problem":"moe","pred":0.025,"meas":-0.0088,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-673600058","problem":"dec","pred":0.0,"meas":-0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-463200014","problem":"attn","pred":0.015,"meas":0.0183,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-496500074","problem":"dec","pred":0.0005,"meas":-0.0017,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-956","problem":"dec","pred":0.001,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-352","problem":"moe","pred":0.008,"meas":-0.0118,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-50348","problem":"pre","pred":0.0001,"meas":-0.0049,"origin":"codex island","status":"TRYING"},{"id":"dec-867700002","problem":"dec","pred":0.0,"meas":-0.0027,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-414000071","problem":"attn","pred":0.003,"meas":0.0101,"origin":"(empty)","status":"OPTIMIZING"},{"id":"dec-1470","problem":"dec","pred":0.001,"meas":-0.01,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50277","problem":"pre","pred":0.0005,"meas":-0.0003,"origin":"codex island","status":"TRYING"},{"id":"dec-673600006","problem":"dec","pred":0.004,"meas":-0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50558","problem":"attn","pred":0.001,"meas":0.0043,"origin":"codex island","status":"TRYING"},{"id":"dec-779600050","problem":"dec","pred":0.0005,"meas":-0.0015,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1525","problem":"dec","pred":0.015,"meas":0.0124,"origin":"combination engine","status":"OPTIMIZING"},{"id":"dec-867700012","problem":"dec","pred":0.03,"meas":0.0014,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50366","problem":"pre","pred":0.0001,"meas":-0.0128,"origin":"codex island","status":"TRYING"},{"id":"attn-463200009","problem":"attn","pred":0.03,"meas":0.0032,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1307","problem":"attn","pred":0.003,"meas":0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-548000026","problem":"attn","pred":0.05,"meas":0.0682,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000090","problem":"dec","pred":0.006,"meas":-0.0033,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700011","problem":"dec","pred":0.003,"meas":-0.0346,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1711","problem":"moe","pred":0.002,"meas":-0.0028,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-50549","problem":"attn","pred":0.001,"meas":-0.004,"origin":"codex island","status":"TRYING"},{"id":"dec-414000026","problem":"dec","pred":0.0006,"meas":-0.0004,"origin":"port / amplify a win","status":"TRYING"},{"id":"moe-840","problem":"moe","pred":0.008,"meas":0.0184,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-673600005","problem":"attn","pred":0.03,"meas":-0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-204900008","problem":"dec","pred":0.0015,"meas":-0.0023,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-194641-prank_last_long","problem":"pre","pred":0.0,"meas":0.0012,"origin":"codex island","status":"TRYING"},{"id":"pre-50320","problem":"pre","pred":0.0001,"meas":-0.0056,"origin":"codex island","status":"TRYING"},{"id":"all-116","problem":"all","pred":0.0,"meas":1.1062,"origin":"first principles / floor","status":"WORKING"},{"id":"dec-2004","problem":"dec","pred":0.0017,"meas":-0.0008,"origin":"codex island","status":"TRYING"},{"id":"attn-50554","problem":"attn","pred":0.001,"meas":-0.0078,"origin":"codex island","status":"TRYING"},{"id":"attn-50244","problem":"attn","pred":0.005,"meas":0.0033,"origin":"first principles / floor","status":"TRYING"},{"id":"pre-456","problem":"pre","pred":0.009,"meas":0.0047,"origin":"learned priors","status":"TRYING"},{"id":"dec-1735","problem":"dec","pred":0.01,"meas":-0.0017,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-132","problem":"dec","pred":0.005,"meas":-0.0006,"origin":"codex island","status":"KILLED"},{"id":"dec-414000121","problem":"dec","pred":0.0015,"meas":0.0002,"origin":"rival / peer code","status":"TRYING"},{"id":"attn-50000009","problem":"attn","pred":0.01,"meas":0.0988,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1639","problem":"attn","pred":0.003,"meas":-0.0058,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-637800018","problem":"dec","pred":0.0005,"meas":-0.0028,"origin":"knob sweep","status":"TRYING"},{"id":"moe-299","problem":"moe","pred":0.004,"meas":-0.0039,"origin":"codex island","status":"KILLED"},{"id":"dec-901","problem":"dec","pred":0.008,"meas":0.019,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-599200016","problem":"dec","pred":0.03,"meas":0.026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-800400050","problem":"dec","pred":0.03,"meas":-0.0014,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1052","problem":"moe","pred":0.002,"meas":-0.0059,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-1090","problem":"moe","pred":0.008,"meas":0.0128,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"pre-079","problem":"pre","pred":0.004,"meas":0.0004,"origin":"codex island","status":"KILLED"},{"id":"dec-370","problem":"dec","pred":0.006,"meas":-0.001,"origin":"codex island","status":"TRYING"},{"id":"dec-548000034","problem":"dec","pred":0.0,"meas":0.0,"origin":"(empty)","status":"TRYING"},{"id":"attn-463200010","problem":"attn","pred":0.005,"meas":0.0031,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000025","problem":"dec","pred":0.001,"meas":-0.0081,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-730100001","problem":"moe","pred":0.0003,"meas":-0.0171,"origin":"codex island","status":"TRYING"},{"id":"attn-548000023","problem":"attn","pred":0.01,"meas":-0.0276,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-60700003","problem":"dec","pred":0.0,"meas":-0.0005,"origin":"(empty)","status":"TRYING"},{"id":"pre-705","problem":"pre","pred":0.001,"meas":-0.004,"origin":"codex island","status":"TRYING"},{"id":"attn-673600001","problem":"attn","pred":0.03,"meas":0.0021,"origin":"(empty)","status":"TRYING"},{"id":"pre-651","problem":"pre","pred":0.007,"meas":0.0022,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-1381","problem":"moe","pred":0.003,"meas":0.0057,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600017","problem":"dec","pred":0.004,"meas":-0.0357,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2805","problem":"dec","pred":0.003,"meas":0.0018,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-714400013","problem":"bf16","pred":0.005,"meas":-0.0007,"origin":"combination engine","status":"TRYING"},{"id":"all-123","problem":"all","pred":0.0,"meas":0.4045,"origin":"first principles / floor","status":"TRYING"},{"id":"pre-704","problem":"pre","pred":0.001,"meas":-0.0002,"origin":"codex island","status":"TRYING"},{"id":"dec-2405","problem":"dec","pred":0.0015,"meas":0.0049,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2431","problem":"dec","pred":0.01,"meas":-0.0385,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-867700029","problem":"attn","pred":0.003,"meas":-0.0798,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-15300017","problem":"moe","pred":0.01,"meas":-0.0234,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-640800011","problem":"dec","pred":0.0,"meas":0.0015,"origin":"(empty)","status":"TRYING"},{"id":"dec-867700057","problem":"dec","pred":0.0,"meas":-0.0041,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-151021-guard_v4_mid","problem":"pre","pred":0.0001,"meas":-0.0028,"origin":"codex island","status":"TRYING"},{"id":"pre-182","problem":"pre","pred":0.01,"meas":0.0048,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-50334","problem":"pre","pred":0.001,"meas":-0.0079,"origin":"codex island","status":"TRYING"},{"id":"dec-888100057","problem":"dec","pred":0.001,"meas":0.1075,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1490","problem":"dec","pred":0.008,"meas":0.0046,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-637800044","problem":"bf16","pred":0.001,"meas":-0.0006,"origin":"knob sweep","status":"TRYING"},{"id":"dec-599200007","problem":"dec","pred":0.3,"meas":-0.0059,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-590000010","problem":"attn","pred":0.0,"meas":0.6294,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"bf16-50321","problem":"bf16","pred":0.0001,"meas":-0.0008,"origin":"codex island","status":"TRYING"},{"id":"pre-50097","problem":"pre","pred":0.002,"meas":-0.0004,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-610","problem":"bf16","pred":0.0005,"meas":-0.0022,"origin":"codex island","status":"TRYING"},{"id":"moe-053","problem":"moe","pred":0.02,"meas":-0.1212,"origin":"codex island","status":"KILLED"},{"id":"dec-429000052","problem":"dec","pred":0.002,"meas":-0.0225,"origin":"(empty)","status":"TRYING"},{"id":"attn-1709","problem":"attn","pred":0.012,"meas":0.0089,"origin":"combination engine","status":"TRYING"},{"id":"dec-1494","problem":"dec","pred":0.006,"meas":0.008,"origin":"plan files / owner","status":"TRYING"},{"id":"dec-867700038","problem":"dec","pred":0.002,"meas":1.1008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-627","problem":"pre","pred":0.0028,"meas":-0.0006,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-2322","problem":"dec","pred":0.002,"meas":-0.1522,"origin":"rival / peer code","status":"KILLED"},{"id":"dec-867700062","problem":"dec","pred":0.0,"meas":0.0108,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50089","problem":"bf16","pred":0.01,"meas":-0.0285,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-414000017","problem":"bf16","pred":0.005,"meas":-0.0364,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-714400020","problem":"bf16","pred":0.002,"meas":0.0043,"origin":"knob sweep","status":"TRYING"},{"id":"pre-096","problem":"pre","pred":0.015,"meas":0.0186,"origin":"codex island","status":"OPTIMIZING"},{"id":"dec-391600082","problem":"dec","pred":0.003,"meas":0.0016,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-777300059","problem":"dec","pred":0.0005,"meas":-0.0083,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-853","problem":"attn","pred":0.008,"meas":0.0048,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500037","problem":"dec","pred":0.03,"meas":0.0,"origin":"(empty)","status":"TRYING"},{"id":"bf16-589","problem":"bf16","pred":0.003,"meas":-0.0024,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-192","problem":"moe","pred":0.025,"meas":0.0307,"origin":"combination engine","status":"KILLED"},{"id":"dec-2766","problem":"dec","pred":0.0,"meas":0.0023,"origin":"(empty)","status":"TRYING"},{"id":"attn-414000125","problem":"attn","pred":0.005,"meas":-0.172,"origin":"dream / reflection / resurface","status":"TRYING"},{"id":"attn-1811","problem":"attn","pred":0.012,"meas":0.0253,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-548000012","problem":"attn","pred":0.005,"meas":0.1112,"origin":"lane's own work (origin not stated)","status":"OPTIMIZING"},{"id":"dec-888100014","problem":"dec","pred":0.001,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-315","problem":"moe","pred":0.0,"meas":-0.0002,"origin":"codex island","status":"KILLED"},{"id":"attn-090","problem":"attn","pred":0.04,"meas":0.0363,"origin":"codex island","status":"OPTIMIZING"},{"id":"pre-345","problem":"pre","pred":0.004,"meas":0.012,"origin":"codex island","status":"WORKING"},{"id":"moe-50000001","problem":"moe","pred":0.01,"meas":0.0068,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1323","problem":"moe","pred":0.015,"meas":-0.0102,"origin":"codex island","status":"TRYING"},{"id":"bf16-414000013","problem":"bf16","pred":0.05,"meas":-0.0426,"origin":"rival / peer code","status":"TRYING"},{"id":"all-117","problem":"all","pred":0.0,"meas":0.3604,"origin":"first principles / floor","status":"WORKING"},{"id":"moe-1037","problem":"moe","pred":0.004,"meas":0.0056,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-888100038","problem":"dec","pred":0.03,"meas":0.1572,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-646","problem":"attn","pred":0.005,"meas":0.0222,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-171855-routeweight_tw_l1last","problem":"pre","pred":0.0,"meas":0.0043,"origin":"codex island","status":"TRYING"},{"id":"pre-333","problem":"pre","pred":0.008,"meas":-0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-142","problem":"dec","pred":0.012,"meas":0.0211,"origin":"codex island","status":"WORKING"},{"id":"attn-414000096","problem":"attn","pred":0.006,"meas":-0.0128,"origin":"(empty)","status":"TRYING"},{"id":"pre-066","problem":"pre","pred":0.004,"meas":-0.001,"origin":"codex island","status":"TRYING"},{"id":"pre-579","problem":"pre","pred":0.007,"meas":-0.0059,"origin":"codex island","status":"TRYING"},{"id":"dec-1728","problem":"dec","pred":0.03,"meas":-0.0043,"origin":"codex island","status":"TRYING"},{"id":"dec-679700011","problem":"dec","pred":0.0,"meas":0.0233,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1520","problem":"dec","pred":0.005,"meas":-0.0538,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-637800034","problem":"bf16","pred":0.004,"meas":0.003,"origin":"knob sweep","status":"TRYING"},{"id":"pre-612","problem":"pre","pred":0.003,"meas":0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500044","problem":"dec","pred":0.03,"meas":-0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-209","problem":"bf16","pred":0.012,"meas":0.0108,"origin":"codex island","status":"WORKING"},{"id":"dec-391600016","problem":"dec","pred":0.003,"meas":-0.0257,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-078","problem":"attn","pred":0.03,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-123100002","problem":"dec","pred":0.004,"meas":0.0077,"origin":"rival / peer code","status":"TRYING"},{"id":"bf16-714400019","problem":"bf16","pred":0.002,"meas":-0.0007,"origin":"knob sweep","status":"TRYING"},{"id":"dec-717600006","problem":"dec","pred":0.0005,"meas":-0.0008,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-1366","problem":"dec","pred":0.006,"meas":-0.0069,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1551","problem":"moe","pred":0.0009,"meas":0.001,"origin":"codex island","status":"TRYING"},{"id":"moe-303","problem":"moe","pred":0.018,"meas":0.0155,"origin":"combination engine","status":"TRYING"},{"id":"pre-629","problem":"pre","pred":0.0045,"meas":0.0004,"origin":"combination engine","status":"TRYING"},{"id":"bf16-269","problem":"bf16","pred":0.001,"meas":0.0004,"origin":"codex island","status":"TRYING"},{"id":"dec-717600018","problem":"dec","pred":0.001,"meas":-0.0007,"origin":"(empty)","status":"TRYING"},{"id":"attn-1730","problem":"attn","pred":0.002,"meas":-0.002,"origin":"combination engine","status":"TRYING"},{"id":"dec-50205","problem":"dec","pred":0.001,"meas":0.0003,"origin":"knob sweep","status":"TRYING"},{"id":"moe-1669","problem":"moe","pred":0.001,"meas":0.0039,"origin":"evolution search","status":"TRYING"},{"id":"dec-1642","problem":"dec","pred":0.004,"meas":0.0025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600008","problem":"dec","pred":0.004,"meas":0.0041,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-496500015","problem":"bf16","pred":0.014,"meas":-0.0533,"origin":"(empty)","status":"TRYING"},{"id":"dec-414000053","problem":"dec","pred":0.011,"meas":0.0132,"origin":"rival / peer code","status":"WORKING"},{"id":"dec-148","problem":"dec","pred":0.012,"meas":0.0108,"origin":"codex island","status":"WORKING"},{"id":"attn-673600006","problem":"attn","pred":0.0,"meas":0.0037,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-677","problem":"attn","pred":0.012,"meas":0.0141,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-888100042","problem":"dec","pred":0.0,"meas":-0.018,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700044","problem":"dec","pred":0.003,"meas":-0.0023,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1805","problem":"dec","pred":0.01,"meas":-0.0107,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"dec-888100055","problem":"dec","pred":0.002,"meas":0.0115,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-151021-ranks_v4_post","problem":"pre","pred":0.0001,"meas":-0.0027,"origin":"codex island","status":"TRYING"},{"id":"dec-204900006","problem":"dec","pred":0.0005,"meas":-0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-414000045","problem":"attn","pred":0.003,"meas":-0.0009,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-744","problem":"bf16","pred":0.0015,"meas":0.0007,"origin":"rival / peer code","status":"TRYING"},{"id":"attn-50162","problem":"attn","pred":0.006,"meas":-0.0372,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700060","problem":"dec","pred":0.0,"meas":0.0341,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600028","problem":"dec","pred":0.03,"meas":-2.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1308","problem":"attn","pred":0.003,"meas":0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300046","problem":"dec","pred":0.001,"meas":-0.6407,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-248","problem":"moe","pred":0.012,"meas":-0.0035,"origin":"codex island","status":"KILLED"},{"id":"attn-599200021","problem":"attn","pred":0.03,"meas":0.0299,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50030","problem":"moe","pred":0.015,"meas":0.0025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50036","problem":"pre","pred":0.001,"meas":0.0008,"origin":"first principles / floor","status":"TRYING"},{"id":"bf16-50329","problem":"bf16","pred":0.001,"meas":-0.0075,"origin":"knob sweep","status":"TRYING"},{"id":"dec-714500038","problem":"dec","pred":0.01,"meas":0.0,"origin":"manager hypothesis generator","status":"TRYING"},{"id":"dec-50000057","problem":"dec","pred":0.0,"meas":0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-336","problem":"pre","pred":0.015,"meas":-0.6006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-673600007","problem":"attn","pred":0.01,"meas":-0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-470500088","problem":"dec","pred":0.008,"meas":0.0108,"origin":"evolution search","status":"TRYING"},{"id":"bf16-517600010","problem":"bf16","pred":0.003,"meas":-0.0044,"origin":"dream / reflection / resurface","status":"TRYING"},{"id":"bf16-50112","problem":"bf16","pred":0.012,"meas":-0.0158,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-046","problem":"attn","pred":0.02,"meas":-0.0514,"origin":"codex island","status":"TRYING"},{"id":"dec-2415","problem":"dec","pred":0.001,"meas":0.0043,"origin":"combination engine","status":"TRYING"},{"id":"dec-429000004","problem":"dec","pred":0.001,"meas":-0.0092,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-734800001","problem":"attn","pred":0.002,"meas":0.0101,"origin":"combination engine","status":"WORKING"},{"id":"attn-1878","problem":"attn","pred":0.0003,"meas":-0.0099,"origin":"combination engine","status":"TRYING"},{"id":"attn-15300001","problem":"attn","pred":0.002,"meas":-0.0025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50196","problem":"dec","pred":0.0001,"meas":-0.0017,"origin":"codex island","status":"TRYING"},{"id":"attn-463200027","problem":"attn","pred":0.0,"meas":-0.0017,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50395","problem":"pre","pred":0.0005,"meas":0.0004,"origin":"combination engine","status":"TRYING"},{"id":"dec-772200001","problem":"dec","pred":0.0003,"meas":0.0015,"origin":"codex island","status":"TRYING"},{"id":"attn-463300022","problem":"attn","pred":0.003,"meas":-0.021,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-619","problem":"bf16","pred":0.0025,"meas":0.0025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-548","problem":"bf16","pred":0.01,"meas":-0.2413,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1854","problem":"attn","pred":0.0,"meas":-0.031,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-734800084","problem":"dec","pred":0.3,"meas":0.0008,"origin":"(empty)","status":"TRYING"},{"id":"bf16-041","problem":"bf16","pred":0.012,"meas":-0.0254,"origin":"codex island","status":"KILLED"},{"id":"attn-827200004","problem":"attn","pred":0.0005,"meas":-0.0045,"origin":"codex island","status":"TRYING"},{"id":"dec-777300020","problem":"dec","pred":0.0016,"meas":-0.0035,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-340","problem":"moe","pred":0.01,"meas":0.0124,"origin":"codex island","status":"OPTIMIZING"},{"id":"pre-50383","problem":"pre","pred":0.0001,"meas":0.0021,"origin":"codex island","status":"TRYING"},{"id":"dec-800400028","problem":"dec","pred":0.002,"meas":-0.0001,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-171","problem":"dec","pred":0.012,"meas":0.0077,"origin":"codex island","status":"KILLED"},{"id":"dec-779600006","problem":"dec","pred":0.008,"meas":0.008,"origin":"(empty)","status":"TRYING"},{"id":"pre-50244","problem":"pre","pred":0.0,"meas":-0.0064,"origin":"codex island","status":"TRYING"},{"id":"bf16-512","problem":"bf16","pred":0.003,"meas":0.0018,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50491","problem":"attn","pred":0.016,"meas":-0.0004,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-50000023","problem":"dec","pred":0.0012,"meas":0.0018,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2424","problem":"dec","pred":0.001,"meas":-0.0014,"origin":"port / amplify a win","status":"TRYING"},{"id":"pre-sweeper-140533-plan_expert_last","problem":"pre","pred":0.0001,"meas":-0.0006,"origin":"codex island","status":"TRYING"},{"id":"moe-414000061","problem":"moe","pred":0.006,"meas":0.004,"origin":"dream / reflection / resurface","status":"TRYING"},{"id":"attn-50550","problem":"attn","pred":0.001,"meas":-0.0129,"origin":"codex island","status":"TRYING"},{"id":"dec-888100011","problem":"dec","pred":0.0001,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1615","problem":"moe","pred":0.0005,"meas":-0.0031,"origin":"evolution search","status":"TRYING"},{"id":"moe-1712","problem":"moe","pred":0.0,"meas":0.0374,"origin":"(empty)","status":"TRYING"},{"id":"moe-765700003","problem":"moe","pred":0.004,"meas":-0.0501,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-468500012","problem":"attn","pred":0.0,"meas":0.0027,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-776","problem":"attn","pred":0.01,"meas":0.009,"origin":"plan files / owner","status":"TRYING"},{"id":"dec-2962","problem":"dec","pred":0.0015,"meas":-0.416,"origin":"first principles / floor","status":"KILLED"},{"id":"dec-391600092","problem":"dec","pred":0.002,"meas":-0.0012,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-194641-prank_guardL1last_all","problem":"pre","pred":0.0,"meas":-0.0025,"origin":"codex island","status":"TRYING"},{"id":"dec-777300031","problem":"dec","pred":0.03,"meas":-0.0036,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2371","problem":"dec","pred":0.001,"meas":0.0081,"origin":"(empty)","status":"TRYING"},{"id":"moe-15300027","problem":"moe","pred":0.03,"meas":0.0605,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-371","problem":"dec","pred":0.004,"meas":0.0031,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-006","problem":"bf16","pred":0.015,"meas":-0.0073,"origin":"knob sweep","status":"TRYING"},{"id":"dec-777300006","problem":"dec","pred":0.03,"meas":0.0054,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1752","problem":"dec","pred":0.002,"meas":-0.006,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-429000021","problem":"dec","pred":0.0,"meas":0.0003,"origin":"port / amplify a win","status":"TRYING"},{"id":"pre-50375","problem":"pre","pred":0.0001,"meas":0.0048,"origin":"codex island","status":"TRYING"},{"id":"attn-50076","problem":"attn","pred":0.03,"meas":-0.01,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50475","problem":"attn","pred":0.004,"meas":0.0071,"origin":"first principles / floor","status":"TRYING"},{"id":"pre-636","problem":"pre","pred":0.0038,"meas":-0.0029,"origin":"learned priors","status":"TRYING"},{"id":"attn-599200013","problem":"attn","pred":0.002,"meas":0.0049,"origin":"(empty)","status":"TRYING"},{"id":"dec-470500094","problem":"dec","pred":0.0004,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-184136-pexp_cs_long","problem":"pre","pred":0.0,"meas":-0.0032,"origin":"codex island","status":"TRYING"},{"id":"dec-391600058","problem":"dec","pred":0.003,"meas":0.0421,"origin":"first principles / floor","status":"OPTIMIZING"},{"id":"attn-779600012","problem":"attn","pred":0.002,"meas":-0.0007,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2744","problem":"dec","pred":0.003,"meas":-0.0302,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-391600050","problem":"dec","pred":0.5,"meas":0.0178,"origin":"measurement / oracle / profiling","status":"OPTIMIZING"},{"id":"moe-140","problem":"moe","pred":0.01,"meas":0.1271,"origin":"codex island","status":"WORKING"},{"id":"dec-2359","problem":"dec","pred":0.001,"meas":-0.001,"origin":"(empty)","status":"TRYING"},{"id":"dec-779600043","problem":"dec","pred":0.0,"meas":-0.0129,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500069","problem":"dec","pred":0.001,"meas":-0.0011,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1745","problem":"attn","pred":0.015,"meas":-0.0011,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-517600013","problem":"dec","pred":0.0025,"meas":-0.0001,"origin":"combination engine","status":"TRYING"},{"id":"dec-123100021","problem":"dec","pred":0.1,"meas":0.0009,"origin":"combination engine","status":"TRYING"},{"id":"bf16-170","problem":"bf16","pred":0.3,"meas":-0.0009,"origin":"codex island","status":"KILLED"},{"id":"bf16-50073","problem":"bf16","pred":0.012,"meas":-0.0046,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-123100004","problem":"attn","pred":0.006,"meas":-0.0521,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-288","problem":"moe","pred":0.004,"meas":0.0102,"origin":"codex island","status":"KILLED"},{"id":"dec-2133","problem":"dec","pred":0.0045,"meas":0.0048,"origin":"codex island","status":"TRYING"},{"id":"dec-779600042","problem":"dec","pred":0.15,"meas":0.0017,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1175","problem":"moe","pred":0.015,"meas":-0.0033,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-470500088","problem":"moe","pred":0.0002,"meas":0.0026,"origin":"port / amplify a win","status":"TRYING"},{"id":"attn-714500013","problem":"attn","pred":0.0,"meas":0.0574,"origin":"(empty)","status":"TRYING"},{"id":"dec-2425","problem":"dec","pred":0.0025,"meas":0.0017,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700013","problem":"dec","pred":0.001,"meas":0.0026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000102","problem":"dec","pred":0.002,"meas":-0.0056,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-888100002","problem":"dec","pred":0.005,"meas":0.0087,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-001","problem":"pre","pred":0.04,"meas":0.0676,"origin":"alpha (go one level lower)","status":"SHIPPED"},{"id":"attn-114","problem":"attn","pred":0.004,"meas":-0.0013,"origin":"codex island","status":"TRYING"},{"id":"moe-470500114","problem":"moe","pred":0.001,"meas":-0.033,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-50341","problem":"pre","pred":0.0015,"meas":0.001,"origin":"codex island","status":"TRYING"},{"id":"bf16-50274","problem":"bf16","pred":0.0001,"meas":-0.0131,"origin":"codex island","status":"TRYING"},{"id":"pre-414000020","problem":"pre","pred":0.004,"meas":-0.0023,"origin":"port / amplify a win","status":"TRYING"},{"id":"pre-091","problem":"pre","pred":0.012,"meas":-0.0042,"origin":"codex island","status":"KILLED"},{"id":"bf16-714400006","problem":"bf16","pred":0.003,"meas":-0.0009,"origin":"knob sweep","status":"TRYING"},{"id":"attn-1216","problem":"attn","pred":0.004,"meas":-0.0024,"origin":"codex island","status":"TRYING"},{"id":"dec-673600003","problem":"dec","pred":0.005,"meas":-1.6829,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-589300002","problem":"moe","pred":0.05,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-625","problem":"dec","pred":0.024,"meas":0.0145,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-679700100","problem":"dec","pred":0.005,"meas":-0.1338,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1053","problem":"moe","pred":0.005,"meas":0.0084,"origin":"plan files / owner","status":"TRYING"},{"id":"bf16-046","problem":"bf16","pred":0.012,"meas":0.0613,"origin":"codex island","status":"TRYING"},{"id":"dec-2357","problem":"dec","pred":0.0005,"meas":0.0007,"origin":"(empty)","status":"TRYING"},{"id":"dec-599200015","problem":"dec","pred":0.03,"meas":0.0017,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2375","problem":"dec","pred":0.015,"meas":0.0,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-391600077","problem":"dec","pred":0.0,"meas":0.396,"origin":"measurement / oracle / profiling","status":"OPTIMIZING"},{"id":"bf16-011","problem":"bf16","pred":0.18,"meas":0.1783,"origin":"measurement / oracle / profiling","status":"SHIPPED"},{"id":"moe-1374","problem":"moe","pred":0.003,"meas":0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-045","problem":"bf16","pred":0.008,"meas":-0.0035,"origin":"interviews / talks","status":"TRYING"},{"id":"dec-679700008","problem":"dec","pred":0.002,"meas":-0.0067,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-003","problem":"moe","pred":0.01,"meas":-0.0037,"origin":"knob sweep","status":"KILLED"},{"id":"dec-640800002","problem":"dec","pred":0.001,"meas":-0.0004,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-2495","problem":"dec","pred":0.0015,"meas":-0.0388,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-414000120","problem":"dec","pred":0.001,"meas":-0.0022,"origin":"rival / peer code","status":"TRYING"},{"id":"moe-1015","problem":"moe","pred":0.004,"meas":-0.0116,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50249","problem":"attn","pred":0.007,"meas":0.0126,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-1507","problem":"dec","pred":0.005,"meas":0.0078,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-414000106","problem":"attn","pred":0.004,"meas":-0.0206,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-590000011","problem":"attn","pred":0.03,"meas":0.0576,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1800","problem":"dec","pred":0.002,"meas":-0.0119,"origin":"codex island","status":"TRYING"},{"id":"pre-50281","problem":"pre","pred":0.0005,"meas":-0.0063,"origin":"codex island","status":"TRYING"},{"id":"attn-1726","problem":"attn","pred":0.001,"meas":0.0061,"origin":"codex island","status":"TRYING"},{"id":"dec-468500026","problem":"dec","pred":0.035,"meas":-0.0472,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1193","problem":"attn","pred":0.006,"meas":-0.0042,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-1017","problem":"moe","pred":0.015,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000067","problem":"dec","pred":0.004,"meas":0.0061,"origin":"rival / peer code","status":"TRYING"},{"id":"moe-326300028","problem":"moe","pred":0.0002,"meas":0.0047,"origin":"codex island","status":"TRYING"},{"id":"dec-590000002","problem":"dec","pred":0.001,"meas":0.0143,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-355","problem":"moe","pred":0.012,"meas":0.0214,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"bf16-083","problem":"bf16","pred":0.012,"meas":0.0052,"origin":"codex island","status":"TRYING"},{"id":"attn-1720","problem":"attn","pred":0.001,"meas":-0.0116,"origin":"codex island","status":"TRYING"},{"id":"dec-470500098","problem":"dec","pred":0.0,"meas":0.0002,"origin":"evolution search","status":"TRYING"},{"id":"attn-726","problem":"attn","pred":0.012,"meas":-0.0005,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"dec-2102","problem":"dec","pred":0.0015,"meas":0.002,"origin":"codex island","status":"TRYING"},{"id":"bf16-503","problem":"bf16","pred":0.015,"meas":-0.0028,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-548000020","problem":"dec","pred":0.008,"meas":-0.0051,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-123100001","problem":"dec","pred":0.003,"meas":0.0026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600049","problem":"dec","pred":0.0008,"meas":-0.0516,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-717600016","problem":"attn","pred":0.008,"meas":0.0231,"origin":"combination engine","status":"WORKING"},{"id":"pre-sweeper-122444-large_m8","problem":"pre","pred":0.0001,"meas":-0.0032,"origin":"codex island","status":"TRYING"},{"id":"attn-725","problem":"attn","pred":0.005,"meas":-0.007,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"pre-50001","problem":"pre","pred":0.005,"meas":0.001,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-50204","problem":"moe","pred":0.0001,"meas":0.0,"origin":"codex island","status":"TRYING"},{"id":"bf16-062","problem":"bf16","pred":0.08,"meas":0.0,"origin":"port / amplify a win","status":"TRYING"},{"id":"bf16-150","problem":"bf16","pred":0.016,"meas":0.0148,"origin":"codex island","status":"KILLED"},{"id":"dec-153","problem":"dec","pred":0.012,"meas":0.0207,"origin":"codex island","status":"WORKING"},{"id":"dec-1831","problem":"dec","pred":0.003,"meas":-0.002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-429000003","problem":"dec","pred":0.001,"meas":0.0013,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-2138","problem":"dec","pred":0.002,"meas":0.002,"origin":"codex island","status":"TRYING"},{"id":"dec-2663","problem":"dec","pred":0.001,"meas":-0.0035,"origin":"first principles / floor","status":"KILLED"},{"id":"pre-50312","problem":"pre","pred":0.0001,"meas":-0.0037,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-164924-rank_l1first_mid","problem":"pre","pred":0.0,"meas":-0.002,"origin":"codex island","status":"TRYING"},{"id":"dec-800400060","problem":"dec","pred":0.03,"meas":-0.0023,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-156","problem":"dec","pred":0.012,"meas":0.0054,"origin":"codex island","status":"TRYING"},{"id":"dec-779600041","problem":"dec","pred":0.003,"meas":0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-414000103","problem":"attn","pred":0.002,"meas":-0.0049,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-463200017","problem":"dec","pred":0.005,"meas":0.0588,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-589300013","problem":"dec","pred":0.002,"meas":-0.0213,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600064","problem":"dec","pred":0.03,"meas":-0.0368,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-717600013","problem":"dec","pred":0.003,"meas":0.0004,"origin":"(empty)","status":"TRYING"},{"id":"dec-414000041","problem":"dec","pred":0.002,"meas":0.0032,"origin":"port / amplify a win","status":"TRYING"},{"id":"moe-15300012","problem":"moe","pred":0.01,"meas":-0.0212,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-637800006","problem":"dec","pred":0.0005,"meas":0.0049,"origin":"knob sweep","status":"TRYING"},{"id":"attn-1734","problem":"attn","pred":0.022,"meas":0.0165,"origin":"rival / peer code","status":"OPTIMIZING"},{"id":"attn-867700011","problem":"attn","pred":0.03,"meas":-0.0137,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"bf16-50125","problem":"bf16","pred":0.03,"meas":-0.5251,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1853","problem":"dec","pred":0.001,"meas":0.0005,"origin":"codex island","status":"TRYING"},{"id":"dec-2329","problem":"dec","pred":0.003,"meas":-0.0016,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-673600031","problem":"dec","pred":0.0,"meas":0.0069,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-765700007","problem":"attn","pred":0.002,"meas":0.0174,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-030","problem":"attn","pred":0.04,"meas":-0.3677,"origin":"codex island","status":"TRYING"},{"id":"moe-1082","problem":"moe","pred":0.008,"meas":-0.0077,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-859","problem":"attn","pred":0.003,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2274","problem":"dec","pred":0.0015,"meas":0.0006,"origin":"evolution search","status":"TRYING"},{"id":"pre-50370","problem":"pre","pred":0.0001,"meas":0.0143,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-091758-tanh64","problem":"pre","pred":0.0001,"meas":-0.0022,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-164924-rank_l1first_plan","problem":"pre","pred":0.0,"meas":-0.0019,"origin":"codex island","status":"TRYING"},{"id":"dec-679700013","problem":"dec","pred":0.002,"meas":0.002,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-963","problem":"moe","pred":0.012,"meas":-0.0067,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50323","problem":"bf16","pred":0.0001,"meas":-0.0047,"origin":"codex island","status":"TRYING"},{"id":"dec-779600004","problem":"dec","pred":0.03,"meas":0.1603,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"bf16-129","problem":"bf16","pred":0.012,"meas":0.0,"origin":"codex island","status":"KILLED"},{"id":"moe-170","problem":"moe","pred":0.015,"meas":0.0238,"origin":"codex island","status":"TRYING"},{"id":"moe-1601","problem":"moe","pred":0.0039,"meas":0.0041,"origin":"codex island","status":"TRYING"},{"id":"pre-50060","problem":"pre","pred":0.0045,"meas":0.0029,"origin":"manager hypothesis generator","status":"TRYING"},{"id":"dec-888100047","problem":"dec","pred":0.001,"meas":-0.0026,"origin":"(empty)","status":"TRYING"},{"id":"dec-50000053","problem":"dec","pred":0.004,"meas":0.0122,"origin":"(empty)","status":"TRYING"},{"id":"moe-50216","problem":"moe","pred":0.0003,"meas":-0.0013,"origin":"codex island","status":"TRYING"},{"id":"dec-2315","problem":"dec","pred":0.01,"meas":0.0074,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-772200018","problem":"dec","pred":0.0001,"meas":0.0003,"origin":"codex island","status":"TRYING"},{"id":"dec-679700140","problem":"dec","pred":0.0006,"meas":-0.0032,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-763","problem":"attn","pred":0.005,"meas":0.0061,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50233","problem":"moe","pred":0.0003,"meas":0.0018,"origin":"codex island","status":"TRYING"},{"id":"attn-724","problem":"attn","pred":0.006,"meas":0.003,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"moe-637800009","problem":"moe","pred":0.0002,"meas":-0.0052,"origin":"knob sweep","status":"KILLED"},{"id":"attn-50489","problem":"attn","pred":0.008,"meas":-0.006,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-208","problem":"moe","pred":0.014,"meas":-0.003,"origin":"codex island","status":"TRYING"},{"id":"pre-50289","problem":"pre","pred":0.001,"meas":-0.0011,"origin":"codex island","status":"TRYING"},{"id":"moe-1294","problem":"moe","pred":0.008,"meas":-0.0081,"origin":"combination engine","status":"TRYING"},{"id":"moe-867700003","problem":"moe","pred":0.0015,"meas":-0.0053,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-306","problem":"moe","pred":0.005,"meas":0.0002,"origin":"codex island","status":"KILLED"},{"id":"pre-50282","problem":"pre","pred":0.0005,"meas":-0.0036,"origin":"codex island","status":"TRYING"},{"id":"moe-1647","problem":"moe","pred":0.02,"meas":0.0232,"origin":"evolution search","status":"OPTIMIZING"},{"id":"bf16-50317","problem":"bf16","pred":0.0002,"meas":0.005,"origin":"codex island","status":"TRYING"},{"id":"moe-50235","problem":"moe","pred":0.0003,"meas":0.0053,"origin":"codex island","status":"TRYING"},{"id":"dec-2372","problem":"dec","pred":0.002,"meas":0.0006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-701","problem":"attn","pred":0.0027,"meas":-0.0018,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-658","problem":"attn","pred":0.012,"meas":0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1856","problem":"attn","pred":0.002,"meas":-0.0333,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-2421","problem":"dec","pred":0.0015,"meas":0.006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-637800014","problem":"dec","pred":0.0005,"meas":0.0011,"origin":"knob sweep","status":"TRYING"},{"id":"pre-50387","problem":"pre","pred":0.0001,"meas":0.0006,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-112510-score_gather","problem":"pre","pred":0.0001,"meas":-0.003,"origin":"codex island","status":"TRYING"},{"id":"dec-867700050","problem":"dec","pred":0.002,"meas":-0.0026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300066","problem":"dec","pred":0.03,"meas":-0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-765700016","problem":"dec","pred":0.0,"meas":0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000036","problem":"dec","pred":0.0,"meas":0.0,"origin":"(empty)","status":"KILLED"},{"id":"pre-496500056","problem":"pre","pred":0.001,"meas":-0.0335,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-414000013","problem":"moe","pred":0.004,"meas":0.0094,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-144","problem":"dec","pred":0.008,"meas":0.0017,"origin":"codex island","status":"TRYING"},{"id":"pre-50279","problem":"pre","pred":0.0005,"meas":-0.0021,"origin":"codex island","status":"TRYING"},{"id":"bf16-50320","problem":"bf16","pred":0.0001,"meas":-0.0134,"origin":"codex island","status":"TRYING"},{"id":"dec-468500018","problem":"dec","pred":0.03,"meas":0.1507,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1622","problem":"dec","pred":0.002,"meas":-0.0206,"origin":"papers (arXiv)","status":"TRYING"},{"id":"attn-50154","problem":"attn","pred":0.012,"meas":-0.0031,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-717600009","problem":"attn","pred":0.003,"meas":-0.0055,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2697","problem":"dec","pred":0.002,"meas":-0.0394,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-800400017","problem":"dec","pred":0.03,"meas":-0.002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-15300006","problem":"moe","pred":0.03,"meas":0.0606,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-432","problem":"attn","pred":0.005,"meas":-0.0086,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-278","problem":"dec","pred":0.01,"meas":0.0048,"origin":"codex island","status":"TRYING"},{"id":"pre-257","problem":"pre","pred":0.0115,"meas":0.0065,"origin":"port / amplify a win","status":"TRYING"},{"id":"attn-1295","problem":"attn","pred":0.012,"meas":-0.0146,"origin":"first principles / floor","status":"TRYING"},{"id":"pre-631","problem":"pre","pred":0.01,"meas":-0.0,"origin":"codex island","status":"TRYING"},{"id":"dec-777300058","problem":"dec","pred":0.001,"meas":0.0006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-60700047","problem":"dec","pred":0.001,"meas":-0.0233,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-013","problem":"dec","pred":0.06,"meas":0.0003,"origin":"plan files / owner","status":"SHIPPED"},{"id":"dec-1270","problem":"dec","pred":0.002,"meas":0.0111,"origin":"first principles / floor","status":"WORKING"},{"id":"bf16-734800006","problem":"bf16","pred":0.0008,"meas":0.0032,"origin":"combination engine","status":"TRYING"},{"id":"dec-2636","problem":"dec","pred":0.0005,"meas":-0.0008,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"attn-779600001","problem":"attn","pred":0.02,"meas":0.0304,"origin":"(empty)","status":"TRYING"},{"id":"pre-50258","problem":"pre","pred":0.002,"meas":0.0038,"origin":"codex island","status":"TRYING"},{"id":"bf16-637800015","problem":"bf16","pred":0.002,"meas":0.0077,"origin":"knob sweep","status":"TRYING"},{"id":"moe-414000041","problem":"moe","pred":0.0003,"meas":-0.0005,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-2800","problem":"dec","pred":0.0006,"meas":0.0029,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2939","problem":"dec","pred":0.004,"meas":-0.0032,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-779600007","problem":"dec","pred":0.03,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1333","problem":"attn","pred":0.003,"meas":0.0114,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1663","problem":"moe","pred":0.002,"meas":0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-496","problem":"moe","pred":0.01,"meas":0.0027,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1620","problem":"moe","pred":0.003,"meas":-0.0036,"origin":"evolution search","status":"TRYING"},{"id":"attn-1343","problem":"attn","pred":0.045,"meas":-0.0072,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-2153","problem":"dec","pred":0.0013,"meas":0.001,"origin":"codex island","status":"TRYING"},{"id":"attn-599200003","problem":"attn","pred":0.03,"meas":0.0294,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-679700080","problem":"dec","pred":0.0,"meas":-0.0022,"origin":"(empty)","status":"TRYING"},{"id":"moe-496500049","problem":"moe","pred":0.0004,"meas":-0.0066,"origin":"codex island","status":"TRYING"},{"id":"moe-1327","problem":"moe","pred":0.01,"meas":-0.0168,"origin":"codex island","status":"TRYING"},{"id":"dec-468500058","problem":"dec","pred":0.0,"meas":0.0003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-679700010","problem":"moe","pred":0.003,"meas":-0.0021,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-800400053","problem":"dec","pred":0.0,"meas":-0.0429,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600044","problem":"dec","pred":0.003,"meas":0.0127,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-124","problem":"moe","pred":0.005,"meas":0.0005,"origin":"codex island","status":"TRYING"},{"id":"dec-2466","problem":"dec","pred":0.002,"meas":0.0023,"origin":"learned priors","status":"TRYING"},{"id":"bf16-637800021","problem":"bf16","pred":0.0015,"meas":-0.0006,"origin":"knob sweep","status":"TRYING"},{"id":"attn-599200030","problem":"attn","pred":0.03,"meas":-0.0752,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1823","problem":"dec","pred":0.0003,"meas":0.0003,"origin":"codex island","status":"KILLED"},{"id":"dec-679700129","problem":"dec","pred":0.001,"meas":-0.0003,"origin":"(empty)","status":"KILLED"},{"id":"moe-496500114","problem":"moe","pred":0.001,"meas":-0.0277,"origin":"(empty)","status":"TRYING"},{"id":"dec-2346","problem":"dec","pred":0.0004,"meas":0.004,"origin":"(empty)","status":"TRYING"},{"id":"dec-679700125","problem":"dec","pred":0.001,"meas":-0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-608","problem":"pre","pred":0.005,"meas":-0.0007,"origin":"combination engine","status":"TRYING"},{"id":"moe-15300016","problem":"moe","pred":0.02,"meas":-0.0212,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-800400031","problem":"dec","pred":0.03,"meas":-0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-583","problem":"bf16","pred":0.004,"meas":0.0011,"origin":"codex island","status":"KILLED"},{"id":"attn-867700018","problem":"attn","pred":0.0,"meas":-0.0721,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-867700061","problem":"dec","pred":0.0,"meas":0.0195,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-681","problem":"attn","pred":0.0,"meas":-0.0017,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-1294","problem":"attn","pred":0.018,"meas":0.0264,"origin":"first principles / floor","status":"WORKING"},{"id":"pre-414000017","problem":"pre","pred":0.001,"meas":0.0046,"origin":"combination engine","status":"TRYING"},{"id":"pre-sweeper-184136-pexp_cs_short","problem":"pre","pred":0.0,"meas":-0.0009,"origin":"codex island","status":"TRYING"},{"id":"dec-2317","problem":"dec","pred":0.0015,"meas":0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700010","problem":"dec","pred":0.004,"meas":-0.0045,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-407","problem":"pre","pred":0.005,"meas":0.0042,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1022","problem":"moe","pred":0.005,"meas":0.009,"origin":"combination engine","status":"TRYING"},{"id":"dec-589300006","problem":"dec","pred":0.003,"meas":0.0034,"origin":"(empty)","status":"TRYING"},{"id":"moe-1685","problem":"moe","pred":0.0015,"meas":0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1573","problem":"dec","pred":0.0025,"meas":0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-771","problem":"pre","pred":0.0025,"meas":0.0011,"origin":"combination engine","status":"TRYING"},{"id":"moe-470500008","problem":"moe","pred":0.0,"meas":0.0007,"origin":"codex island","status":"TRYING"},{"id":"attn-599200018","problem":"attn","pred":0.002,"meas":-0.0161,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600018","problem":"dec","pred":0.002,"meas":-0.0053,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-742","problem":"pre","pred":0.0015,"meas":0.0012,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-548000002","problem":"attn","pred":0.0,"meas":0.007,"origin":"(empty)","status":"TRYING"},{"id":"moe-762300001","problem":"moe","pred":0.0,"meas":-0.0159,"origin":"(empty)","status":"TRYING"},{"id":"dec-679700007","problem":"dec","pred":0.002,"meas":-0.0047,"origin":"port / amplify a win","status":"TRYING"},{"id":"moe-50254","problem":"moe","pred":0.0003,"meas":-0.0011,"origin":"codex island","status":"TRYING"},{"id":"dec-2438","problem":"dec","pred":0.004,"meas":-0.0007,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1691","problem":"moe","pred":0.003,"meas":-0.0534,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-50000005","problem":"attn","pred":0.01,"meas":0.0184,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-765700010","problem":"dec","pred":0.003,"meas":-0.0009,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-576","problem":"bf16","pred":0.005,"meas":-0.0162,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-463200004","problem":"dec","pred":0.3,"meas":-0.0108,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-414000059","problem":"moe","pred":0.002,"meas":0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700059","problem":"dec","pred":0.0,"meas":0.056,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-032","problem":"attn","pred":0.04,"meas":0.1794,"origin":"codex island","status":"OPTIMIZING"},{"id":"pre-sweeper-105205-poll128","problem":"pre","pred":0.0002,"meas":-0.009,"origin":"codex island","status":"TRYING"},{"id":"pre-50327","problem":"pre","pred":0.0001,"meas":0.0091,"origin":"codex island","status":"TRYING"},{"id":"pre-50352","problem":"pre","pred":0.0005,"meas":0.0044,"origin":"codex island","status":"TRYING"},{"id":"moe-153","problem":"moe","pred":0.01,"meas":-0.0037,"origin":"codex island","status":"KILLED"},{"id":"dec-2458","problem":"dec","pred":0.0015,"meas":0.0015,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600055","problem":"dec","pred":0.002,"meas":0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2116","problem":"dec","pred":0.0004,"meas":-0.0002,"origin":"evolution search","status":"TRYING"},{"id":"attn-548000020","problem":"attn","pred":0.003,"meas":0.1071,"origin":"lane's own work (origin not stated)","status":"OPTIMIZING"},{"id":"dec-1624","problem":"dec","pred":0.0025,"meas":-0.0087,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"pre-sweeper-184136-prank_cs_short","problem":"pre","pred":0.0,"meas":0.0002,"origin":"codex island","status":"TRYING"},{"id":"dec-2794","problem":"dec","pred":0.0017,"meas":0.0038,"origin":"port / amplify a win","status":"TRYING"},{"id":"pre-50266","problem":"pre","pred":0.001,"meas":-0.0185,"origin":"codex island","status":"TRYING"},{"id":"dec-468500054","problem":"dec","pred":0.001,"meas":-0.0622,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-293","problem":"moe","pred":0.013,"meas":-0.1339,"origin":"codex island","status":"KILLED"},{"id":"dec-679700058","problem":"dec","pred":0.008,"meas":0.0086,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000085","problem":"dec","pred":0.002,"meas":0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300055","problem":"dec","pred":0.0003,"meas":-0.0015,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-679700006","problem":"moe","pred":0.1,"meas":-0.0049,"origin":"(empty)","status":"TRYING"},{"id":"pre-640","problem":"pre","pred":0.015,"meas":-0.0082,"origin":"codex island","status":"TRYING"},{"id":"dec-599200023","problem":"dec","pred":0.0,"meas":-0.0034,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-679700115","problem":"dec","pred":0.002,"meas":-0.0031,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50275","problem":"pre","pred":0.001,"meas":-0.002,"origin":"codex island","status":"TRYING"},{"id":"attn-049","problem":"attn","pred":0.02,"meas":0.3022,"origin":"codex island","status":"SHIPPED"},{"id":"dec-204900001","problem":"dec","pred":0.0015,"meas":0.0014,"origin":"first principles / floor","status":"TRYING"},{"id":"pre-50361","problem":"pre","pred":0.0002,"meas":-0.0012,"origin":"codex island","status":"TRYING"},{"id":"attn-50093","problem":"attn","pred":0.001,"meas":0.0057,"origin":"manager hypothesis generator","status":"TRYING"},{"id":"attn-827","problem":"attn","pred":0.007,"meas":-0.0045,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-15300007","problem":"dec","pred":0.03,"meas":-0.0112,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000044","problem":"dec","pred":0.0,"meas":0.6524,"origin":"rival / peer code","status":"WORKING"},{"id":"dec-2320","problem":"dec","pred":0.002,"meas":-0.0206,"origin":"evolution search","status":"TRYING"},{"id":"bf16-470500030","problem":"bf16","pred":0.0126,"meas":-0.0602,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-185950-prank_wt_long","problem":"pre","pred":0.0,"meas":0.0037,"origin":"codex island","status":"TRYING"},{"id":"attn-509400006","problem":"attn","pred":0.002,"meas":0.013,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-496500058","problem":"pre","pred":0.0438,"meas":-0.0043,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-104","problem":"attn","pred":0.06,"meas":0.0806,"origin":"codex island","status":"WORKING"},{"id":"dec-673600044","problem":"dec","pred":0.0,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-438","problem":"bf16","pred":0.01,"meas":0.0042,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-599200019","problem":"dec","pred":0.0,"meas":-0.0057,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-888100006","problem":"attn","pred":0.01,"meas":0.0244,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-468500010","problem":"dec","pred":0.0067,"meas":0.0078,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-334","problem":"dec","pred":0.01,"meas":0.0061,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000009","problem":"dec","pred":0.0005,"meas":-0.0012,"origin":"rival / peer code","status":"TRYING"},{"id":"moe-1363","problem":"moe","pred":0.03,"meas":0.0,"origin":"codex island","status":"TRYING"},{"id":"attn-50553","problem":"attn","pred":0.001,"meas":-0.012,"origin":"codex island","status":"TRYING"},{"id":"attn-463200003","problem":"attn","pred":0.005,"meas":0.0193,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50000008","problem":"dec","pred":0.03,"meas":-0.0064,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2376","problem":"dec","pred":0.015,"meas":0.0014,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-414000068","problem":"dec","pred":0.001,"meas":0.0012,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000055","problem":"dec","pred":0.03,"meas":-0.0014,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-800400029","problem":"dec","pred":0.002,"meas":-0.0017,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-599200007","problem":"attn","pred":0.03,"meas":-0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50197","problem":"pre","pred":0.03,"meas":-0.2383,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-067","problem":"bf16","pred":0.02,"meas":0.0042,"origin":"combination engine","status":"KILLED"},{"id":"dec-589300009","problem":"dec","pred":0.03,"meas":0.0557,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-589300012","problem":"dec","pred":0.002,"meas":-0.0166,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1192","problem":"attn","pred":0.008,"meas":-0.0434,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-800400045","problem":"dec","pred":0.002,"meas":0.0024,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-470500042","problem":"pre","pred":0.0438,"meas":-0.005,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-517600003","problem":"bf16","pred":0.004,"meas":0.0393,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2094","problem":"dec","pred":0.0017,"meas":0.0025,"origin":"codex island","status":"TRYING"},{"id":"dec-867700069","problem":"dec","pred":0.0,"meas":-0.0009,"origin":"(empty)","status":"TRYING"},{"id":"dec-277","problem":"dec","pred":0.03,"meas":0.0343,"origin":"codex island","status":"OPTIMIZING"},{"id":"dec-391600020","problem":"dec","pred":0.0,"meas":-0.0041,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1642","problem":"attn","pred":0.02,"meas":0.017,"origin":"codex island","status":"WORKING"},{"id":"pre-50283","problem":"pre","pred":0.0005,"meas":-0.0098,"origin":"codex island","status":"TRYING"},{"id":"pre-50397","problem":"pre","pred":0.001,"meas":0.0004,"origin":"combination engine","status":"TRYING"},{"id":"pre-sweeper-194641-pexp_last_all","problem":"pre","pred":0.0,"meas":0.0006,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-194641-pexp_last_mid","problem":"pre","pred":0.0,"meas":0.0296,"origin":"codex island","status":"WORKING"},{"id":"dec-391600066","problem":"dec","pred":0.001,"meas":0.0138,"origin":"(empty)","status":"WORKING"},{"id":"bf16-50066","problem":"bf16","pred":0.002,"meas":0.0036,"origin":"combination engine","status":"TRYING"},{"id":"dec-123100023","problem":"dec","pred":0.0008,"meas":0.001,"origin":"(empty)","status":"TRYING"},{"id":"dec-1533","problem":"dec","pred":0.0005,"meas":0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50190","problem":"moe","pred":0.0075,"meas":-0.0013,"origin":"first principles / floor","status":"TRYING"},{"id":"pre-50252","problem":"pre","pred":0.002,"meas":0.0004,"origin":"codex island","status":"TRYING"},{"id":"pre-50373","problem":"pre","pred":0.0001,"meas":0.0133,"origin":"codex island","status":"TRYING"},{"id":"moe-242","problem":"moe","pred":0.02,"meas":-0.1453,"origin":"codex island","status":"KILLED"},{"id":"moe-154","problem":"moe","pred":0.01,"meas":0.0037,"origin":"codex island","status":"KILLED"},{"id":"dec-2422","problem":"dec","pred":0.0009,"meas":0.0009,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-15300010","problem":"dec","pred":0.001,"meas":0.0052,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-523","problem":"bf16","pred":0.008,"meas":0.007,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-417900001","problem":"moe","pred":0.0,"meas":-0.0113,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-888100066","problem":"dec","pred":0.002,"meas":0.0323,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-373","problem":"dec","pred":0.002,"meas":-0.0063,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-777300047","problem":"dec","pred":0.0001,"meas":-0.0036,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1597","problem":"dec","pred":0.0015,"meas":0.0017,"origin":"combination engine","status":"TRYING"},{"id":"pre-sweeper-140533-plan_bias_last","problem":"pre","pred":0.0001,"meas":-0.006,"origin":"codex island","status":"TRYING"},{"id":"attn-734800002","problem":"attn","pred":0.0101,"meas":-0.0041,"origin":"combination engine","status":"OPTIMIZING"},{"id":"moe-152","problem":"moe","pred":0.012,"meas":-0.0121,"origin":"codex island","status":"KILLED"},{"id":"moe-414000065","problem":"moe","pred":0.003,"meas":-0.0015,"origin":"combination engine","status":"TRYING"},{"id":"dec-414000050","problem":"dec","pred":0.0,"meas":0.8553,"origin":"rival / peer code","status":"WORKING"},{"id":"pre-sweeper-194641-prank_L1last_all","problem":"pre","pred":0.0,"meas":0.0025,"origin":"codex island","status":"TRYING"},{"id":"moe-496500102","problem":"moe","pred":0.0025,"meas":0.0099,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-137","problem":"attn","pred":0.008,"meas":0.0035,"origin":"codex island","status":"KILLED"},{"id":"dec-779600003","problem":"dec","pred":0.003,"meas":-0.0026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2410","problem":"dec","pred":0.0015,"meas":0.0031,"origin":"port / amplify a win","status":"TRYING"},{"id":"moe-517600027","problem":"moe","pred":0.002,"meas":0.001,"origin":"combination engine","status":"TRYING"},{"id":"bf16-066","problem":"bf16","pred":0.02,"meas":0.0175,"origin":"codex island","status":"KILLED"},{"id":"bf16-117","problem":"bf16","pred":0.013,"meas":0.0143,"origin":"measurement / oracle / profiling","status":"WORKING"},{"id":"dec-50000007","problem":"dec","pred":0.03,"meas":-0.0043,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-717600001","problem":"dec","pred":0.002,"meas":-0.0034,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-360","problem":"pre","pred":0.004,"meas":0.0044,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-590000009","problem":"attn","pred":0.001,"meas":-0.1069,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50333","problem":"pre","pred":0.001,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"dec-1419","problem":"dec","pred":0.001,"meas":-0.0115,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700023","problem":"dec","pred":0.004,"meas":0.0044,"origin":"(empty)","status":"TRYING"},{"id":"dec-548000057","problem":"dec","pred":0.03,"meas":0.0068,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-122444-route_high","problem":"pre","pred":0.0001,"meas":-0.0046,"origin":"codex island","status":"TRYING"},{"id":"dec-1782","problem":"dec","pred":0.006,"meas":-0.1401,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-194641-pexp_last_short","problem":"pre","pred":0.0,"meas":0.0015,"origin":"codex island","status":"TRYING"},{"id":"dec-824200001","problem":"dec","pred":0.001,"meas":-0.0015,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1478","problem":"dec","pred":0.002,"meas":-0.0021,"origin":"combination engine","status":"TRYING"},{"id":"attn-184","problem":"attn","pred":0.01,"meas":-0.0003,"origin":"codex island","status":"TRYING"},{"id":"dec-777300050","problem":"dec","pred":0.0013,"meas":-0.5596,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300056","problem":"dec","pred":0.0003,"meas":-0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000125","problem":"dec","pred":0.001,"meas":0.0006,"origin":"dream / reflection / resurface","status":"TRYING"},{"id":"attn-589300003","problem":"attn","pred":0.0,"meas":-0.0103,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-091758-tanh32","problem":"pre","pred":0.0001,"meas":-0.0159,"origin":"codex island","status":"TRYING"},{"id":"moe-15300025","problem":"moe","pred":0.03,"meas":0.0606,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-463300007","problem":"dec","pred":0.0,"meas":-0.001,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-777300014","problem":"attn","pred":0.03,"meas":-0.0029,"origin":"(empty)","status":"TRYING"},{"id":"moe-50227","problem":"moe","pred":0.0005,"meas":-0.0019,"origin":"codex island","status":"TRYING"},{"id":"moe-1040","problem":"moe","pred":0.005,"meas":-0.0036,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-194641-prank_guardL1last_mid","problem":"pre","pred":0.0,"meas":-0.0012,"origin":"codex island","status":"TRYING"},{"id":"pre-50265","problem":"pre","pred":0.001,"meas":0.0024,"origin":"codex island","status":"TRYING"},{"id":"dec-1510","problem":"dec","pred":0.016,"meas":0.0158,"origin":"combination engine","status":"OPTIMIZING"},{"id":"dec-679700024","problem":"dec","pred":0.003,"meas":-0.0078,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-455","problem":"pre","pred":0.005,"meas":-0.0008,"origin":"codex island","status":"TRYING"},{"id":"bf16-135","problem":"bf16","pred":0.011,"meas":-0.0102,"origin":"codex island","status":"KILLED"},{"id":"dec-772200012","problem":"dec","pred":0.0001,"meas":0.0006,"origin":"codex island","status":"TRYING"},{"id":"attn-496500026","problem":"attn","pred":0.0025,"meas":0.0028,"origin":"evolution search","status":"TRYING"},{"id":"dec-2321","problem":"dec","pred":0.0006,"meas":0.0006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-679700131","problem":"dec","pred":0.0006,"meas":0.0002,"origin":"(empty)","status":"TRYING"},{"id":"moe-414000030","problem":"moe","pred":0.008,"meas":0.0046,"origin":"dream / reflection / resurface","status":"TRYING"},{"id":"pre-sweeper-144713-post_weights_early","problem":"pre","pred":0.0001,"meas":0.0001,"origin":"codex island","status":"TRYING"},{"id":"moe-717600005","problem":"moe","pred":0.001,"meas":-0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-137","problem":"pre","pred":0.008,"meas":0.0042,"origin":"codex island","status":"SHIPPED"},{"id":"dec-167","problem":"dec","pred":0.025,"meas":0.0207,"origin":"measurement / oracle / profiling","status":"SHIPPED"},{"id":"attn-115","problem":"attn","pred":0.015,"meas":0.0093,"origin":"codex island","status":"KILLED"},{"id":"bf16-637800043","problem":"bf16","pred":0.001,"meas":0.0051,"origin":"knob sweep","status":"TRYING"},{"id":"moe-1534","problem":"moe","pred":0.004,"meas":0.0034,"origin":"combination engine","status":"TRYING"},{"id":"dec-867700051","problem":"dec","pred":0.005,"meas":0.0049,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-765700002","problem":"dec","pred":0.005,"meas":0.0054,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-777300005","problem":"moe","pred":0.003,"meas":0.0195,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-548000060","problem":"dec","pred":0.03,"meas":-0.0013,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-517600012","problem":"dec","pred":0.0003,"meas":-0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2922","problem":"dec","pred":0.002,"meas":-0.0252,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-50126","problem":"moe","pred":0.03,"meas":-0.0078,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-391600062","problem":"dec","pred":0.004,"meas":0.2604,"origin":"(empty)","status":"OPTIMIZING"},{"id":"dec-590000001","problem":"dec","pred":0.0,"meas":0.0083,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000027","problem":"dec","pred":0.005,"meas":-0.0127,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700005","problem":"dec","pred":0.001,"meas":-0.0037,"origin":"(empty)","status":"TRYING"},{"id":"moe-1655","problem":"moe","pred":0.015,"meas":0.002,"origin":"plan files / owner","status":"TRYING"},{"id":"dec-470500141","problem":"dec","pred":0.002,"meas":-0.0159,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-50579","problem":"attn","pred":0.0005,"meas":0.0112,"origin":"codex island","status":"WORKING"},{"id":"dec-1483","problem":"dec","pred":0.0025,"meas":0.0001,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-401","problem":"pre","pred":0.003,"meas":0.0007,"origin":"(empty)","status":"TRYING"},{"id":"attn-714400001","problem":"attn","pred":0.009,"meas":-0.004,"origin":"combination engine","status":"TRYING"},{"id":"dec-589300001","problem":"dec","pred":0.002,"meas":-0.0028,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2381","problem":"dec","pred":0.0006,"meas":-0.0014,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-534","problem":"moe","pred":0.006,"meas":0.0058,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600058","problem":"dec","pred":0.001,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-714400010","problem":"bf16","pred":0.003,"meas":0.0101,"origin":"knob sweep","status":"WORKING"},{"id":"dec-468500073","problem":"dec","pred":0.0015,"meas":0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-657","problem":"attn","pred":0.016,"meas":0.0229,"origin":"measurement / oracle / profiling","status":"WORKING"},{"id":"dec-429000076","problem":"dec","pred":0.0,"meas":0.0653,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-1625","problem":"dec","pred":0.0025,"meas":-0.0201,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-093","problem":"attn","pred":0.004,"meas":0.0387,"origin":"codex island","status":"WORKING"},{"id":"dec-414000031","problem":"dec","pred":0.0012,"meas":0.0028,"origin":"port / amplify a win","status":"TRYING"},{"id":"moe-496500048","problem":"moe","pred":0.0004,"meas":0.0005,"origin":"codex island","status":"TRYING"},{"id":"dec-2373","problem":"dec","pred":0.015,"meas":0.002,"origin":"rival / peer code","status":"TRYING"},{"id":"moe-134","problem":"moe","pred":0.02,"meas":0.1034,"origin":"codex island","status":"WORKING"},{"id":"bf16-496500011","problem":"bf16","pred":0.001,"meas":-0.0667,"origin":"(empty)","status":"TRYING"},{"id":"pre-50334","problem":"pre","pred":0.0002,"meas":-0.0001,"origin":"codex island","status":"TRYING"},{"id":"attn-1620","problem":"attn","pred":0.011,"meas":0.0163,"origin":"codex island","status":"WORKING"},{"id":"moe-338","problem":"moe","pred":0.015,"meas":-0.0214,"origin":"codex island","status":"KILLED"},{"id":"pre-sweeper-140533-plan_rank_last","problem":"pre","pred":0.0001,"meas":-0.0023,"origin":"codex island","status":"TRYING"},{"id":"moe-304","problem":"moe","pred":0.005,"meas":-0.0079,"origin":"codex island","status":"KILLED"},{"id":"dec-779600027","problem":"dec","pred":0.03,"meas":-0.0116,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000077","problem":"dec","pred":0.03,"meas":0.0141,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-133","problem":"moe","pred":0.02,"meas":0.0157,"origin":"codex island","status":"KILLED"},{"id":"pre-sweeper-120109-cut28","problem":"pre","pred":0.0001,"meas":-0.0002,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-124406-o_short","problem":"pre","pred":0.0001,"meas":-0.003,"origin":"codex island","status":"TRYING"},{"id":"attn-50223","problem":"attn","pred":0.01,"meas":-0.013,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-036","problem":"dec","pred":0.05,"meas":0.3979,"origin":"codex island","status":"OPTIMIZING"},{"id":"moe-50130","problem":"moe","pred":0.025,"meas":-0.0011,"origin":"first principles / floor","status":"TRYING"},{"id":"pre-sweeper-164924-rank_l1_mid","problem":"pre","pred":0.0,"meas":-0.0044,"origin":"codex island","status":"TRYING"},{"id":"attn-867700027","problem":"attn","pred":0.002,"meas":0.0021,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300052","problem":"dec","pred":0.0005,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-396","problem":"pre","pred":0.0045,"meas":0.0026,"origin":"combination engine","status":"TRYING"},{"id":"dec-800400018","problem":"dec","pred":0.03,"meas":0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300036","problem":"dec","pred":0.03,"meas":0.004,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-1827","problem":"dec","pred":0.003,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1518","problem":"dec","pred":0.002,"meas":-0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-030","problem":"moe","pred":0.005,"meas":-0.0077,"origin":"books","status":"TRYING"},{"id":"dec-468500079","problem":"dec","pred":0.008,"meas":0.5931,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-sweeper-171855-layout_bigc_l1first","problem":"pre","pred":0.0,"meas":0.0037,"origin":"codex island","status":"TRYING"},{"id":"attn-085","problem":"attn","pred":0.012,"meas":0.0,"origin":"codex island","status":"WORKING"},{"id":"dec-414000108","problem":"dec","pred":0.002,"meas":-0.0029,"origin":"(empty)","status":"TRYING"},{"id":"moe-50159","problem":"moe","pred":0.004,"meas":0.0057,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-837","problem":"attn","pred":0.004,"meas":-0.0398,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-449","problem":"pre","pred":0.01,"meas":-0.0028,"origin":"codex island","status":"TRYING"},{"id":"dec-414000035","problem":"dec","pred":0.001,"meas":0.0001,"origin":"port / amplify a win","status":"TRYING"},{"id":"attn-721","problem":"attn","pred":0.012,"meas":-0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50000055","problem":"dec","pred":0.009,"meas":0.0074,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-114822-score_umax","problem":"pre","pred":0.0001,"meas":-0.0042,"origin":"codex island","status":"TRYING"},{"id":"dec-2123","problem":"dec","pred":0.0005,"meas":0.0008,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-888100048","problem":"dec","pred":0.0005,"meas":-0.0013,"origin":"(empty)","status":"TRYING"},{"id":"moe-125","problem":"moe","pred":0.008,"meas":-0.0008,"origin":"codex island","status":"TRYING"},{"id":"pre-610","problem":"pre","pred":0.006,"meas":-0.0001,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-153552-ranks_v2_post","problem":"pre","pred":0.0001,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"moe-679700002","problem":"moe","pred":0.0025,"meas":0.0013,"origin":"(empty)","status":"TRYING"},{"id":"attn-414000063","problem":"attn","pred":0.005,"meas":0.0045,"origin":"(empty)","status":"TRYING"},{"id":"pre-50260","problem":"pre","pred":0.0003,"meas":0.0064,"origin":"codex island","status":"TRYING"},{"id":"dec-517600028","problem":"dec","pred":0.004,"meas":0.0012,"origin":"combination engine","status":"TRYING"},{"id":"dec-2607","problem":"dec","pred":0.0004,"meas":-0.0075,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50000002","problem":"moe","pred":0.003,"meas":0.0031,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"pre-50345","problem":"pre","pred":0.0002,"meas":0.002,"origin":"codex island","status":"TRYING"},{"id":"dec-888100061","problem":"dec","pred":0.002,"meas":0.0045,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300001","problem":"dec","pred":0.03,"meas":0.0113,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-800400070","problem":"dec","pred":0.03,"meas":-0.0007,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-1021","problem":"moe","pred":0.004,"meas":-0.003,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-589300003","problem":"dec","pred":0.03,"meas":0.0041,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2787","problem":"dec","pred":0.0,"meas":0.7127,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"pre-50289","problem":"pre","pred":0.0002,"meas":0.0081,"origin":"codex island","status":"TRYING"},{"id":"attn-944","problem":"attn","pred":0.004,"meas":-0.0028,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-181756-loc_mask_short","problem":"pre","pred":0.0,"meas":0.0019,"origin":"codex island","status":"TRYING"},{"id":"dec-888100063","problem":"dec","pred":0.002,"meas":0.75,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1526","problem":"dec","pred":0.002,"meas":0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-635","problem":"pre","pred":0.003,"meas":0.0002,"origin":"learned priors","status":"TRYING"},{"id":"attn-1732","problem":"attn","pred":0.01,"meas":0.0172,"origin":"combination engine","status":"WORKING"},{"id":"moe-15300002","problem":"moe","pred":0.009,"meas":0.0159,"origin":"evolution search","status":"OPTIMIZING"},{"id":"dec-2015","problem":"dec","pred":0.0015,"meas":0.0015,"origin":"codex island","status":"TRYING"},{"id":"moe-1481","problem":"moe","pred":0.008,"meas":0.0089,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-590000013","problem":"attn","pred":0.0,"meas":-0.0767,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-800400015","problem":"dec","pred":0.03,"meas":-0.0037,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-800400021","problem":"dec","pred":0.03,"meas":-0.0031,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2378","problem":"dec","pred":0.025,"meas":0.0033,"origin":"plan files / owner","status":"TRYING"},{"id":"dec-178","problem":"dec","pred":0.012,"meas":0.0281,"origin":"lane's own work (origin not stated)","status":"SHIPPED"},{"id":"attn-774","problem":"attn","pred":0.01,"meas":-0.0014,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-548000009","problem":"attn","pred":0.001,"meas":0.0024,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-067","problem":"attn","pred":0.01,"meas":0.021,"origin":"codex island","status":"WORKING"},{"id":"dec-679700029","problem":"dec","pred":0.004,"meas":-0.0012,"origin":"(empty)","status":"TRYING"},{"id":"moe-335","problem":"moe","pred":0.007,"meas":0.0002,"origin":"codex island","status":"TRYING"},{"id":"attn-209","problem":"attn","pred":0.012,"meas":0.0108,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-1553","problem":"dec","pred":0.003,"meas":0.0025,"origin":"plan files / owner","status":"TRYING"},{"id":"attn-1746","problem":"attn","pred":0.01,"meas":0.0012,"origin":"plan files / owner","status":"TRYING"},{"id":"moe-50237","problem":"moe","pred":0.0002,"meas":-0.0049,"origin":"codex island","status":"TRYING"},{"id":"pre-456300003","problem":"pre","pred":0.0001,"meas":-0.0053,"origin":"codex island","status":"TRYING"},{"id":"moe-470500039","problem":"moe","pred":0.0002,"meas":-0.0004,"origin":"codex island","status":"TRYING"},{"id":"moe-1103","problem":"moe","pred":0.003,"meas":0.0025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2814","problem":"dec","pred":0.0045,"meas":0.0046,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2382","problem":"dec","pred":0.002,"meas":-0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50569","problem":"attn","pred":0.0003,"meas":-0.0136,"origin":"codex island","status":"TRYING"},{"id":"attn-599200012","problem":"attn","pred":0.005,"meas":-0.0026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-142648-scatter_mid_meta_last","problem":"pre","pred":0.0001,"meas":-0.0006,"origin":"codex island","status":"TRYING"},{"id":"dec-468500064","problem":"dec","pred":0.03,"meas":-0.0105,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-661","problem":"pre","pred":0.005,"meas":-0.0001,"origin":"codex island","status":"TRYING"},{"id":"pre-744","problem":"pre","pred":0.0025,"meas":0.005,"origin":"(empty)","status":"TRYING"},{"id":"attn-589300005","problem":"attn","pred":0.01,"meas":-0.0014,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"pre-sweeper-171855-layout_segb_l1last","problem":"pre","pred":0.0,"meas":0.0024,"origin":"codex island","status":"TRYING"},{"id":"attn-673600023","problem":"attn","pred":0.008,"meas":0.0154,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-714500034","problem":"dec","pred":0.004,"meas":0.0081,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-247","problem":"moe","pred":0.01,"meas":-0.0016,"origin":"codex island","status":"KILLED"},{"id":"pre-726","problem":"pre","pred":0.035,"meas":0.0368,"origin":"rival / peer code","status":"OPTIMIZING"},{"id":"bf16-734800001","problem":"bf16","pred":0.0005,"meas":-0.002,"origin":"combination engine","status":"TRYING"},{"id":"dec-015","problem":"dec","pred":0.03,"meas":-0.0025,"origin":"books","status":"SHIPPED"},{"id":"dec-867700035","problem":"dec","pred":0.03,"meas":0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50251","problem":"moe","pred":0.001,"meas":0.0068,"origin":"codex island","status":"TRYING"},{"id":"dec-470500078","problem":"dec","pred":0.0005,"meas":0.0,"origin":"codex island","status":"TRYING"},{"id":"attn-1310","problem":"attn","pred":0.003,"meas":0.0009,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-697","problem":"bf16","pred":0.02,"meas":0.0256,"origin":"rival / peer code","status":"WORKING"},{"id":"dec-50000045","problem":"dec","pred":0.0001,"meas":0.0024,"origin":"(empty)","status":"TRYING"},{"id":"dec-779600048","problem":"dec","pred":0.002,"meas":0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-829","problem":"moe","pred":0.006,"meas":0.0408,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-1731","problem":"dec","pred":0.01,"meas":0.0,"origin":"codex island","status":"TRYING"},{"id":"attn-692","problem":"attn","pred":0.008,"meas":-0.0032,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"moe-1613","problem":"moe","pred":0.0003,"meas":0.0035,"origin":"codex island","status":"TRYING"},{"id":"dec-1480","problem":"dec","pred":0.0025,"meas":0.0071,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-2656","problem":"dec","pred":0.0,"meas":-0.0002,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-204900004","problem":"dec","pred":0.001,"meas":-0.0027,"origin":"(empty)","status":"TRYING"},{"id":"attn-50010","problem":"attn","pred":0.003,"meas":-0.0037,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-470500090","problem":"dec","pred":0.0,"meas":-0.0001,"origin":"plan files / owner","status":"TRYING"},{"id":"dec-391600059","problem":"dec","pred":0.002,"meas":-0.018,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-50000003","problem":"moe","pred":0.002,"meas":0.0131,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2379","problem":"dec","pred":0.002,"meas":0.0033,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-496500051","problem":"moe","pred":0.0004,"meas":-0.0078,"origin":"codex island","status":"TRYING"},{"id":"pre-50081","problem":"pre","pred":0.007,"meas":0.0022,"origin":"combination engine","status":"TRYING"},{"id":"dec-2045","problem":"dec","pred":0.004,"meas":0.0002,"origin":"codex island","status":"TRYING"},{"id":"pre-414000013","problem":"pre","pred":0.002,"meas":-0.0029,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-105205-poll64","problem":"pre","pred":0.0001,"meas":-0.0128,"origin":"codex island","status":"TRYING"},{"id":"moe-867700004","problem":"moe","pred":0.0,"meas":-0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500065","problem":"dec","pred":0.002,"meas":0.03,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50259","problem":"moe","pred":0.005,"meas":-0.0075,"origin":"knob sweep","status":"TRYING"},{"id":"bf16-637800026","problem":"bf16","pred":0.2,"meas":0.0004,"origin":"(empty)","status":"TRYING"},{"id":"moe-603","problem":"moe","pred":0.006,"meas":0.0111,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-414000071","problem":"dec","pred":0.0035,"meas":0.0048,"origin":"rival / peer code","status":"TRYING"},{"id":"attn-1309","problem":"attn","pred":0.003,"meas":-0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50268","problem":"pre","pred":0.0001,"meas":-0.0121,"origin":"codex island","status":"TRYING"},{"id":"bf16-008","problem":"bf16","pred":0.015,"meas":-0.056,"origin":"papers (arXiv)","status":"TRYING"},{"id":"bf16-637800017","problem":"bf16","pred":0.002,"meas":-0.0034,"origin":"knob sweep","status":"TRYING"},{"id":"moe-463200003","problem":"moe","pred":0.0,"meas":-0.2354,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-414000019","problem":"pre","pred":0.002,"meas":0.0024,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"dec-012","problem":"dec","pred":0.03,"meas":0.0055,"origin":"plan files / owner","status":"SHIPPED"},{"id":"bf16-588","problem":"bf16","pred":0.003,"meas":-0.0012,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50258","problem":"bf16","pred":0.0003,"meas":-0.0103,"origin":"codex island","status":"TRYING"},{"id":"moe-1028","problem":"moe","pred":0.004,"meas":0.0024,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50000058","problem":"dec","pred":0.0,"meas":-0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-599200001","problem":"dec","pred":0.005,"meas":0.0241,"origin":"(empty)","status":"WORKING"},{"id":"dec-414000118","problem":"dec","pred":0.003,"meas":-0.0032,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-517600014","problem":"dec","pred":0.0003,"meas":-0.0032,"origin":"(empty)","status":"TRYING"},{"id":"dec-714400001","problem":"dec","pred":0.001,"meas":-0.0013,"origin":"combination engine","status":"TRYING"},{"id":"moe-414000037","problem":"moe","pred":0.003,"meas":0.0063,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-1026","problem":"dec","pred":0.012,"meas":0.0422,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"moe-679700005","problem":"moe","pred":0.0,"meas":-0.0083,"origin":"(empty)","status":"TRYING"},{"id":"pre-50153","problem":"pre","pred":0.002,"meas":0.0003,"origin":"port / amplify a win","status":"TRYING"},{"id":"attn-463200018","problem":"attn","pred":0.5,"meas":0.0111,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-734800019","problem":"dec","pred":0.0002,"meas":0.0096,"origin":"combination engine","status":"TRYING"},{"id":"pre-50241","problem":"pre","pred":0.0002,"meas":-0.0008,"origin":"codex island","status":"TRYING"},{"id":"dec-204900003","problem":"dec","pred":0.0005,"meas":0.0006,"origin":"(empty)","status":"TRYING"},{"id":"attn-743","problem":"attn","pred":0.004,"meas":-0.0062,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"dec-2261","problem":"dec","pred":0.0012,"meas":0.0005,"origin":"evolution search","status":"TRYING"},{"id":"dec-548000033","problem":"dec","pred":0.001,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-001","problem":"moe","pred":0.12,"meas":0.0645,"origin":"nature / fractals","status":"SHIPPED"},{"id":"dec-468500042","problem":"dec","pred":0.0,"meas":0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2347","problem":"dec","pred":0.004,"meas":-0.0002,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-2470","problem":"dec","pred":0.003,"meas":0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"all-126","problem":"all","pred":0.0,"meas":0.8128,"origin":"first principles / floor","status":"WORKING"},{"id":"dec-772200023","problem":"dec","pred":0.0,"meas":-0.0011,"origin":"codex island","status":"TRYING"},{"id":"moe-15300011","problem":"moe","pred":0.02,"meas":-0.0212,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1019","problem":"moe","pred":0.008,"meas":0.0029,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-867700030","problem":"attn","pred":0.003,"meas":-0.0327,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600081","problem":"dec","pred":0.0,"meas":-0.0022,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-50541","problem":"attn","pred":0.0005,"meas":0.003,"origin":"codex island","status":"TRYING"},{"id":"attn-1278","problem":"attn","pred":0.015,"meas":-0.0122,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-777300002","problem":"attn","pred":0.0,"meas":0.2291,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-394","problem":"attn","pred":0.0135,"meas":0.0049,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-590000008","problem":"dec","pred":0.0,"meas":0.0382,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-589300010","problem":"dec","pred":0.03,"meas":-0.0016,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-1560","problem":"moe","pred":0.004,"meas":0.0035,"origin":"codex island","status":"TRYING"},{"id":"attn-50580","problem":"attn","pred":0.0005,"meas":-0.0099,"origin":"codex island","status":"TRYING"},{"id":"attn-714500097","problem":"attn","pred":0.005,"meas":-0.0191,"origin":"manager hypothesis generator","status":"TRYING"},{"id":"bf16-693","problem":"bf16","pred":0.001,"meas":-0.0001,"origin":"combination engine","status":"TRYING"},{"id":"attn-738","problem":"attn","pred":0.008,"meas":0.0008,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"dec-714500041","problem":"dec","pred":0.0,"meas":0.0248,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-663","problem":"pre","pred":0.0085,"meas":-0.0041,"origin":"codex island","status":"TRYING"},{"id":"moe-50072","problem":"moe","pred":0.01,"meas":0.0003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-287","problem":"dec","pred":0.007,"meas":0.0052,"origin":"codex island","status":"KILLED"},{"id":"attn-1744","problem":"attn","pred":0.0012,"meas":0.0017,"origin":"combination engine","status":"TRYING"},{"id":"moe-517600030","problem":"moe","pred":0.003,"meas":-0.005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50248","problem":"pre","pred":0.002,"meas":-0.0039,"origin":"codex island","status":"TRYING"},{"id":"dec-861200001","problem":"dec","pred":0.006,"meas":0.0072,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-794","problem":"attn","pred":0.003,"meas":0.0064,"origin":"web / blogs / deep research","status":"WORKING"},{"id":"dec-2767","problem":"dec","pred":0.0015,"meas":0.0009,"origin":"combination engine","status":"TRYING"},{"id":"dec-800400009","problem":"dec","pred":0.001,"meas":-0.0073,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-611","problem":"pre","pred":0.003,"meas":0.0017,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-714500059","problem":"dec","pred":0.008,"meas":-0.0011,"origin":"manager hypothesis generator","status":"KILLED"},{"id":"dec-599200011","problem":"dec","pred":0.03,"meas":0.0103,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-734800004","problem":"dec","pred":0.0005,"meas":-0.0177,"origin":"knob sweep","status":"TRYING"},{"id":"attn-463200016","problem":"attn","pred":0.015,"meas":0.0183,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2361","problem":"dec","pred":0.0,"meas":-0.0008,"origin":"combination engine","status":"TRYING"},{"id":"dec-2289","problem":"dec","pred":0.0001,"meas":0.0026,"origin":"codex island","status":"TRYING"},{"id":"pre-50259","problem":"pre","pred":0.0002,"meas":-0.0071,"origin":"codex island","status":"TRYING"},{"id":"dec-548000067","problem":"dec","pred":0.03,"meas":0.0055,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-599200011","problem":"attn","pred":0.006,"meas":0.006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50000020","problem":"dec","pred":0.0,"meas":0.2621,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"pre-378","problem":"pre","pred":0.0028,"meas":0.0046,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-552","problem":"bf16","pred":0.003,"meas":-0.0561,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-748","problem":"attn","pred":0.004,"meas":0.002,"origin":"combination engine","status":"TRYING"},{"id":"dec-1509","problem":"dec","pred":0.003,"meas":-0.0156,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-185950-pexp_wt_all","problem":"pre","pred":0.0,"meas":0.0012,"origin":"codex island","status":"TRYING"},{"id":"pre-734800007","problem":"pre","pred":0.035,"meas":0.0379,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"moe-1713","problem":"moe","pred":0.004,"meas":-0.3217,"origin":"(empty)","status":"TRYING"},{"id":"dec-414000055","problem":"dec","pred":0.001,"meas":-0.0025,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-1569","problem":"dec","pred":0.008,"meas":0.0016,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600083","problem":"dec","pred":0.002,"meas":-0.0034,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-50234","problem":"moe","pred":0.0002,"meas":0.0085,"origin":"codex island","status":"TRYING"},{"id":"pre-50394","problem":"pre","pred":0.0045,"meas":0.0006,"origin":"combination engine","status":"TRYING"},{"id":"bf16-637800030","problem":"bf16","pred":0.005,"meas":-0.0033,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-517600006","problem":"moe","pred":0.002,"meas":-0.0034,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1388","problem":"moe","pred":0.003,"meas":0.0085,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-15300003","problem":"dec","pred":0.002,"meas":-0.0085,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-414000078","problem":"dec","pred":0.002,"meas":0.0021,"origin":"(empty)","status":"TRYING"},{"id":"attn-496500107","problem":"attn","pred":0.0,"meas":-0.0352,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-429000025","problem":"dec","pred":0.0002,"meas":-0.0002,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-414000075","problem":"dec","pred":0.0,"meas":-0.0006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-679700138","problem":"dec","pred":0.0007,"meas":-0.0013,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-637800002","problem":"attn","pred":0.006,"meas":0.0044,"origin":"combination engine","status":"WORKING"},{"id":"pre-50274","problem":"pre","pred":0.001,"meas":0.0011,"origin":"codex island","status":"TRYING"},{"id":"pre-395","problem":"pre","pred":0.008,"meas":0.0016,"origin":"plan files / owner","status":"KILLED"},{"id":"dec-2919","problem":"dec","pred":0.0,"meas":0.0936,"origin":"first principles / floor","status":"OPTIMIZING"},{"id":"moe-229","problem":"moe","pred":0.016,"meas":-0.0189,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"pre-620","problem":"pre","pred":0.005,"meas":-0.0014,"origin":"codex island","status":"TRYING"},{"id":"moe-116","problem":"moe","pred":0.01,"meas":-0.0153,"origin":"codex island","status":"KILLED"},{"id":"attn-463300012","problem":"attn","pred":0.003,"meas":-0.001,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-548000063","problem":"dec","pred":0.03,"meas":-0.0184,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-772200011","problem":"dec","pred":0.0002,"meas":-0.0012,"origin":"codex island","status":"TRYING"},{"id":"bf16-032","problem":"bf16","pred":0.03,"meas":-0.1762,"origin":"codex island","status":"TRYING"},{"id":"attn-50000004","problem":"attn","pred":0.01,"meas":-0.0072,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-115","problem":"pre","pred":0.005,"meas":-0.0016,"origin":"manager hypothesis generator","status":"TRYING"},{"id":"dec-800400038","problem":"dec","pred":0.03,"meas":-0.0071,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-772200019","problem":"dec","pred":0.0001,"meas":0.0011,"origin":"codex island","status":"TRYING"},{"id":"moe-010","problem":"moe","pred":0.03,"meas":-0.2787,"origin":"plan files / owner","status":"KILLED"},{"id":"moe-1618","problem":"moe","pred":0.002,"meas":-0.0097,"origin":"evolution search","status":"TRYING"},{"id":"dec-2964","problem":"dec","pred":0.0008,"meas":-0.0008,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-888100023","problem":"dec","pred":0.0005,"meas":-0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-015","problem":"bf16","pred":0.02,"meas":0.0073,"origin":"first principles / floor","status":"KILLED"},{"id":"pre-694","problem":"pre","pred":0.0008,"meas":0.0033,"origin":"codex island","status":"TRYING"},{"id":"moe-463200008","problem":"moe","pred":0.003,"meas":0.0027,"origin":"(empty)","status":"TRYING"},{"id":"pre-643","problem":"pre","pred":0.003,"meas":0.0007,"origin":"codex island","status":"TRYING"},{"id":"attn-835","problem":"attn","pred":0.015,"meas":-0.0324,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1883","problem":"attn","pred":0.0005,"meas":0.0008,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-777300048","problem":"dec","pred":0.003,"meas":-0.724,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-1744","problem":"dec","pred":0.01,"meas":-0.0091,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-414000123","problem":"attn","pred":0.001,"meas":0.0009,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-637800006","problem":"bf16","pred":0.002,"meas":-0.0032,"origin":"knob sweep","status":"TRYING"},{"id":"dec-2904","problem":"dec","pred":0.0009,"meas":0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600087","problem":"dec","pred":0.0005,"meas":-0.0013,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-468500002","problem":"moe","pred":0.03,"meas":-0.0232,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"bf16-50080","problem":"bf16","pred":0.004,"meas":-0.0212,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2909","problem":"dec","pred":0.0,"meas":0.0557,"origin":"first principles / floor","status":"OPTIMIZING"},{"id":"moe-717600006","problem":"moe","pred":0.001,"meas":-0.0004,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-637800005","problem":"dec","pred":0.0005,"meas":-0.0018,"origin":"knob sweep","status":"TRYING"},{"id":"attn-826","problem":"attn","pred":0.015,"meas":0.0078,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-637800001","problem":"moe","pred":0.001,"meas":0.0024,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1089","problem":"moe","pred":0.009,"meas":0.0028,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1695","problem":"moe","pred":0.002,"meas":-0.0004,"origin":"learned priors","status":"TRYING"},{"id":"dec-2309","problem":"dec","pred":0.0007,"meas":0.0011,"origin":"combination engine","status":"TRYING"},{"id":"attn-762","problem":"attn","pred":0.003,"meas":-0.0026,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1323","problem":"attn","pred":0.003,"meas":0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-463200014","problem":"dec","pred":0.0,"meas":-0.0175,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-496500069","problem":"moe","pred":0.0003,"meas":0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-125844-route_plan_ef","problem":"pre","pred":0.0001,"meas":-0.0009,"origin":"codex island","status":"TRYING"},{"id":"dec-772200002","problem":"dec","pred":0.0004,"meas":0.0016,"origin":"codex island","status":"TRYING"},{"id":"dec-414000012","problem":"dec","pred":0.004,"meas":0.0027,"origin":"combination engine","status":"TRYING"},{"id":"moe-749100001","problem":"moe","pred":0.001,"meas":0.0047,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-772200005","problem":"dec","pred":0.0006,"meas":-0.0009,"origin":"codex island","status":"TRYING"},{"id":"dec-679700001","problem":"dec","pred":0.002,"meas":-0.0026,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-1279","problem":"attn","pred":0.018,"meas":-0.0046,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-496500055","problem":"moe","pred":0.001,"meas":0.001,"origin":"codex island","status":"TRYING"},{"id":"attn-1311","problem":"attn","pred":0.003,"meas":0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-599200005","problem":"dec","pred":0.003,"meas":0.0017,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-188","problem":"dec","pred":0.012,"meas":0.0133,"origin":"codex island","status":"OPTIMIZING"},{"id":"bf16-711","problem":"bf16","pred":0.001,"meas":0.0065,"origin":"combination engine","status":"TRYING"},{"id":"bf16-122","problem":"bf16","pred":0.008,"meas":0.0033,"origin":"codex island","status":"KILLED"},{"id":"pre-50203","problem":"pre","pred":0.012,"meas":0.0007,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-391600078","problem":"dec","pred":0.0,"meas":0.2261,"origin":"measurement / oracle / profiling","status":"OPTIMIZING"},{"id":"dec-888100059","problem":"dec","pred":0.001,"meas":0.1047,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-160104-guard_v2_long","problem":"pre","pred":0.0001,"meas":-0.0017,"origin":"codex island","status":"TRYING"},{"id":"moe-246","problem":"moe","pred":0.022,"meas":0.02,"origin":"combination engine","status":"WORKING"},{"id":"moe-50099","problem":"moe","pred":0.005,"meas":-0.0087,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50184","problem":"pre","pred":0.01,"meas":0.0004,"origin":"combination engine","status":"TRYING"},{"id":"attn-50567","problem":"attn","pred":0.0003,"meas":-0.0114,"origin":"codex island","status":"TRYING"},{"id":"attn-463300010","problem":"attn","pred":0.003,"meas":0.0309,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"pre-496500061","problem":"pre","pred":0.0,"meas":-0.0036,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-50181","problem":"dec","pred":0.001,"meas":-0.0024,"origin":"codex island","status":"TRYING"},{"id":"bf16-50334","problem":"bf16","pred":0.0001,"meas":-0.0029,"origin":"knob sweep","status":"TRYING"},{"id":"dec-734800006","problem":"dec","pred":0.0005,"meas":-0.0016,"origin":"knob sweep","status":"TRYING"},{"id":"dec-548000039","problem":"dec","pred":0.008,"meas":0.1681,"origin":"(empty)","status":"TRYING"},{"id":"moe-238","problem":"moe","pred":0.015,"meas":-0.0233,"origin":"codex island","status":"KILLED"},{"id":"attn-749100004","problem":"attn","pred":0.001,"meas":0.0,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-517600036","problem":"moe","pred":0.004,"meas":0.0174,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-2391","problem":"dec","pred":0.0003,"meas":0.0017,"origin":"(empty)","status":"TRYING"},{"id":"moe-15300022","problem":"moe","pred":0.03,"meas":0.0606,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-414000027","problem":"moe","pred":0.0012,"meas":-0.0015,"origin":"(empty)","status":"KILLED"},{"id":"dec-867700001","problem":"dec","pred":0.002,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1622","problem":"attn","pred":0.004,"meas":0.004,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"dec-777300023","problem":"dec","pred":0.001,"meas":-0.0022,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-1382","problem":"moe","pred":0.003,"meas":0.0045,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-091758-tanh128","problem":"pre","pred":0.0002,"meas":-0.0101,"origin":"codex island","status":"TRYING"},{"id":"dec-2198","problem":"dec","pred":0.0009,"meas":0.0008,"origin":"codex island","status":"TRYING"},{"id":"dec-762300008","problem":"dec","pred":0.003,"meas":0.0141,"origin":"(empty)","status":"TRYING"},{"id":"attn-126","problem":"attn","pred":0.022,"meas":0.0431,"origin":"lane's own work (origin not stated)","status":"OPTIMIZING"},{"id":"attn-163","problem":"attn","pred":0.009,"meas":0.0246,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-50122","problem":"attn","pred":0.008,"meas":0.0025,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-50203","problem":"dec","pred":0.001,"meas":-0.0022,"origin":"knob sweep","status":"TRYING"},{"id":"dec-589300014","problem":"dec","pred":0.002,"meas":-0.0036,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-429000069","problem":"dec","pred":0.0003,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-463200018","problem":"dec","pred":0.004,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1581","problem":"dec","pred":0.004,"meas":0.0012,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-287","problem":"moe","pred":0.006,"meas":0.0024,"origin":"codex island","status":"KILLED"},{"id":"moe-414000017","problem":"moe","pred":0.006,"meas":0.0151,"origin":"(empty)","status":"WORKING"},{"id":"moe-1399","problem":"moe","pred":0.01,"meas":-0.0004,"origin":"codex island","status":"TRYING"},{"id":"dec-800400040","problem":"dec","pred":0.001,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1837","problem":"attn","pred":0.006,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-679700055","problem":"dec","pred":0.002,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-414000014","problem":"pre","pred":0.003,"meas":-0.0037,"origin":"dream / reflection / resurface","status":"TRYING"},{"id":"attn-50250","problem":"attn","pred":0.005,"meas":0.0181,"origin":"first principles / floor","status":"WORKING"},{"id":"pre-sweeper-094846-both_last","problem":"pre","pred":0.0003,"meas":-0.0108,"origin":"codex island","status":"TRYING"},{"id":"moe-1020","problem":"moe","pred":0.006,"meas":-0.0018,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-407","problem":"attn","pred":0.008,"meas":0.0213,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50330","problem":"pre","pred":0.0001,"meas":-0.0022,"origin":"codex island","status":"TRYING"},{"id":"pre-50331","problem":"pre","pred":0.0001,"meas":0.0004,"origin":"codex island","status":"TRYING"},{"id":"attn-867700006","problem":"attn","pred":0.001,"meas":-0.0111,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000038","problem":"dec","pred":0.002,"meas":0.0059,"origin":"combination engine","status":"TRYING"},{"id":"pre-50336","problem":"pre","pred":0.001,"meas":-0.0135,"origin":"codex island","status":"TRYING"},{"id":"attn-765700010","problem":"attn","pred":0.03,"meas":-0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-131","problem":"moe","pred":0.012,"meas":0.004,"origin":"combination engine","status":"TRYING"},{"id":"moe-1694","problem":"moe","pred":0.004,"meas":-0.0019,"origin":"learned priors","status":"TRYING"},{"id":"dec-166","problem":"dec","pred":0.032,"meas":0.0324,"origin":"codex island","status":"WORKING"},{"id":"pre-50096","problem":"pre","pred":0.004,"meas":0.001,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-sweeper-171855-layout_bigc_l1last","problem":"pre","pred":0.0,"meas":0.0043,"origin":"codex island","status":"TRYING"},{"id":"dec-888100043","problem":"dec","pred":0.002,"meas":-0.0072,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-1326","problem":"attn","pred":0.003,"meas":0.0009,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-888100027","problem":"dec","pred":0.01,"meas":-0.0011,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-120109-cut12","problem":"pre","pred":0.0001,"meas":-0.0071,"origin":"codex island","status":"TRYING"},{"id":"moe-1318","problem":"moe","pred":0.007,"meas":-0.0097,"origin":"codex island","status":"TRYING"},{"id":"attn-1302","problem":"attn","pred":0.002,"meas":-0.0047,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-664","problem":"pre","pred":0.006,"meas":-0.0658,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-765700003","problem":"dec","pred":0.005,"meas":-0.0192,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-888100001","problem":"dec","pred":0.001,"meas":-0.014,"origin":"(empty)","status":"TRYING"},{"id":"attn-391600003","problem":"attn","pred":0.003,"meas":0.0203,"origin":"(empty)","status":"WORKING"},{"id":"pre-sweeper-194641-prank_last_mid","problem":"pre","pred":0.0,"meas":-0.0467,"origin":"codex island","status":"TRYING"},{"id":"muse_decode_layer-779600001","problem":"dec","pred":0.002,"meas":0.0198,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-1767","problem":"attn","pred":0.008,"meas":0.006,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-sweeper-112510-ids_cs","problem":"pre","pred":0.0001,"meas":-0.001,"origin":"codex island","status":"TRYING"},{"id":"attn-50490","problem":"attn","pred":0.016,"meas":0.0094,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-765700008","problem":"dec","pred":0.003,"meas":0.0032,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50581","problem":"attn","pred":0.0005,"meas":0.0142,"origin":"codex island","status":"WORKING"},{"id":"dec-463200006","problem":"dec","pred":0.0,"meas":0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-310","problem":"moe","pred":0.007,"meas":0.0009,"origin":"codex island","status":"KILLED"},{"id":"bf16-031","problem":"bf16","pred":0.035,"meas":-0.135,"origin":"codex island","status":"TRYING"},{"id":"attn-548000010","problem":"attn","pred":0.002,"meas":-0.2738,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-824200002","problem":"attn","pred":0.003,"meas":0.0012,"origin":"port / amplify a win","status":"TRYING"},{"id":"attn-470500020","problem":"attn","pred":0.0001,"meas":0.0007,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"bf16-179","problem":"bf16","pred":0.0,"meas":0.0,"origin":"codex island","status":"KILLED"},{"id":"dec-734800009","problem":"dec","pred":0.0005,"meas":0.0002,"origin":"knob sweep","status":"TRYING"},{"id":"dec-468500020","problem":"dec","pred":0.004,"meas":1.4979,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000043","problem":"dec","pred":0.0005,"meas":-0.0016,"origin":"rival / peer code","status":"TRYING"},{"id":"moe-414000024","problem":"moe","pred":0.002,"meas":-0.0022,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-473","problem":"bf16","pred":0.008,"meas":0.0057,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"dec-429000054","problem":"dec","pred":0.0,"meas":0.0129,"origin":"(empty)","status":"WORKING"},{"id":"dec-414000110","problem":"dec","pred":0.0007,"meas":-0.0015,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50169","problem":"dec","pred":0.003,"meas":-0.0006,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-391600091","problem":"dec","pred":0.02,"meas":-0.0021,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"moe-1134","problem":"moe","pred":0.006,"meas":-0.0219,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-714400002","problem":"bf16","pred":0.0,"meas":0.0056,"origin":"knob sweep","status":"TRYING"},{"id":"dec-777300035","problem":"dec","pred":0.03,"meas":0.0005,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-470500024","problem":"attn","pred":0.0002,"meas":0.0107,"origin":"codex island","status":"TRYING"},{"id":"moe-332","problem":"moe","pred":0.01,"meas":-0.0027,"origin":"codex island","status":"TRYING"},{"id":"moe-470500063","problem":"moe","pred":0.0003,"meas":0.0089,"origin":"codex island","status":"TRYING"},{"id":"dec-2671","problem":"dec","pred":0.0,"meas":0.0013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-734800008","problem":"bf16","pred":0.02,"meas":0.0259,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-429000022","problem":"dec","pred":0.0005,"meas":0.0003,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-3028","problem":"dec","pred":0.004,"meas":0.0037,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2061","problem":"dec","pred":0.0013,"meas":-0.0003,"origin":"codex island","status":"TRYING"},{"id":"dec-2803","problem":"dec","pred":0.0005,"meas":-0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600057","problem":"dec","pred":0.0,"meas":-0.0013,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-756","problem":"attn","pred":0.006,"meas":-0.016,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2290","problem":"dec","pred":0.0001,"meas":0.0005,"origin":"codex island","status":"TRYING"},{"id":"dec-762300018","problem":"dec","pred":0.1,"meas":-0.002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50347","problem":"pre","pred":0.0001,"meas":-0.0057,"origin":"codex island","status":"TRYING"},{"id":"dec-637800004","problem":"dec","pred":0.0005,"meas":-0.0008,"origin":"knob sweep","status":"TRYING"},{"id":"attn-683","problem":"attn","pred":0.012,"meas":0.0147,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"dec-2645","problem":"dec","pred":0.0,"meas":-0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50217","problem":"moe","pred":0.0003,"meas":0.0004,"origin":"codex island","status":"TRYING"},{"id":"dec-463200001","problem":"dec","pred":0.03,"meas":-0.0108,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1359","problem":"moe","pred":0.018,"meas":-0.0237,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-50229","problem":"moe","pred":0.001,"meas":-0.0006,"origin":"codex island","status":"TRYING"},{"id":"bf16-496500002","problem":"bf16","pred":0.0,"meas":-0.0089,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-468500001","problem":"attn","pred":0.02,"meas":0.0265,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-590000006","problem":"attn","pred":0.003,"meas":0.0071,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-867700007","problem":"attn","pred":0.015,"meas":0.0157,"origin":"lane's own work (origin not stated)","status":"OPTIMIZING"},{"id":"dec-2154","problem":"dec","pred":0.0038,"meas":0.0018,"origin":"codex island","status":"TRYING"},{"id":"pre-698","problem":"pre","pred":0.0003,"meas":0.0003,"origin":"codex island","status":"TRYING"},{"id":"pre-50117","problem":"pre","pred":0.002,"meas":-0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-054","problem":"moe","pred":0.12,"meas":0.0302,"origin":"codex island","status":"WORKING"},{"id":"attn-740","problem":"attn","pred":0.008,"meas":0.002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1321","problem":"dec","pred":0.0,"meas":0.0173,"origin":"measurement / oracle / profiling","status":"WORKING"},{"id":"dec-50058","problem":"dec","pred":0.005,"meas":-0.0003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-779600009","problem":"dec","pred":0.002,"meas":0.0025,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-2054","problem":"dec","pred":0.003,"meas":-0.0018,"origin":"codex island","status":"KILLED"},{"id":"bf16-586","problem":"bf16","pred":0.003,"meas":-0.0009,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1923","problem":"attn","pred":0.006,"meas":0.0058,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-15300005","problem":"dec","pred":0.002,"meas":-0.016,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-517600043","problem":"dec","pred":0.003,"meas":0.0008,"origin":"combination engine","status":"TRYING"},{"id":"moe-749100004","problem":"moe","pred":0.0,"meas":-0.0455,"origin":"(empty)","status":"TRYING"},{"id":"dec-414000016","problem":"dec","pred":0.004,"meas":0.0078,"origin":"(empty)","status":"TRYING"},{"id":"dec-777300008","problem":"dec","pred":0.03,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-477","problem":"bf16","pred":0.012,"meas":0.0089,"origin":"port / amplify a win","status":"WORKING"},{"id":"all-121","problem":"all","pred":0.0,"meas":0.1042,"origin":"first principles / floor","status":"WORKING"},{"id":"dec-2705","problem":"dec","pred":0.003,"meas":0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-463200024","problem":"attn","pred":0.06,"meas":0.0621,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-151021-weights_v4_post","problem":"pre","pred":0.0001,"meas":-0.0015,"origin":"codex island","status":"TRYING"},{"id":"dec-777300015","problem":"dec","pred":0.0,"meas":-0.1892,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-777300004","problem":"attn","pred":0.03,"meas":0.0731,"origin":"measurement / oracle / profiling","status":"WORKING"},{"id":"moe-50210","problem":"moe","pred":0.002,"meas":-0.0015,"origin":"codex island","status":"TRYING"},{"id":"attn-1346","problem":"attn","pred":0.03,"meas":-0.005,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-391600051","problem":"dec","pred":0.004,"meas":0.0055,"origin":"first principles / floor","status":"TRYING"},{"id":"bf16-229","problem":"bf16","pred":0.0269,"meas":-0.0056,"origin":"manager hypothesis generator","status":"TRYING"},{"id":"attn-463300003","problem":"attn","pred":0.01,"meas":0.013,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"bf16-50264","problem":"bf16","pred":0.0003,"meas":-0.0033,"origin":"codex island","status":"TRYING"},{"id":"pre-50376","problem":"pre","pred":0.0001,"meas":-0.0056,"origin":"codex island","status":"TRYING"},{"id":"dec-734800078","problem":"dec","pred":0.0015,"meas":0.0023,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2326","problem":"dec","pred":0.0015,"meas":-0.0021,"origin":"evolution search","status":"TRYING"},{"id":"dec-2419","problem":"dec","pred":0.005,"meas":0.0057,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2297","problem":"dec","pred":0.0005,"meas":0.0008,"origin":"combination engine","status":"TRYING"},{"id":"moe-1256","problem":"moe","pred":0.012,"meas":-0.001,"origin":"first principles / floor","status":"WORKING"},{"id":"moe-1077","problem":"moe","pred":0.007,"meas":0.0028,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-044","problem":"pre","pred":0.02,"meas":0.0388,"origin":"alpha (go one level lower)","status":"SHIPPED"},{"id":"dec-779600035","problem":"dec","pred":0.005,"meas":-0.0006,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-50000051","problem":"dec","pred":0.005,"meas":0.0085,"origin":"(empty)","status":"TRYING"},{"id":"bf16-50315","problem":"bf16","pred":0.0002,"meas":-0.0019,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-164924-rank_l1first_post","problem":"pre","pred":0.0,"meas":-0.0022,"origin":"codex island","status":"TRYING"},{"id":"moe-113","problem":"moe","pred":0.0,"meas":0.0,"origin":"codex island","status":"TRYING"},{"id":"dec-429000048","problem":"dec","pred":0.0,"meas":0.1713,"origin":"(empty)","status":"WORKING"},{"id":"dec-772200021","problem":"dec","pred":0.0001,"meas":0.0016,"origin":"codex island","status":"TRYING"},{"id":"attn-927","problem":"attn","pred":0.008,"meas":0.0085,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50082","problem":"dec","pred":0.012,"meas":0.0109,"origin":"lane's own work (origin not stated)","status":"OPTIMIZING"},{"id":"attn-414000077","problem":"attn","pred":0.006,"meas":0.0141,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"pre-142","problem":"pre","pred":0.012,"meas":0.0101,"origin":"codex island","status":"SHIPPED"},{"id":"attn-803","problem":"attn","pred":0.003,"meas":-0.0257,"origin":"web / blogs / deep research","status":"TRYING"},{"id":"pre-607","problem":"pre","pred":0.006,"meas":0.0007,"origin":"codex island","status":"TRYING"},{"id":"attn-867700032","problem":"attn","pred":0.01,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-777300017","problem":"dec","pred":0.2,"meas":-0.007,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-650","problem":"pre","pred":0.009,"meas":0.004,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-717600015","problem":"dec","pred":0.002,"meas":-0.0001,"origin":"port / amplify a win","status":"TRYING"},{"id":"attn-50426","problem":"attn","pred":0.03,"meas":-0.3134,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-589300015","problem":"dec","pred":0.002,"meas":-0.0059,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50335","problem":"bf16","pred":0.0001,"meas":-0.0013,"origin":"knob sweep","status":"TRYING"},{"id":"dec-800400075","problem":"dec","pred":0.03,"meas":-0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-640800015","problem":"dec","pred":0.001,"meas":0.0003,"origin":"(empty)","status":"TRYING"},{"id":"bf16-445","problem":"bf16","pred":0.008,"meas":0.0117,"origin":"plan files / owner","status":"OPTIMIZING"},{"id":"moe-120","problem":"moe","pred":0.01,"meas":-0.0127,"origin":"codex island","status":"KILLED"},{"id":"attn-50000002","problem":"attn","pred":0.01,"meas":-0.0072,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2980","problem":"dec","pred":0.002,"meas":0.0017,"origin":"learned priors","status":"TRYING"},{"id":"moe-414000043","problem":"moe","pred":0.0,"meas":-0.0733,"origin":"(empty)","status":"TRYING"},{"id":"dec-2790","problem":"dec","pred":0.0004,"meas":0.0018,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50291","problem":"pre","pred":0.0002,"meas":0.0061,"origin":"codex island","status":"TRYING"},{"id":"attn-045","problem":"attn","pred":0.05,"meas":0.0858,"origin":"alpha (go one level lower)","status":"SHIPPED"},{"id":"dec-2753","problem":"dec","pred":0.0015,"meas":0.0014,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-599200008","problem":"dec","pred":0.4,"meas":0.0006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-463300001","problem":"dec","pred":0.003,"meas":0.0084,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1760","problem":"attn","pred":0.004,"meas":0.0,"origin":"(empty)","status":"TRYING"},{"id":"dec-867700009","problem":"dec","pred":0.0027,"meas":0.0027,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50265","problem":"bf16","pred":0.001,"meas":-0.0175,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-339","problem":"moe","pred":0.02,"meas":-0.0829,"origin":"codex island","status":"KILLED"},{"id":"dec-800400052","problem":"dec","pred":0.03,"meas":-0.0021,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2733","problem":"dec","pred":0.0045,"meas":0.0036,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-599200021","problem":"dec","pred":0.0,"meas":0.0004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500072","problem":"dec","pred":0.0,"meas":-0.0005,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50318","problem":"pre","pred":0.0001,"meas":-0.0023,"origin":"codex island","status":"TRYING"},{"id":"attn-1729","problem":"attn","pred":0.001,"meas":-0.0002,"origin":"combination engine","status":"TRYING"},{"id":"dec-548000013","problem":"dec","pred":0.0,"meas":0.0209,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-257","problem":"moe","pred":0.015,"meas":0.0001,"origin":"codex island","status":"KILLED"},{"id":"attn-1322","problem":"attn","pred":0.003,"meas":-0.0003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-888100007","problem":"dec","pred":0.008,"meas":-0.0418,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50313","problem":"pre","pred":0.0001,"meas":-0.0023,"origin":"codex island","status":"TRYING"},{"id":"pre-786","problem":"pre","pred":0.06,"meas":0.0439,"origin":"combination engine","status":"TRYING"},{"id":"bf16-50332","problem":"bf16","pred":0.001,"meas":-0.0007,"origin":"knob sweep","status":"TRYING"},{"id":"dec-50000037","problem":"dec","pred":0.001,"meas":-0.1251,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1924","problem":"dec","pred":0.006,"meas":0.0019,"origin":"combination engine","status":"TRYING"},{"id":"dec-468500060","problem":"dec","pred":0.0,"meas":0.0047,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-112510-tw_cs","problem":"pre","pred":0.0001,"meas":-0.0017,"origin":"codex island","status":"TRYING"},{"id":"attn-50104","problem":"attn","pred":0.003,"meas":0.0048,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50253","problem":"bf16","pred":0.001,"meas":-0.0093,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-052","problem":"dec","pred":0.01,"meas":0.0011,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-144713-post_keep_ids","problem":"pre","pred":0.0001,"meas":-0.0017,"origin":"codex island","status":"TRYING"},{"id":"attn-463200005","problem":"attn","pred":0.0,"meas":-0.0136,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-249","problem":"moe","pred":0.015,"meas":-0.0032,"origin":"codex island","status":"KILLED"},{"id":"dec-779600059","problem":"dec","pred":0.001,"meas":0.0115,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2778","problem":"dec","pred":0.001,"meas":0.0006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-069","problem":"pre","pred":0.01,"meas":0.0013,"origin":"codex island","status":"KILLED"},{"id":"dec-762300001","problem":"dec","pred":0.001,"meas":-0.0122,"origin":"(empty)","status":"TRYING"},{"id":"moe-269","problem":"moe","pred":0.0085,"meas":0.0168,"origin":"codex island","status":"TRYING"},{"id":"dec-548000021","problem":"dec","pred":0.003,"meas":-0.01,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-50514","problem":"attn","pred":0.4,"meas":-0.0105,"origin":"first principles / floor","status":"TRYING"},{"id":"moe-191","problem":"moe","pred":0.01,"meas":0.0322,"origin":"codex island","status":"TRYING"},{"id":"dec-548000047","problem":"dec","pred":0.003,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1378","problem":"attn","pred":0.011,"meas":0.0114,"origin":"combination engine","status":"WORKING"},{"id":"pre-50344","problem":"pre","pred":0.0002,"meas":-0.0028,"origin":"codex island","status":"TRYING"},{"id":"attn-414000126","problem":"attn","pred":0.001,"meas":-0.0045,"origin":"(empty)","status":"TRYING"},{"id":"attn-468500015","problem":"attn","pred":0.008,"meas":0.013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-114822-score_uor","problem":"pre","pred":0.0001,"meas":-0.0037,"origin":"codex island","status":"TRYING"},{"id":"moe-1658","problem":"moe","pred":0.0,"meas":0.001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-370","problem":"pre","pred":0.006,"meas":0.001,"origin":"web / blogs / deep research","status":"TRYING"},{"id":"dec-1870","problem":"dec","pred":0.003,"meas":-0.0001,"origin":"codex island","status":"KILLED"},{"id":"bf16-637800016","problem":"bf16","pred":0.002,"meas":0.0017,"origin":"knob sweep","status":"TRYING"},{"id":"dec-468500009","problem":"dec","pred":0.002,"meas":0.0002,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-463200015","problem":"dec","pred":0.003,"meas":0.0292,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50000054","problem":"dec","pred":0.004,"meas":0.0032,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50577","problem":"attn","pred":0.0005,"meas":-0.0007,"origin":"codex island","status":"TRYING"},{"id":"dec-2625","problem":"dec","pred":0.002,"meas":0.0022,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50318","problem":"bf16","pred":0.0002,"meas":-0.0023,"origin":"codex island","status":"TRYING"},{"id":"attn-765700007","problem":"attn","pred":0.03,"meas":-0.0,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"bf16-709","problem":"bf16","pred":0.03,"meas":0.044,"origin":"rival / peer code","status":"WORKING"},{"id":"dec-429000027","problem":"dec","pred":0.0015,"meas":-0.0062,"origin":"(empty)","status":"TRYING"},{"id":"bf16-734800004","problem":"bf16","pred":0.0001,"meas":0.0295,"origin":"knob sweep","status":"TRYING"},{"id":"dec-2183","problem":"dec","pred":0.001,"meas":-0.0039,"origin":"codex island","status":"TRYING"},{"id":"attn-1807","problem":"attn","pred":0.012,"meas":-0.1349,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"pre-717600001","problem":"pre","pred":0.0,"meas":0.0,"origin":"(empty)","status":"TRYING"},{"id":"moe-496500109","problem":"moe","pred":0.0002,"meas":0.0033,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-173","problem":"moe","pred":0.02,"meas":0.0353,"origin":"codex island","status":"OPTIMIZING"},{"id":"attn-746","problem":"attn","pred":0.012,"meas":0.0103,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"bf16-734800002","problem":"bf16","pred":0.0001,"meas":-0.0027,"origin":"knob sweep","status":"TRYING"},{"id":"dec-772200008","problem":"dec","pred":0.0002,"meas":-0.0024,"origin":"codex island","status":"TRYING"},{"id":"pre-755","problem":"pre","pred":0.004,"meas":-0.1912,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-008","problem":"pre","pred":0.33,"meas":0.3355,"origin":"papers (arXiv)","status":"SHIPPED"},{"id":"moe-496500119","problem":"moe","pred":0.001,"meas":-0.0263,"origin":"(empty)","status":"TRYING"},{"id":"dec-1567","problem":"dec","pred":0.004,"meas":0.0118,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-589300001","problem":"attn","pred":0.02,"meas":0.0209,"origin":"lane's own work (origin not stated)","status":"OPTIMIZING"},{"id":"dec-468500008","problem":"dec","pred":0.0044,"meas":0.0055,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-347","problem":"moe","pred":0.004,"meas":0.0021,"origin":"codex island","status":"TRYING"},{"id":"pre-50360","problem":"pre","pred":0.0001,"meas":-0.0089,"origin":"codex island","status":"TRYING"},{"id":"bf16-603","problem":"bf16","pred":0.008,"meas":-0.0208,"origin":"codex island","status":"KILLED"},{"id":"bf16-50119","problem":"bf16","pred":0.035,"meas":-0.1025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-660","problem":"pre","pred":0.004,"meas":0.0049,"origin":"codex island","status":"TRYING"},{"id":"dec-463200002","problem":"dec","pred":0.03,"meas":0.0007,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50082","problem":"bf16","pred":0.008,"meas":-0.0217,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500012","problem":"dec","pred":0.03,"meas":0.1004,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-673600047","problem":"dec","pred":0.0,"meas":0.0,"origin":"(empty)","status":"TRYING"},{"id":"moe-470500066","problem":"moe","pred":0.0001,"meas":-0.0024,"origin":"codex island","status":"TRYING"},{"id":"dec-1414","problem":"dec","pred":0.0,"meas":-0.001,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-218000005","problem":"moe","pred":0.0002,"meas":-0.0033,"origin":"codex island","status":"TRYING"},{"id":"attn-828","problem":"attn","pred":0.008,"meas":0.0101,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"bf16-140","problem":"bf16","pred":0.01,"meas":0.0016,"origin":"codex island","status":"TRYING"},{"id":"attn-777300015","problem":"attn","pred":0.03,"meas":0.0063,"origin":"(empty)","status":"TRYING"},{"id":"dec-429000026","problem":"dec","pred":0.003,"meas":0.0037,"origin":"(empty)","status":"TRYING"},{"id":"moe-717600014","problem":"moe","pred":0.0,"meas":0.0019,"origin":"(empty)","status":"TRYING"},{"id":"bf16-105","problem":"bf16","pred":0.004,"meas":0.0057,"origin":"codex island","status":"TRYING"},{"id":"moe-155","problem":"moe","pred":0.02,"meas":-0.0352,"origin":"codex island","status":"KILLED"},{"id":"dec-496500111","problem":"dec","pred":0.001,"meas":-0.0079,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-888100052","problem":"dec","pred":0.002,"meas":0.0041,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-151021-ranks_v4_mid","problem":"pre","pred":0.0001,"meas":-0.0009,"origin":"codex island","status":"TRYING"},{"id":"pre-50261","problem":"pre","pred":0.0003,"meas":-0.0031,"origin":"codex island","status":"TRYING"},{"id":"attn-858","problem":"attn","pred":0.0,"meas":-0.0032,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000039","problem":"dec","pred":0.002,"meas":-0.0005,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-1992","problem":"dec","pred":0.006,"meas":0.0024,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"dec-765700001","problem":"dec","pred":0.001,"meas":-0.0011,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-50209","problem":"dec","pred":0.004,"meas":-0.0013,"origin":"combination engine","status":"TRYING"},{"id":"attn-463200015","problem":"attn","pred":0.01,"meas":0.0183,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2364","problem":"dec","pred":0.01,"meas":0.0055,"origin":"(empty)","status":"TRYING"},{"id":"dec-50000012","problem":"dec","pred":0.03,"meas":-0.0006,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-688","problem":"pre","pred":0.0003,"meas":0.0003,"origin":"codex island","status":"TRYING"},{"id":"moe-417900002","problem":"moe","pred":0.03,"meas":-0.0142,"origin":"(empty)","status":"TRYING"},{"id":"dec-772200015","problem":"dec","pred":0.0001,"meas":0.0023,"origin":"codex island","status":"TRYING"},{"id":"dec-599200022","problem":"dec","pred":0.0,"meas":-0.0264,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50299","problem":"pre","pred":0.0003,"meas":0.0024,"origin":"codex island","status":"TRYING"},{"id":"attn-50551","problem":"attn","pred":0.001,"meas":0.0172,"origin":"codex island","status":"WORKING"},{"id":"dec-414000057","problem":"dec","pred":0.001,"meas":-0.0011,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-138","problem":"dec","pred":0.012,"meas":0.0042,"origin":"codex island","status":"TRYING"},{"id":"dec-2165","problem":"dec","pred":0.0005,"meas":-0.0007,"origin":"codex island","status":"TRYING"},{"id":"dec-590000003","problem":"dec","pred":0.03,"meas":0.0143,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-414000016","problem":"pre","pred":0.001,"meas":0.0045,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-414000093","problem":"attn","pred":0.0,"meas":-0.0012,"origin":"(empty)","status":"TRYING"},{"id":"pre-340","problem":"pre","pred":0.008,"meas":0.0103,"origin":"first principles / floor","status":"WORKING"},{"id":"dec-2362","problem":"dec","pred":0.001,"meas":-0.0018,"origin":"combination engine","status":"TRYING"},{"id":"dec-779600066","problem":"dec","pred":0.03,"meas":-0.0539,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-468500002","problem":"pre","pred":0.003,"meas":0.0029,"origin":"(empty)","status":"TRYING"},{"id":"bf16-186","problem":"bf16","pred":0.004,"meas":-0.0025,"origin":"codex island","status":"KILLED"},{"id":"dec-800400025","problem":"dec","pred":0.002,"meas":-0.0095,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"pre-702","problem":"pre","pred":0.0003,"meas":0.0005,"origin":"codex island","status":"TRYING"},{"id":"bf16-125","problem":"bf16","pred":0.013,"meas":0.0142,"origin":"codex island","status":"WORKING"},{"id":"dec-2089","problem":"dec","pred":0.0005,"meas":-0.0022,"origin":"evolution search","status":"TRYING"},{"id":"pre-717600002","problem":"pre","pred":0.01,"meas":-0.321,"origin":"(empty)","status":"TRYING"},{"id":"dec-468500041","problem":"dec","pred":0.0,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-1919","problem":"attn","pred":0.004,"meas":0.0035,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-714500016","problem":"dec","pred":0.004,"meas":-0.2894,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"pre-sweeper-194641-prank_L1last_mid","problem":"pre","pred":0.0,"meas":-0.002,"origin":"codex island","status":"KILLED"},{"id":"dec-548000070","problem":"dec","pred":0.03,"meas":0.0028,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-888100051","problem":"dec","pred":0.0001,"meas":0.0884,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000103","problem":"dec","pred":0.002,"meas":-0.0018,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50359","problem":"pre","pred":0.0001,"meas":-0.0026,"origin":"codex island","status":"TRYING"},{"id":"dec-2011","problem":"dec","pred":0.0,"meas":-0.0188,"origin":"codex island","status":"TRYING"},{"id":"dec-463200003","problem":"dec","pred":0.03,"meas":-0.0108,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-50114","problem":"attn","pred":0.008,"meas":0.0214,"origin":"first principles / floor","status":"OPTIMIZING"},{"id":"bf16-545","problem":"bf16","pred":0.009,"meas":-0.0526,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-811","problem":"pre","pred":0.0012,"meas":-0.0007,"origin":"port / amplify a win","status":"TRYING"},{"id":"attn-1895","problem":"attn","pred":0.004,"meas":0.0059,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-765700005","problem":"moe","pred":0.0,"meas":-0.1237,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600023","problem":"dec","pred":0.004,"meas":-0.3692,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-114822-score_uadd","problem":"pre","pred":0.0001,"meas":-0.0031,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-160104-guard_v2_post","problem":"pre","pred":0.0001,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"attn-463300020","problem":"attn","pred":0.03,"meas":-0.031,"origin":"(empty)","status":"TRYING"},{"id":"dec-779600023","problem":"dec","pred":0.0,"meas":0.0214,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50331","problem":"pre","pred":0.0002,"meas":0.0031,"origin":"codex island","status":"TRYING"},{"id":"dec-391600064","problem":"dec","pred":0.002,"meas":0.0245,"origin":"(empty)","status":"WORKING"},{"id":"dec-50000018","problem":"dec","pred":0.0001,"meas":-0.0038,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700036","problem":"dec","pred":0.03,"meas":0.0067,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-50085","problem":"bf16","pred":0.01,"meas":0.0024,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-391600021","problem":"dec","pred":0.003,"meas":0.0033,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-548000022","problem":"attn","pred":0.0,"meas":-0.0389,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-187","problem":"moe","pred":0.01,"meas":0.0617,"origin":"codex island","status":"WORKING"},{"id":"attn-589300002","problem":"attn","pred":0.001,"meas":0.0012,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1516","problem":"dec","pred":0.008,"meas":0.0073,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50391","problem":"pre","pred":0.0015,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"attn-031","problem":"attn","pred":0.05,"meas":-0.5585,"origin":"codex island","status":"KILLED"},{"id":"dec-679700009","problem":"dec","pred":0.002,"meas":0.0016,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-2121","problem":"dec","pred":0.0013,"meas":-0.0002,"origin":"codex island","status":"TRYING"},{"id":"pre-450","problem":"pre","pred":0.01,"meas":-0.0014,"origin":"codex island","status":"TRYING"},{"id":"dec-296","problem":"dec","pred":0.004,"meas":-0.0,"origin":"codex island","status":"TRYING"},{"id":"moe-470500056","problem":"moe","pred":0.0002,"meas":-0.0036,"origin":"codex island","status":"TRYING"},{"id":"attn-414000118","problem":"attn","pred":0.003,"meas":-0.019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-1093","problem":"attn","pred":0.008,"meas":0.0239,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"attn-50168","problem":"attn","pred":0.015,"meas":-0.025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-286","problem":"moe","pred":0.004,"meas":0.0015,"origin":"codex island","status":"KILLED"},{"id":"moe-1174","problem":"moe","pred":0.012,"meas":0.0072,"origin":"combination engine","status":"TRYING"},{"id":"attn-714","problem":"attn","pred":0.015,"meas":0.0182,"origin":"lane's own work (origin not stated)","status":"WORKING"},{"id":"bf16-470500028","problem":"bf16","pred":0.001,"meas":-0.0571,"origin":"(empty)","status":"TRYING"},{"id":"dec-888100005","problem":"dec","pred":0.03,"meas":-0.0059,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-637800003","problem":"dec","pred":0.0005,"meas":-0.011,"origin":"knob sweep","status":"TRYING"},{"id":"dec-2197","problem":"dec","pred":0.0003,"meas":-0.0003,"origin":"codex island","status":"TRYING"},{"id":"dec-548000058","problem":"dec","pred":0.03,"meas":-0.0048,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50253","problem":"pre","pred":0.0002,"meas":-0.0058,"origin":"codex island","status":"TRYING"},{"id":"dec-048","problem":"dec","pred":0.03,"meas":-0.0187,"origin":"codex island","status":"KILLED"},{"id":"pre-414000005","problem":"pre","pred":0.002,"meas":0.0005,"origin":"(empty)","status":"TRYING"},{"id":"pre-50242","problem":"pre","pred":0.0001,"meas":-0.0006,"origin":"codex island","status":"TRYING"},{"id":"moe-15300032","problem":"moe","pred":0.03,"meas":0.0606,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50326","problem":"pre","pred":0.0001,"meas":-0.0011,"origin":"codex island","status":"TRYING"},{"id":"bf16-50328","problem":"bf16","pred":0.002,"meas":0.0057,"origin":"knob sweep","status":"TRYING"},{"id":"dec-2806","problem":"dec","pred":0.0005,"meas":0.002,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-599200026","problem":"dec","pred":0.0,"meas":-0.0109,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-836","problem":"attn","pred":0.006,"meas":-0.0332,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-023","problem":"attn","pred":0.03,"meas":-0.0712,"origin":"books","status":"KILLED"},{"id":"moe-1153","problem":"moe","pred":0.01,"meas":0.0053,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-599200002","problem":"dec","pred":0.003,"meas":0.0011,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-429000051","problem":"dec","pred":0.0015,"meas":-0.2731,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-637800004","problem":"moe","pred":0.0002,"meas":-0.001,"origin":"knob sweep","status":"TRYING"},{"id":"dec-2476","problem":"dec","pred":0.004,"meas":0.003,"origin":"combination engine","status":"TRYING"},{"id":"dec-734800007","problem":"dec","pred":0.0005,"meas":-0.0002,"origin":"knob sweep","status":"TRYING"},{"id":"pre-50107","problem":"pre","pred":0.005,"meas":0.0003,"origin":"first principles / floor","status":"WORKING"},{"id":"moe-051","problem":"moe","pred":0.067,"meas":-0.0625,"origin":"codex island","status":"KILLED"},{"id":"moe-463200009","problem":"moe","pred":0.004,"meas":0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-50031","problem":"moe","pred":0.015,"meas":0.0057,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1781","problem":"dec","pred":0.005,"meas":-0.0052,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-sweeper-142648-scatter_plan_meta_last","problem":"pre","pred":0.0001,"meas":-0.001,"origin":"codex island","status":"TRYING"},{"id":"dec-496500112","problem":"dec","pred":0.0022,"meas":-0.0151,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-072","problem":"moe","pred":0.02,"meas":0.044,"origin":"codex island","status":"WORKING"},{"id":"attn-590000007","problem":"attn","pred":0.03,"meas":0.0048,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-3084","problem":"dec","pred":0.001,"meas":0.0006,"origin":"learned priors","status":"TRYING"},{"id":"attn-1719","problem":"attn","pred":0.001,"meas":-0.0116,"origin":"codex island","status":"TRYING"},{"id":"pre-443","problem":"pre","pred":0.01,"meas":0.0004,"origin":"codex island","status":"TRYING"},{"id":"attn-517600009","problem":"attn","pred":0.003,"meas":-0.1274,"origin":"(empty)","status":"TRYING"},{"id":"dec-50000047","problem":"dec","pred":0.005,"meas":0.0002,"origin":"(empty)","status":"TRYING"},{"id":"dec-2296","problem":"dec","pred":0.0025,"meas":-0.0001,"origin":"evolution search","status":"TRYING"},{"id":"pre-638","problem":"pre","pred":0.003,"meas":-0.0012,"origin":"codex island","status":"TRYING"},{"id":"pre-679700002","problem":"pre","pred":0.0,"meas":-0.0091,"origin":"(empty)","status":"TRYING"},{"id":"dec-3035","problem":"dec","pred":0.0004,"meas":0.0003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-144713-scatter_mid_ranks_early","problem":"pre","pred":0.0001,"meas":-0.0049,"origin":"codex island","status":"TRYING"},{"id":"pre-639","problem":"pre","pred":0.01,"meas":0.0016,"origin":"codex island","status":"TRYING"},{"id":"dec-429000018","problem":"dec","pred":0.0,"meas":-0.0668,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-468500025","problem":"dec","pred":0.004,"meas":-0.7143,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500007","problem":"dec","pred":0.004,"meas":0.0719,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-414000079","problem":"moe","pred":0.003,"meas":-0.0042,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-2366","problem":"dec","pred":0.0003,"meas":-0.0121,"origin":"(empty)","status":"TRYING"},{"id":"attn-867700028","problem":"attn","pred":0.003,"meas":-0.0117,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-192","problem":"dec","pred":0.024,"meas":-0.0638,"origin":"plan files / owner","status":"TRYING"},{"id":"bf16-637800033","problem":"bf16","pred":0.002,"meas":-0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-441","problem":"bf16","pred":0.006,"meas":-0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-544","problem":"dec","pred":0.015,"meas":0.0165,"origin":"port / amplify a win","status":"OPTIMIZING"},{"id":"pre-626","problem":"pre","pred":0.35,"meas":-0.0054,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-414000035","problem":"attn","pred":0.0015,"meas":0.007,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-131","problem":"dec","pred":0.018,"meas":0.0046,"origin":"codex island","status":"KILLED"},{"id":"moe-496500112","problem":"moe","pred":0.0002,"meas":0.0042,"origin":"codex island","status":"TRYING"},{"id":"moe-1373","problem":"moe","pred":0.003,"meas":0.0012,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50321","problem":"pre","pred":0.0001,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"attn-1286","problem":"attn","pred":0.02,"meas":-0.0636,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-414000062","problem":"dec","pred":0.0005,"meas":0.0002,"origin":"(empty)","status":"TRYING"},{"id":"dec-2567","problem":"dec","pred":0.0008,"meas":-0.0025,"origin":"(empty)","status":"TRYING"},{"id":"dec-1517","problem":"dec","pred":0.017,"meas":0.0138,"origin":"alpha (go one level lower)","status":"WORKING"},{"id":"attn-590000004","problem":"attn","pred":0.03,"meas":-0.013,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-470500025","problem":"attn","pred":0.0002,"meas":0.0031,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-184136-pexp_cs_all","problem":"pre","pred":0.0,"meas":0.0005,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-194641-pexp_guardL1last_long","problem":"pre","pred":0.0,"meas":0.0001,"origin":"codex island","status":"TRYING"},{"id":"pre-50089","problem":"pre","pred":0.0,"meas":0.002,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-468500057","problem":"dec","pred":0.0,"meas":-0.0172,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-429000058","problem":"dec","pred":0.001,"meas":0.0049,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-867700058","problem":"dec","pred":0.0,"meas":-0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000052","problem":"dec","pred":0.002,"meas":-0.0014,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"all-016","problem":"all","pred":0.21,"meas":0.487,"origin":"plan files / owner","status":"WORKING"},{"id":"pre-657","problem":"pre","pred":0.006,"meas":-0.0006,"origin":"codex island","status":"TRYING"},{"id":"attn-021","problem":"attn","pred":0.02,"meas":0.0055,"origin":"books","status":"KILLED"},{"id":"pre-sweeper-185950-pexp_wt_long","problem":"pre","pred":0.0,"meas":0.001,"origin":"codex island","status":"TRYING"},{"id":"attn-752","problem":"attn","pred":0.012,"meas":0.0062,"origin":"port / amplify a win","status":"TRYING"},{"id":"dec-391600019","problem":"dec","pred":0.002,"meas":-0.0024,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-468500051","problem":"dec","pred":0.001,"meas":0.0015,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-1803","problem":"dec","pred":0.015,"meas":0.0023,"origin":"alpha (go one level lower)","status":"TRYING"},{"id":"dec-1756","problem":"dec","pred":0.01,"meas":0.0061,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-640800001","problem":"dec","pred":0.001,"meas":0.0019,"origin":"rival / peer code","status":"TRYING"},{"id":"pre-sweeper-194641-prank_guardL1last_long","problem":"pre","pred":0.0,"meas":-0.0002,"origin":"codex island","status":"KILLED"},{"id":"pre-50323","problem":"pre","pred":0.0001,"meas":0.0002,"origin":"codex island","status":"TRYING"},{"id":"pre-50305","problem":"pre","pred":0.0001,"meas":0.0002,"origin":"codex island","status":"TRYING"},{"id":"pre-50385","problem":"pre","pred":0.0001,"meas":-0.0007,"origin":"codex island","status":"TRYING"},{"id":"dec-317200004","problem":"dec","pred":0.0001,"meas":0.0003,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-125844-reset_b","problem":"pre","pred":0.0001,"meas":-0.0016,"origin":"codex island","status":"TRYING"},{"id":"dec-429000015","problem":"dec","pred":0.3,"meas":-0.0061,"origin":"(empty)","status":"TRYING"},{"id":"dec-429000008","problem":"dec","pred":0.0013,"meas":0.0028,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-496500120","problem":"moe","pred":0.001,"meas":-0.0278,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-1728","problem":"attn","pred":0.002,"meas":-0.008,"origin":"combination engine","status":"TRYING"},{"id":"pre-50172","problem":"pre","pred":0.03,"meas":-0.0037,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"bf16-50124","problem":"bf16","pred":0.035,"meas":-0.0077,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-717600002","problem":"dec","pred":0.0019,"meas":0.0011,"origin":"combination engine","status":"TRYING"},{"id":"dec-50180","problem":"dec","pred":0.0,"meas":-0.0006,"origin":"codex island","status":"KILLED"},{"id":"moe-47800004","problem":"moe","pred":0.0,"meas":-0.0305,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"attn-509400011","problem":"attn","pred":0.002,"meas":-0.0099,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"pre-414000031","problem":"pre","pred":0.004,"meas":0.0064,"origin":"(empty)","status":"TRYING"},{"id":"dec-50000050","problem":"dec","pred":0.005,"meas":-0.0002,"origin":"(empty)","status":"TRYING"},{"id":"moe-47800003","problem":"moe","pred":0.001,"meas":-0.0373,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-734800005","problem":"dec","pred":0.0005,"meas":-0.0165,"origin":"knob sweep","status":"TRYING"},{"id":"attn-044","problem":"attn","pred":0.02,"meas":0.0757,"origin":"codex island","status":"OPTIMIZING"},{"id":"pre-598","problem":"pre","pred":0.012,"meas":0.0005,"origin":"combination engine","status":"TRYING"},{"id":"moe-1482","problem":"moe","pred":0.004,"meas":0.0,"origin":"codex island","status":"TRYING"},{"id":"attn-463300006","problem":"attn","pred":0.002,"meas":-0.0298,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-800400074","problem":"dec","pred":0.03,"meas":-0.0003,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-286","problem":"dec","pred":0.003,"meas":0.0041,"origin":"codex island","status":"KILLED"},{"id":"moe-080","problem":"moe","pred":0.02,"meas":0.0538,"origin":"codex island","status":"WORKING"},{"id":"moe-765700001","problem":"moe","pred":0.003,"meas":-0.0135,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50249","problem":"pre","pred":0.002,"meas":-0.0075,"origin":"codex island","status":"TRYING"},{"id":"bf16-637800001","problem":"bf16","pred":0.002,"meas":-0.0076,"origin":"knob sweep","status":"TRYING"},{"id":"bf16-50104","problem":"bf16","pred":0.008,"meas":-0.0272,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-50091","problem":"pre","pred":0.008,"meas":0.001,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-679700139","problem":"dec","pred":0.0004,"meas":-0.0041,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-107","problem":"attn","pred":0.01,"meas":0.065,"origin":"codex island","status":"WORKING"},{"id":"pre-297","problem":"pre","pred":0.015,"meas":0.0186,"origin":"alpha (go one level lower)","status":"OPTIMIZING"},{"id":"pre-sweeper-151021-ranks_v4_plan","problem":"pre","pred":0.0001,"meas":-0.0019,"origin":"codex island","status":"TRYING"},{"id":"moe-241","problem":"moe","pred":0.013,"meas":-0.0412,"origin":"codex island","status":"KILLED"},{"id":"dec-590000011","problem":"dec","pred":0.03,"meas":-0.025,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-414000099","problem":"dec","pred":0.0005,"meas":-0.0092,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-673600062","problem":"dec","pred":0.0,"meas":0.0041,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"attn-717600008","problem":"attn","pred":0.002,"meas":-0.0228,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-548000059","problem":"dec","pred":0.03,"meas":-0.0013,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"dec-470500142","problem":"dec","pred":0.001,"meas":-0.0036,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"pre-738","problem":"pre","pred":0.0016,"meas":-0.0008,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"moe-1657","problem":"moe","pred":0.0005,"meas":0.0019,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-637800021","problem":"dec","pred":0.0005,"meas":0.0011,"origin":"knob sweep","status":"TRYING"},{"id":"moe-1682","problem":"moe","pred":0.002,"meas":0.0032,"origin":"evolution search","status":"TRYING"},{"id":"dec-867700025","problem":"dec","pred":0.005,"meas":0.0073,"origin":"(empty)","status":"TRYING"},{"id":"moe-637800010","problem":"moe","pred":0.0002,"meas":0.0045,"origin":"knob sweep","status":"TRYING"},{"id":"attn-1281","problem":"attn","pred":0.01,"meas":0.0172,"origin":"first principles / floor","status":"WORKING"},{"id":"bf16-637800025","problem":"bf16","pred":0.001,"meas":-0.0029,"origin":"knob sweep","status":"TRYING"},{"id":"moe-274","problem":"moe","pred":0.005,"meas":-0.0027,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-185950-prank_wt_all","problem":"pre","pred":0.0,"meas":-0.0036,"origin":"codex island","status":"TRYING"},{"id":"attn-1771","problem":"attn","pred":0.009,"meas":0.0095,"origin":"evolution search","status":"TRYING"},{"id":"bf16-508","problem":"bf16","pred":0.004,"meas":0.0001,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-517600018","problem":"bf16","pred":0.0015,"meas":-0.0826,"origin":"rival / peer code","status":"TRYING"},{"id":"dec-673600028","problem":"dec","pred":0.001,"meas":0.0,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"dec-300","problem":"dec","pred":0.008,"meas":0.0029,"origin":"measurement / oracle / profiling","status":"KILLED"},{"id":"attn-50568","problem":"attn","pred":0.0003,"meas":-0.0069,"origin":"codex island","status":"TRYING"},{"id":"dec-391600089","problem":"dec","pred":0.002,"meas":-0.0427,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-762300006","problem":"dec","pred":0.002,"meas":0.0022,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"bf16-637800004","problem":"bf16","pred":0.002,"meas":-0.0035,"origin":"knob sweep","status":"TRYING"},{"id":"attn-749100003","problem":"attn","pred":0.03,"meas":-0.0007,"origin":"measurement / oracle / profiling","status":"TRYING"},{"id":"moe-1536","problem":"moe","pred":0.005,"meas":0.003,"origin":"combination engine","status":"TRYING"},{"id":"dec-121","problem":"dec","pred":0.006,"meas":0.0123,"origin":"codex island","status":"WORKING"},{"id":"dec-2762","problem":"dec","pred":0.004,"meas":-0.027,"origin":"first principles / floor","status":"TRYING"},{"id":"attn-414000065","problem":"attn","pred":0.017,"meas":0.0125,"origin":"(empty)","status":"OPTIMIZING"},{"id":"dec-800400049","problem":"dec","pred":0.003,"meas":-0.0148,"origin":"lane's own work (origin not stated)","status":"KILLED"},{"id":"attn-200","problem":"attn","pred":0.01,"meas":0.0313,"origin":"lane's own work (origin not stated)","status":"OPTIMIZING"},{"id":"moe-414000029","problem":"moe","pred":0.001,"meas":-0.0012,"origin":"(empty)","status":"KILLED"},{"id":"moe-50196","problem":"moe","pred":0.4,"meas":-0.0046,"origin":"first principles / floor","status":"TRYING"},{"id":"dec-2846","problem":"dec","pred":0.002,"meas":0.0007,"origin":"plan files / owner","status":"TRYING"},{"id":"moe-50252","problem":"moe","pred":0.0005,"meas":0.0001,"origin":"codex island","status":"TRYING"},{"id":"attn-548000005","problem":"attn","pred":0.0,"meas":-0.0265,"origin":"lane's own work (origin not stated)","status":"TRYING"},{"id":"pre-sweeper-185950-prank_wt_mid","problem":"pre","pred":0.0,"meas":-0.0001,"origin":"codex island","status":"TRYING"},{"id":"dec-123100020","problem":"dec","pred":0.0008,"meas":0.0026,"origin":"(empty)","status":"TRYING"},{"id":"pre-072","problem":"pre","pred":0.004,"meas":0.0029,"origin":"codex island","status":"TRYING"},{"id":"pre-sweeper-162014-guard_l1first_mid","problem":"pre","pred":0.0,"meas":-0.0024,"origin":"codex island","status":"TRYING"}]},"kill":{"kill_events":1018,"ideas_killed_explicitly":1001,"trials_after_explicit_kill":67,"full_hits_after_explicit_kill":0,"revive_events":9,"kill_reason_classes":{"duplicate / subsumed":135,"measured loss / no gain":562,"other":122,"correctness / legality":138,"premise does not bind / ceiling too small":61}},"status_now":{"SHIPPED":46,"KILLED":1263,"TRYING":7488,"NEW":5825,"OPTIMIZING":92,"WORKING":222},"paired":{"by_verdict":{"ABORTED":788,"INVALID":35,"CONFIRMED":130,"INCONCLUSIVE":371,"REFUTED":166,"JUDGE_CHECK":106,"COLD_SLOW":11,"?":5},"per_problem":{"bf16":{"ABORTED":84,"INVALID":1,"CONFIRMED":6,"INCONCLUSIVE":45,"JUDGE_CHECK":5,"REFUTED":19},"dec":{"CONFIRMED":70,"JUDGE_CHECK":50,"ABORTED":324,"INCONCLUSIVE":141,"COLD_SLOW":10,"REFUTED":23,"INVALID":22,"?":5},"moe":{"REFUTED":22,"INCONCLUSIVE":45,"ABORTED":66,"CONFIRMED":15,"JUDGE_CHECK":15,"INVALID":2},"attn":{"CONFIRMED":31,"JUDGE_CHECK":29,"ABORTED":258,"INCONCLUSIVE":116,"COLD_SLOW":1,"REFUTED":95,"INVALID":6},"pre":{"INCONCLUSIVE":24,"ABORTED":56,"CONFIRMED":8,"JUDGE_CHECK":7,"REFUTED":7,"INVALID":4}},"confirmed_of_decided":{"k":130,"n":667,"rate":0.1949,"ci90":[0.1709,0.2213]},"aborted_share":{"k":788,"n":1612,"rate":0.4888,"ci90":[0.4684,0.5093]},"src":"bench/manager/paired.jsonl (paired_confirm.py, 4 ABBA blocks, 95% t CI)"},"idea_to_trial_h":{"n":8931,"p25":0.0292,"median":0.1086,"p75":0.2578,"p90":0.7839,"within":{"<= 0.25 h":{"k":6630,"n":8931,"rate":0.7424,"ci90":[0.7347,0.7499]},"<= 1 h":{"k":8203,"n":8931,"rate":0.9185,"ci90":[0.9136,0.9231]},"<= 6 h":{"k":8652,"n":8931,"rate":0.9688,"ci90":[0.9656,0.9716]},"<= 24 h":{"k":8793,"n":8931,"rate":0.9845,"ci90":[0.9823,0.9866]}},"per_problem":{"dec":{"n":3499,"median_h":0.09},"attn":{"n":2036,"median_h":0.0967},"moe":{"n":1484,"median_h":0.136},"pre":{"n":976,"median_h":0.1283},"bf16":{"n":907,"median_h":0.1108},"all":{"n":29,"median_h":0.1803}}},"never_tried":{"p":0.402,"lo":0.3955,"hi":0.4087,"k":6005,"n":14936},"research":{"papers":337,"arxiv":215,"deep_reports":434,"blogs":214,"books":24,"interviews":46,"questions":80,"hunches":15,"hunches_with_win":13},"coordination":{"notes_by_type":{"trigger":445,"check":513,"comment":62482,"todo":9688},"forum":13687,"groups":23,"owner_asks":961,"owner_answered":409,"owner_median_h":null,"owner_days":["2026-10-23","2026-10-24","2026-10-25","2026-10-26","2026-10-27","2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-05"],"owner_asks_per_day":[18,4,23,3,75,476,55,59,416,22,19,9,1],"owner_answers_per_day":[11,0,0,2,0,0,0,1,396,0,0,0,0]},"components":[{"name":"Modal judge clone","layer":"measure","file":"bench/modal/judge.py","introduced":"2026-10-22T16:09:10-07:00","introduced_src":"git first-add 68e45985b25","date_note":null,"what":"Re-runs the organizers' scoring method (cold L2, CUDA graphs, paired SGLang ratios) on a Modal H200 for any file.","measured_effect":{"value":"Judge / clone ratio for the same bytes: decode x0.969, MoE x0.967, attention x0.959, bf16 x1.00, prefill x1.03; so a clone gain is not a judge gain","kind":"measured","src":"AGENTS.md:74-76"},"file_exists":true,"git_first_add":"2026-10-22T16:09:10-07:00 68e45985b25 bf16: E6 (L2 prefetch)"},{"name":"Clone and random-draw inputs","layer":"measure","file":"bench/tools/clone_inputs.py","introduced":"2026-10-23T00:44:52-07:00","introduced_src":"git first-add 8878520c43e","date_note":null,"what":"Gives labs the clone's exact per-case inputs plus >= 20 random routing draws, so a lab predicts the judge.","measured_effect":{"value":"Clone's 4 fixed draws read w2_wg48sl +0.83%; the random-draw predictor read +0.45%. Random-routing labs had misjudged wide MoE kernels by ~4%","kind":"measured","src":"bench/tools/clone_inputs.py:3-8; AGENTS.md:141"},"file_exists":true,"git_first_add":"2026-10-23T00:44:52-07:00 8878520c43e bench/tools/clone_inpu"},{"name":"Paired ABBA confirmation","layer":"measure","file":"bench/tools/paired_confirm.py","introduced":"2026-10-27T02:55:44-07:00","introduced_src":"git first-add 2e91ee068b1","date_note":"A one-off confirm script (confirm-decode.md, first-add 2026-10-26T16:18 2aad6607c4f) preceded the tool.","what":"Predeclared same-box A-B-B-A blocks with a 95% CI decide whether a candidate beats the current base.","measured_effect":{"value":"72 h to 10-29: 1,011 paired verdicts, 81 CONFIRMED, 544 (54%) ABORTED for infrastructure. Prototype run 10-26: decode composite C screened +0.289%, confirmed +0.222% (95% CI 0.180..0.264)","kind":"measured","src":"bench/books/POSTMORTEM_1029.md:14-19; bench/reports/evolution/2026-10-26-01a17a50/confirm-decode.md:3-5"},"file_exists":true,"git_first_add":"2026-10-27T02:55:44-07:00 2e91ee068b1 paired_confirm: predec"},{"name":"Scoped judge-exact A/B (jexact_scoped)","layer":"measure","file":"bench/tools/jexact_scoped.py","introduced":"2026-10-31T21:54:58-07:00","introduced_src":"git first-add 70b7db71f17","date_note":"First-add is a bulk 'keep: this disk's work, 5819 files' commit; the file may have existed on disk earlier. SPEC_V3 rule 9 made it the verdict tool on 11-01.","what":"Times only the judge cases a candidate touches; route_identity proves the other cases identical, so they count zero.","measured_effect":{"value":"Resolution ~0.02% vs whole-x noise 0.3-0.6%. Whole-x verdicts only saw wins >= ~1%: hit rate 31.8% below row 2.0, 0.5% at 2.2-2.3, 0% at >= 2.3","kind":"measured","src":"bench/agents/SPEC_V3.md:73-85"},"file_exists":true,"git_first_add":"2026-10-31T22:29:13-07:00 edb7c27b03a sync: main's content a"},{"name":"Internal judge","layer":"measure","file":"bench/tools/internal_judge.py","introduced":"2026-11-01T15:53:28-08:00","introduced_src":"git first-add ad3b38bf87d","date_note":"First-add is a push_sync squash commit ('main's content as one commit'). Docstring dates it 11-01.","what":"Scores decode chain candidates with the judge's own timing sequence; real judge reads become an external check only.","measured_effect":{"value":"Timing protocol vs 37 judge runs: bias -0.34%, MAE 0.66%, 86% of judge-ordered pairs ordered right","kind":"measured","src":"bench/tools/internal_judge.py:9-10 (cites judge_gap/LEARNINGS.md)"},"owner_quote":{"text":"we must have an internal source of truth and an external one, which we trust a bit less","src":"bench/agents/SPEC_V3.md:102"},"file_exists":true,"git_first_add":"2026-11-01T15:53:28-08:00 ad3b38bf87d sync: main's content a"},{"name":"Judge-calibrated screen","layer":"measure","file":"bench/tools/judge_cal.py","introduced":"2026-10-30T17:58:32-07:00","introduced_src":"git first-add 26ce4af52eb","date_note":null,"what":"Fits per-case box-to-judge factors and code-feature offsets so box screens rank files the way the judge does.","measured_effect":{"value":"Box vs judge rank agreement Spearman 0.28 -> 0.63 (leave-one-out); bias +3.6% -> -0.2%","kind":"measured","src":"git commit 26ce4af52eb message; bench/tools/judge_cal.py:4-5"},"owner_quote":{"text":"why is our decode layer so low? ... we are #36","src":"bench/tools/judge_cal.py:2"},"file_exists":true,"git_first_add":"2026-10-30T17:59:12-07:00 cd1141c0df1 sync: main's content a"},{"name":"Instruction budget (ISA)","layer":"measure","file":"bench/tools/isa_budget.py","introduced":"2026-10-23T13:47:50-07:00","introduced_src":"git first-add 4639753c54c","date_note":"First-add is a batch commit ('main 20:55: sub-goal workflow done').","what":"Derives the minimal instructions per pipe for a kernel and compares them with SASS counts to name the waste.","measured_effect":{"value":"First use: attention verify-full ran INT at 9x the needed count, latency-bound at 69% of the HBM floor","kind":"measured","src":"AGENTS.md:145"},"owner_quote":{"text":"decompose into the minimal set of operations and derive what instructions to run","src":"AGENTS.md:145 (owner's advisor, 10-23)"},"file_exists":true,"git_first_add":"2026-10-23T13:47:50-07:00 4639753c54c main 20:55: sub-goal w"},{"name":"Scientist ledger and kill rule","layer":"decide","file":"bench/tools/sci.py","introduced":"2026-10-22T17:35:59-07:00","introduced_src":"git first-add ff8c7643af5","date_note":null,"what":"Append-only ledger of every idea (predicted gain, p) and trial; replays state and retires ideas after 3 misses or 2 BAD.","measured_effect":{"value":"To 11-05: 14,667 ideas, 13,638 trials: 1,201 hit (8.8%), 6,179 miss, 2,936 flat, 2,316 error, 1,006 bad; 1,018 kill events, 9 revives","kind":"measured","src":"bench/sci/events.jsonl (read-only count of kind=add/trial/kill/revive, 2026-11-05)"},"owner_quote":{"text":"after misses kill ideas, move on; keep pushing the ones that start to work","src":"bench/tools/sci.py:7"},"file_exists":true,"git_first_add":"2026-10-22T17:35:59-07:00 ff8c7643af5 bench: auto-scientist "},{"name":"Racing screen and gene bank (race + keeps)","layer":"decide","file":"bench/tools/race.py","introduced":"2026-10-26T16:10:26-07:00","introduced_src":"git first-add 271ceb360eb (keeps.py; race.py 7ea0a869726 at 16:23)","date_note":null,"what":"Ranks small unpaired candidates for paired confirmation and keeps near-best files as genes to stack; no % cutoff.","measured_effect":{"value":"Gene bank at work: PF0RH (+0.009%) and STG (+1.1..1.4 us) were stacked into decode cand63","kind":"measured","src":"bench/agents/SPEC_V3.md:40-42"},"owner_quote":{"text":"below 1% actually starts to matter and we win by sheer scale of experiments ... weighted by the impact","src":"AGENTS.md:266-267"},"file_exists":true,"git_first_add":"2026-10-26T16:23:48-07:00 7ea0a869726 Racing replaces % thre"},{"name":"Gap to physical floor","layer":"decide","file":"bench/tools/gap.py","introduced":"2026-10-22T17:26:21-07:00","introduced_src":"git first-add 654e0e351cf","date_note":null,"what":"Prices every judge case against its physical floor (bytes / 4.8 TB/s, FLOPs / peak) and ranks work by score gain to floor.","measured_effect":{"value":"10-24 read: attention 58%, prefill 61%, decode 68%, MoE 74%, bf16 88% of physical ceiling. Recipe 'floor-first' won 11 of 157 clean tests (7%) vs 'compose' 130 of 812 (16%)","kind":"measured","src":"AGENTS.md:62; bench/WORKED.md:7,13"},"file_exists":true,"git_first_add":"2026-10-22T17:26:21-07:00 654e0e351cf bench: GAP.md ledger ("},{"name":"Candidate chain (one merge queue)","layer":"decide","file":"bench/tools/chain.py","introduced":"2026-10-30T14:41:18-07:00","introduced_src":"git first-add 022c5af0fdf","date_note":"First-add is a push_sync squash commit; the ledger's first row is 2026-10-30T22:37:12Z, consistent.","what":"One linear chain per problem: each landing is a diff confirmed paired against the CURRENT head, written by one merge queue.","measured_effect":{"value":"Decode: 115 merged landings 10-30 22:37Z to 11-05 05:03Z (cand0 x 1.168 -> cand126). Gain per landing on 10-31 fell +1.08 -> +0.59 -> +0.72 -> +0.31 -> +0.18 -> +0.12%","kind":"measured","src":"bench/chain/muse_decode_layer.jsonl (read-only count); bench/SYSTEMS_THINKING.md:12-13"},"owner_quote":{"text":"one source of truth, one main branch ... every idea is a branch you can bisect","src":"bench/tools/chain.py:4"},"file_exists":true,"git_first_add":"2026-10-30T14:41:18-07:00 022c5af0fdf sync: main's content a"},{"name":"Win loss audit","layer":"decide","file":"bench/tools/win_audit.py","introduced":"2026-10-30T15:35:03-07:00","introduced_src":"git first-add 70577bb5a29","date_note":"First-add is a cloud-box bulk commit ('swarm box: agents' work 22:35 UTC').","what":"Follows every CONFIRMED paired win and names the stage where it stopped before the scored row.","measured_effect":null,"why_null":"It prints its audit on demand; no stored audit output was found to count losses before and after it shipped.","owner_quote":{"text":"12 of 13 changes to the best file had no confirmed test; a confirmed +1.34% win was dropped","src":"bench/tools/win_audit.py:4-5"},"file_exists":true,"git_first_add":"2026-10-30T15:36:40-07:00 18acbeadf18 sync: main's content a"},{"name":"Hunch loop (learning loop)","layer":"decide","file":"bench/tools/hunch_loop.py","introduced":"2026-10-23T20:52:57-07:00","introduced_src":"git first-add ac3cd9e74b9","date_note":null,"what":"Every 30 min rebuilds FLOOR, WORKED and HUNCHES: hit rate per recipe, calibrated hunch odds, routes the top unowned hunch.","measured_effect":{"value":"Its calibration table: ideas registered at p 0.8-1.0 won 35%, p 0.6-0.8 won 29%, p 0.4-0.6 won 7%, p < 0.4 won 2%","kind":"measured","src":"bench/WORKED.md:18-24"},"owner_quote":{"text":"create a HUNCH list ... and then a list with what actually worked ... INVENT learning loops like this","src":"bench/tools/hunch_loop.py:5-6"},"file_exists":true,"git_first_add":"2026-10-23T20:52:57-07:00 ac3cd9e74b9 learning loop: hunch_l"},{"name":"Idea-basin detector","layer":"decide","file":"bench/tools/basin.py","introduced":"2026-11-01T11:21:06-08:00","introduced_src":"git first-add 4ac3a98e8d1","date_note":"First-add is a bulk 'keep: this disk's work' commit.","what":"CPU-only check that flags strike lanes re-touching the same knobs with no gain (concentration, novelty, re-touch).","measured_effect":{"value":"11-05 scan: 8 of 31 decode strike lanes in ALARM (e.g. d32s_norms novelty 0.00, re-touch 0.92). Basis: 72% of 444 attempts were 1-3 knob edits, 47% re-touched a measured knob","kind":"measured","src":"okc/muse_decode_layer/strike/lanes/BASINS.md:1-14; bench/research/fork_flush_1101.md:17-18"},"file_exists":true,"git_first_add":"2026-11-01T19:49:32Z 0c686b8cb02 agent box pi:b26:pi-b26-1=r"},{"name":"Codex islands","layer":"generate","file":"bench/tools/codex_island.py","introduced":"2026-10-22T17:56:14-07:00","introduced_src":"git first-add 05e714d3a92","date_note":"islands.py supervisor first-add 2026-10-22T18:53:34 (86a07516d36, an auto 'growth:' commit).","what":"GPT-6 Codex agents as rotating islands per problem (explore, alpha, combination lanes), context flushed each round.","measured_effect":{"value":"At decode row >= 2.2, 22% of whole-x trials were errors, mostly old Codex lanes. Off from 10-23 07:25 UTC","kind":"measured","src":"bench/agents/SPEC_V3.md:82; AGENTS.md:110"},"owner_quote":{"text":"i dont have anymore codex compute","src":"AGENTS.md:110"},"file_exists":true,"git_first_add":"2026-10-22T17:56:14-07:00 05e714d3a92 bench: Codex island la"},{"name":"pi agent launcher and brain switch","layer":"generate","file":"bench/tools/pi_agent.py","introduced":"2026-10-23T02:29:05-07:00","introduced_src":"git first-add e081eeb0cdb","date_note":"brain.py first-add 2026-10-23T15:48:05-07:00 (f760b4f65d2).","what":"Launches pi-harness agents with hard $ and time caps; brain.py and llm.env pick the model (Argon, Sol, DeepSeek).","measured_effect":{"value":"Same 12 case lanes: Argon GPU-off wave ~$96 -> 1 possible candidate; Sol round $2.83 -> 14 candidate files. pi had shown $506 while the real Argon bill was ~$2k","kind":"measured","src":"bench/SYSTEMS.md:129,132; bench/agents/FAMILY_LEARNINGS.md:115"},"owner_quote":{"text":"use gpt 6.1 sol for many cheap workflows ... no argon, we are spending way too much","src":"AGENTS.md:96"},"file_exists":true,"git_first_add":"2026-10-23T02:29:05-07:00 e081eeb0cdb main: pi_agent.py (pi "},{"name":"pi + Argon islands","layer":"generate","file":"bench/tools/pi_island.py","introduced":"2026-10-23T14:31:46-07:00","introduced_src":"git first-add 0f199e4f263","date_note":"AGENTS.md:90 dates the owner order 10-23 21:55 UTC (14:55 PDT), 24 min after the git first-add.","what":"Codex-island briefs re-run on pi with Gemini 4 Argon; 15 islands live from 10-23 22:00 UTC, $25 cap each.","measured_effect":{"value":"Argon arm: 2 clean gains >= 1% in 136 shots. Argon bill $505 (10-23 UTC), $1,098 (10-24 UTC). All 15 islands hit their $25 totals by 00:00 UTC","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:124-125; bench/SYSTEMS.md:107,131"},"owner_quote":{"text":"replace all codex research with Pi and Argon harness","src":"AGENTS.md:90"},"file_exists":true,"git_first_add":"2026-10-23T14:31:46-07:00 0f199e4f263 owner: replace Codex w"},{"name":"Manager fleet","layer":"generate","file":"bench/tools/manager.py","introduced":"2026-10-24T10:49:51-07:00","introduced_src":"git first-add 1cc290b34e6","date_note":"bench/SYSTEMS.md:134 logs the MANAGER HARNESS decision at 10-24 08:55 UTC (01:55 PDT), ~9 h before the git first-add.","what":"Keeps case lanes, a hypothesis generator and an adversary live in rounds; later holds 30-40 pi agents at all times.","measured_effect":{"value":"11-01 00:01-16:01 UTC: 715 of 1,386 Sonnet runner runs ended in under 1 min; 380 ended BLOCKED. Earlier Sol lanes stopped after 1-9 min of their 75","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:76-77,121"},"owner_quote":{"text":"at least 6 threads always running ... in sync ... never stop","src":"bench/SYSTEMS.md:134"},"file_exists":true,"git_first_add":"2026-10-24T10:49:51-07:00 1cc290b34e6 manager.py: the manage"},{"name":"Deep research and Exa corpus","layer":"generate","file":"bench/tools/deep_research.py","introduced":"2026-10-23T16:09:26-07:00","introduced_src":"git first-add 6ad3a36bf1e","date_note":"exa.py first-add 2026-10-22T18:53:34-07:00 (86a07516d36, an auto 'growth:' commit).","what":"Cited answers to one specific question (Exa deep, Gemini Deep Research) plus a 220-report, ~3,800-page corpus.","measured_effect":{"value":"Gemini Deep Research: 47 reports, 0 of 81 specific claims survived premise checks. Exa: $4.44 for 477 calls, 10-24..10-29","kind":"measured","src":"bench/SYSTEMS.md:117; bench/agents/FAMILY_LEARNINGS.md:99-100"},"file_exists":true,"git_first_add":"2026-10-23T16:09:26-07:00 6ad3a36bf1e deep_research.py (Gemi"},{"name":"Paper library and premise check","layer":"generate","file":"bench/tools/papers.py","introduced":"2026-10-23T14:59:05-07:00","introduced_src":"git first-add 0874a0884ac","date_note":"The --binds gate in sci.py was added 2026-10-24T12:40:33-07:00 (602c4f64f18).","what":"Greppable paper library; from 10-24 a paper-class idea needs a --binds receipt that its target stage binds.","measured_effect":{"value":"20 paper ideas tested: 12 missed (another limit bound), 5 caught by the premise check before GPU, 1 shipped (attn-002: v32 full 481 -> 366 us)","kind":"measured","src":"bench/research/PAPER_AUDIT.md:11-18,52-56"},"file_exists":true,"git_first_add":"2026-10-23T14:59:05-07:00 0874a0884ac Agents set up end to e"},{"name":"Fork-and-Flush lineages","layer":"generate","file":"bench/research/fork_flush_1101.md","introduced":"2026-11-01T10:46:14-08:00","introduced_src":"git first-add d45640e10a9","date_note":"Runs inside bench/tools/strike_lanes.py ('ff'), first-add 2026-10-31. The paper's 0.71 vs 0.45 (arXiv 2610.07447) is the paper's number, not ours.","what":"Every 4 leader runs, fork a structural lane 4 ways from the leader's files; 3 forks start with no chat.","measured_effect":{"value":"ON for only 2 lanes on 11-01. First d32s round (stale base cand59): branches -0.08 / +0.08 / -0.25%, two voided; no confirmed FF win found","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:61; bench/manager/NEXT_ACTIONS.md:249"},"owner_quote":{"text":"our hit rate btw is pretty low now ... rebuild spec","src":"bench/agents/SPEC_V3.md:3-5"},"file_exists":true,"git_first_add":"2026-11-01T11:58:52-08:00 3ed1741de8b sync: main's content a"},{"name":"MAP-Elites map-reduce","layer":"generate","file":"bench/tools/mapreduce.py","introduced":"2026-10-29T13:02:02-07:00","introduced_src":"git first-add 2016cf4bde9","date_note":null,"what":"Finds the best file per judge case, merges case winners into one dispatch file, writes AlphaEvolve-style briefs.","measured_effect":{"value":"68 REDUCE files built; 55 were over the 1,000,000-byte file cap; 0 reached a paired verdict","kind":"measured","src":"bench/mapreduce/reduce.jsonl (read-only count); bench/manager/paired.jsonl (no /mapreduce/ rows)"},"owner_quote":{"text":"do map reduce and do alpha generation concepts in here asap","src":"bench/tools/mapreduce.py:4"},"file_exists":true,"git_first_add":"2026-10-29T13:05:35-07:00 8a020474434 sync: main's content a"},{"name":"Mailbox and ledger notes","layer":"coordinate","file":"bench/tools/mailbox.py","introduced":"2026-10-22T21:12:43-07:00","introduced_src":"git first-add 1b70661a631","date_note":"First-add is a batch commit (message is about statecopy.py).","what":"Per-agent read receipts ('blue dot') on shared files and notes; sci.py notes, todos and triggers carry the messages.","measured_effect":null,"why_null":"Activity only: 78,280 notes, 6,203 closed, 3,885 triggers fired (bench/sci/events.jsonl count). Counter-case: 5 todo notes to 'meta-0c' sat unseen 2.5 h (bench/agents/FAMILY_LEARNINGS.md:40). No outcome link measured.","owner_quote":{"text":"each agent has their own READ mailbox ... kinda like a blue dot","src":"bench/tools/mailbox.py:4"},"file_exists":true,"git_first_add":"2026-10-22T21:12:43-07:00 1b70661a631 tools: statecopy.py an"},{"name":"Agent groups and forum","layer":"coordinate","file":"bench/tools/groups.py","introduced":"2026-10-22T23:45:21-07:00","introduced_src":"git first-add c9385d6a36f","date_note":"forum.py first-add 2026-10-27T16:11:23-07:00 (1278663045d).","what":"Agent-run group chats (open or private) and, from 10-27, a deployed forum with channels, mentions and receipts.","measured_effect":null,"why_null":"Activity only: 23 group folders in bench/groups, 13,676 forum_sent events in the ledger; no link from a post to a gain was measured.","owner_quote":{"text":"let agents organise groups ... private chats for agents","src":"bench/tools/groups.py:4-5"},"file_exists":true,"git_first_add":"2026-10-22T23:45:21-07:00 c9385d6a36f bench/tools/groups.py:"},{"name":"Coordinator, work queue and unblock sweep","layer":"coordinate","file":"bench/manager/COORDINATOR.md","introduced":"2026-10-27T14:43:22-07:00","introduced_src":"git first-add f1366359246 (unblock.py)","date_note":"WORK_QUEUE.json first-add 14:52 (79883886d23) and COORDINATOR.md 14:57 (e3cb176b9d1), same afternoon.","what":"Start-of-session steps and tempo targets; idle lanes pull the next ranked task; unblock.py turns each stall into a flag.","measured_effect":{"value":"Built on: GPU busy 41% with 28 jobs queued; claude5 3/5 and claude6 3/4 lanes idle for hours. Later failure: 92 stale open tasks, one idea queued 25 times","kind":"measured","src":"bench/manager/COORDINATOR.md:11-12; bench/agents/FAMILY_LEARNINGS.md:73"},"owner_quote":{"text":"we are underutilised on resources and are not at our pace or tempo","src":"bench/manager/COORDINATOR.md:3-5"},"file_exists":true,"git_first_add":"2026-10-27T14:57:00-07:00 e3cb176b9d1 COORDINATOR.md: start "},{"name":"Mistake log","layer":"remember","file":"bench/MISTAKES.md","introduced":"2026-10-22T20:55:48-07:00","introduced_src":"git first-add 49e12a79980","date_note":"First-add is a commit about the SASS tool (batch).","what":"One row per judge FAIL, voided trial or retracted claim: the mistake class, its cost, and the gate that now catches it.","measured_effect":null,"why_null":"57,127 lines (38 MB) by 11-05, mostly appended on-policy rows; the repeat rate of mistake classes was never measured.","file_exists":true,"git_first_add":"2026-10-22T20:55:48-07:00 49e12a79980 sass tool: compile wit"},{"name":"Sleep cycle and CORE.md","layer":"remember","file":"bench/tools/sleep.py","introduced":"2026-10-27T15:17:41-07:00","introduced_src":"git first-add 3a72a8913cc","date_note":null,"what":"Every 3 h: consolidate lessons, rebuild a <= 150-line CORE.md, check lane transcripts against rules, dream cross-problem tasks.","measured_effect":{"value":"On-policy checks: 963 rule applications vs 159 violations (86% applied, 158 lane checks). Dreams: 6 hits of 285 resolved (2.1%), 402 dreams queued","kind":"measured","src":"bench/manager/sleep/onpolicy/*.json and bench/manager/sleep/dreams.jsonl (read-only count)"},"owner_quote":{"text":"the student has the knowledge but does not use it (paper line the design targets)","src":"bench/manager/SLEEP.md:14"},"file_exists":true,"git_first_add":"2026-10-27T15:17:41-07:00 3a72a8913cc sleep cycle (arXiv 260"},{"name":"Agent traces database","layer":"remember","file":"bench/tools/traces.py","introduced":"2026-10-24T19:45:17-07:00","introduced_src":"git first-add c20ccf2512c","date_note":null,"what":"Every agent's turns, tool calls, ledgers and briefs into one SQLite db; feeds trajectory and brief-quality models.","measured_effect":{"value":"Brief model (brief text -> P(beat prior)) grouped CV AUC 0.877. Logs showed 0 of 7,914 lane runs did any research","kind":"measured","src":"bench/SYSTEMS.md:152; bench/manager/AB_RESEARCH.md:4"},"file_exists":true,"git_first_add":"2026-10-24T19:45:17-07:00 c20ccf2512c fleet to Sol (owner 10"},{"name":"Send pacing and per-problem lock","layer":"guard","file":"bench/tools/okc_run.py","introduced":"2026-10-23T11:50:29-07:00","introduced_src":"git pickaxe 'def released(kind' 2994d96b770","date_note":"okc_run.py itself first-add 2026-10-22T15:12:16 (96c9d26cc4c); the per-problem lock came 2026-10-23T17:16:22-07:00 (b8bd3ca5d70) after two main sessions sent the same problem at 00:03 UTC.","what":"The only sender: releases 10 scored runs through the day, one send per problem at a time, row bar on scored runs.","measured_effect":{"value":"Built after decode spent 10 of 10 scored runs by 9:50 PM PDT 10-22 and its row froze 19 h (rank 2 -> 8). Later the row bar refused cand59 (2.3044 < 2.3408)","kind":"measured","src":"AGENTS.md:88-89; bench/agents/FAMILY_LEARNINGS.md:138"},"owner_quote":{"text":"be careful and distribute submissions carefully","src":"AGENTS.md:79"},"file_exists":true,"git_first_add":"2026-10-22T15:12:16-07:00 96c9d26cc4c Kernel challenge: okc "},{"name":"State-copy gate","layer":"guard","file":"bench/modal/statecopy.py","introduced":"2026-10-22T20:50:47-07:00","introduced_src":"git first-add d758bcde074","date_note":"First-add is a batch commit (message about OWNER_TODO.md); AGENTS.md dates the finding N1300 10-23 03:55 UTC, the same evening.","what":"Re-runs a kernel on tensor-cloned copies of its prepared state, as the judge does, to catch baked raw addresses.","measured_effect":{"value":"Explained all 5 MoE judge FAILs at bs=128 (rel L2 1.414: the kernel read the previous call's buffers from a state copy)","kind":"measured","src":"bench/modal/statecopy.py:4-9; AGENTS.md:126-129"},"file_exists":true,"git_first_add":"2026-10-22T20:50:47-07:00 d758bcde074 main: bench/OWNER_TODO"},{"name":"PDL load-order gate","layer":"guard","file":"bench/audit/gates/pdl_order_gate.py","introduced":"2026-10-23T15:49:03-07:00","introduced_src":"git first-add 8208c46bfe1","date_note":"First-add is a batch commit ('main tick 22:49'); AGENTS.md credits claude-dec-prewait 10-24, under 1 day later.","what":"Reads the SASS of our modules and fails a kernel whose loads of producer output move above griddepcontrol.wait.","measured_effect":{"value":"[sealed until Nov 19]","kind":"measured","src":"AGENTS.md:144"},"file_exists":true,"git_first_add":"2026-10-23T15:49:03-07:00 8208c46bfe1 main tick 22:49: ledge"},{"name":"Paired-plan guard","layer":"guard","file":"bench/tools/paired_guard.py","introduced":"2026-10-30T15:27:35-07:00","introduced_src":"git first-add 20422a89802","date_note":null,"what":"Never spends a paired plan on a base or box that cannot run: base registry, base-first preflight, fault attribution.","measured_effect":{"value":"ABORTED share of paired verdicts: 271 of 564 (48%) from 10-28 22:00 to 10-30 22:00 -> 27 of 221 (12%) from 10-30 22:00 to 11-03 01:32. Docstring pre-count: 380 of 606 plans (63%)","kind":"measured","src":"bench/manager/paired.jsonl (read-only count by 't' field, timezone unmarked); bench/tools/paired_guard.py:5-8"},"owner_quote":{"text":"we are wasting compute. figure out a way to prevent this from happening","src":"bench/tools/paired_guard.py:2-3"},"file_exists":true,"git_first_add":"2026-10-30T15:36:40-07:00 18acbeadf18 sync: main's content a"},{"name":"Idle-box reaper and spend guard","layer":"guard","file":"bench/tools/boxwatch.py","introduced":"2026-10-22T21:06:22-07:00","introduced_src":"git first-add 4dbe3c50c69","date_note":"spend_guard.py first-add 2026-10-23T11:51:43-07:00 (03b44effc5b). boxwatch's first-add commit is a batch commit.","what":"Samples every H200 box; stops boxes idle 15-30 min and orphan apps; spend_guard caps boxes and alarms on burn.","measured_effect":{"value":"Before: $777 for 171 box-hours, 62% idle (10-23 audit). Reaper 10-23..11-05: 403 idle boxes stopped, est. $3,656.53 refunded to the daily cap","kind":"measured","src":"bench/tools/spend_guard.py:3; bench/modal/reaped.jsonl (sum of saved_usd_est, rc=0 rows)"},"owner_quote":{"text":"have a thing which tracks the idle stuff, we are not $$$ rich","src":"bench/tools/boxwatch.py:14"},"file_exists":true,"git_first_add":"2026-10-22T21:06:22-07:00 4dbe3c50c69 main: MoE scored 2.022"},{"name":"Writer provenance gate","layer":"guard","file":"bench/tools/provenance.py","introduced":"2026-10-27T19:35:06-07:00","introduced_src":"git first-add 24d574ec473","date_note":"writer_registry.jsonl first-add 2026-10-27T19:41:12-07:00 (bf310e8bebc). bench/agents/FAMILY_LEARNINGS.md:25 notes the written 'Opus only' rule differs from the running Sonnet writer tier.","what":"Only registered writer models may write source; every file goes into the writer registry and a compile gate before GPU.","measured_effect":{"value":"Registry 3,862 files (Sonnet 2,046, subscription seats 1,504, Opus 312). First-try gate pass: Opus 99.1% of 651, Sonnet 94.6% of 1,404, Haiku 89.1% of 404","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:68,83"},"owner_quote":{"text":"DeepSeek is the default brain; code is written only by Opus 5.5","src":"AGENTS.md:109"},"file_exists":true,"git_first_add":"2026-10-27T19:35:06-07:00 24d574ec473 pi writer: Claude Sonn"},{"name":"Readable-source send refusal","layer":"guard","file":"bench/tools/okc_run.py","introduced":"2026-11-03T13:49:55-08:00","introduced_src":"git pickaxe 'def encoded_source_reason' 085e86cb1f1","date_note":null,"what":"okc_run.py refuses any file with compressed or encoded source (lzma, zlib, base64, decoded exec); no override.","measured_effect":null,"why_null":"Added the day 46 of our runs were disqualified (11-03), one day before the kernel close; no later refusal count was found.","owner_quote":{"text":"submissions must be readable (organizers' rule, 11-02 08:00 PT)","src":"bench/tools/okc_run.py:628"},"file_exists":true,"git_first_add":"2026-10-22T15:12:16-07:00 96c9d26cc4c Kernel challenge: okc "},{"name":"Stop hooks (limit check, turn end)","layer":"guard","file":"bench/tools/turn_end.py","introduced":"2026-10-24T19:25:49Z","introduced_src":"git first-add 273d7d334c6 (limit_check.py)","date_note":"limit_check.py first-add is a cloud-box bulk commit; bench/SYSTEMS.md:136 logs the hook at 10-24 12:15 UTC (7 h earlier). turn_end.py first-add 2026-11-01T17:19:07Z (00a37e36507).","what":"Blocks a session's stop while a problem is below its ceiling or the goal is unmet; the reason becomes the next prompt.","measured_effect":{"value":"Against: about 12 consecutive turns answered the stop hook with a status line and ran no experiment","kind":"measured","src":"bench/books/POSTMORTEM_1029.md:33"},"owner_quote":{"text":"on every single hook ... check if you have actually reached the physical limit ... never stop","src":"bench/tools/limit_check.py:2-3"},"file_exists":true,"git_first_add":"2026-11-01T08:16:15-08:00 d5234baa503 sync: main's content a"},{"name":"H200 boxes","layer":"infra","file":"bench/modal/box.py","introduced":"2026-10-22T17:43:04-07:00","introduced_src":"git first-add 20770817c93","date_note":null,"what":"Rented Modal H200 sandboxes with the judge's image, a FIFO GPU queue, late binding and an in-box judge.","measured_effect":{"value":"1 h rentals died with their queues (10-24: 83 RL runs in 7 h) -> 4 h default. 1 CPU core per box starved nvcc -> 16 cores per box","kind":"measured","src":"bench/SYSTEMS.md:72-73,98"},"owner_quote":{"text":"rent one H200 for exploration ... do it","src":"bench/modal/box.py:3"},"file_exists":true,"git_first_add":"2026-10-22T17:43:04-07:00 20770817c93 bench: H200 box (modal"},{"name":"Object-store sync (push_sync, safe_pull)","layer":"infra","file":"bench/tools/push_sync.py","introduced":"2026-10-23T15:10:45-07:00","introduced_src":"git first-add 424f4a48913","date_note":"safe_pull.py first-add 2026-10-23T15:12:47-07:00 (3ad17dee710).","what":"Push and pull with merge-tree and commit-tree only; ledgers are unioned, so 40 writing agents never lose an append.","measured_effect":{"value":"Built after a git hard reset 10-23 05:09 UTC zeroed island files and lost ~10 ledger writes. Later: 2.27 GB per push try lost every race for > 1 h","kind":"measured","src":"AGENTS.md:214-216; bench/agents/FAMILY_LEARNINGS.md:39"},"file_exists":true,"git_first_add":"2026-10-23T15:10:45-07:00 424f4a48913 push_sync.py: push whe"},{"name":"Error alarm hook","layer":"infra","file":"bench/tools/alarm.py","introduced":"2026-10-23T19:21:41-07:00","introduced_src":"git first-add e9ab9270eef","date_note":null,"what":"Every 2 min checks dead boxes, crashed launches, dead lanes, judge fails, loop tracebacks; Telegram plus inbox.","measured_effect":{"value":"First night: boxes 5 and 8 found dead with live reservations; box.py down refunded $31.42","kind":"measured","src":"bench/SYSTEMS.md:115-116"},"owner_quote":{"text":"is there a HOOK for error or smth u know??","src":"bench/tools/alarm.py:2"},"file_exists":true,"git_first_add":"2026-10-23T19:21:41-07:00 e9ab9270eef alarm.py: the error ho"},{"name":"GPU idle filler","layer":"infra","file":"bench/tools/gpu_fill.py","introduced":"2026-10-24T17:51:19-07:00","introduced_src":"git first-add 8637cdf92b5","date_note":null,"what":"Gives each idle box a strict run of a near-best file, so idle billed GPU time adds evidence.","measured_effect":{"value":"Built when 11 of 13 boxes sat at 0% GPU. Coordinator later reports GPU utilisation 17% -> 31%, and calls it an input, not the objective","kind":"measured","src":"bench/tools/gpu_fill.py:5; bench/books/POSTMORTEM_1029.md:38"},"owner_quote":{"text":"max out so nothing is waiting, faster, but keep pushing","src":"bench/tools/gpu_fill.py:2"},"file_exists":true,"git_first_add":"2026-10-24T17:51:19-07:00 8637cdf92b5 fleet throughput: GPU "}],"method":[{"rule":"Pre-register each idea with a numeric predicted gain and probability before its trial","since":"2026-10-22","enforced_by":"bench/tools/sci.py add (gain, p fields); bench/PROTOCOL.md rule 7","evidence_for":"Predictions carry signal: ideas registered at p 0.8-1.0 won 35%, p < 0.4 won 2% (bench/WORKED.md:20-24)","evidence_against":"Agents are overconfident (p 0.4-0.6 won 7%, bench/WORKED.md:22); strike lanes: 0 of 900 claims carried a prediction (bench/agents/FAMILY_LEARNINGS.md:56)","verdict":"mixed","ledger_evidence":{"value":"14,936 of 14,936 ideas carry a numeric gain and p. First full trial vs prediction: median predicted +0.25%, median measured +0.01%, within 2x 23%, below prediction 72% (n 3423). p bins: p 0.0-0.2 -> 0.7% (n 277), p 0.2-0.4 -> 2.0% (n 1030), p 0.4-0.6 -> 6.5% (n 1704), p 0.6-0.8 -> 25.8% (n 532), p 0.8-1.0 -> 31.4% (n 220)","kind":"measured","src":"bench/sci/events.jsonl via build_science.py (this build)"}},{"rule":"Retire an idea after 3 misses with no hit, or 2 BAD trials in a row","since":"2026-10-22","enforced_by":"bench/tools/sci.py (state replay, auto-kill)","evidence_for":"Mining 435 positive whole-x flats found ~30 worth re-measuring and 1 new gene; the rest were noise or already harvested (bench/agents/SPEC_V3.md:94-97)","evidence_against":"437 flats with a positive estimate (median +0.197%) were never re-measured (bench/agents/SPEC_V3.md:81-82); rule amended: positive observations no longer count toward retirement (bench/tools/sci.py:632); 'Slower is data' after 431 slower / 8 KEEP (bench/agents/FAMILY_LEARNINGS.md:52)","verdict":"mixed","ledger_evidence":{"value":"67 trials ran after an explicit kill with 0 full hits; 339 trials ran after the auto rule tripped on 299 ideas, 0 rescued by a later full hit; 204 of 1001 explicit kills had no trial (duplicates, premise checks); 9 revivals","kind":"measured","src":"bench/sci/events.jsonl via build_science.py (this build)"}},{"rule":"Back every gain claim with a same-box, same-window A/A control","since":"2026-10-23","enforced_by":"bench/tools/sci.py trial --control; bench/PROTOCOL.md rule 4","evidence_for":"\"+0.8%\" was +0.06% on the same box (AGENTS.md:140); abpair read up to +0.5% on identical bytes (bench/agents/SPEC_V3.md:122)","evidence_against":"Per-run A/A subtraction added noise: a slower candidate (4/4 runs slower) pooled as +0.049%; fixed by pooling raw deltas (bench/agents/SPEC_V3.md:118)","verdict":"helped","ledger_evidence":{"value":"share of timed box/modal trials carrying a control or paired receipt, per UTC day: 10-23 11%, 10-24 54%, 10-25 71%, 10-26 96%, 10-27 84%, 10-28 91%, 10-29 80%, 10-30 88%, 10-31 77%, 11-01 91%, 11-02 75%, 11-03 94%, 11-04 94%","kind":"measured","src":"bench/sci/events.jsonl via build_science.py (this build)"}},{"rule":"Confirm with paired A/B in both orientations on >= 20 random draws per shape","since":"2026-10-24","enforced_by":"bench/tools/paired_confirm.py (ABBA blocks); bench/tools/clone_inputs.py","evidence_for":"One orientation read 0.1-0.2% slower for arm B (AGENTS.md:141); clone fixed draws +0.83% vs random-draw +0.45% (bench/tools/clone_inputs.py:6-7); decode C screened +0.289%, confirmed +0.222% CI 0.180..0.264 (bench/reports/evolution/2026-10-26-01a17a50/confirm-decode.md:3-5)","evidence_against":"54% of paired verdicts ABORTED for infrastructure in 72 h to 10-29 (bench/books/POSTMORTEM_1029.md:18); 2 router salts proved under-sampled, >= 6 draws needed (bench/agents/SPEC_V3.md:131)","verdict":"helped"},{"rule":"Rank work by each case's gap to its physical floor","since":"2026-10-22","enforced_by":"bench/tools/gap.py; bench/tools/limit_check.py","evidence_for":"Located the ceilings per problem (attn 58% .. bf16 88% of ceiling, AGENTS.md:62)","evidence_against":"Recipe 'floor-first' won 7% of 157 clean tests vs 16% for 'compose' (bench/WORKED.md:7,13); SPEC v3 ranks TWEAK and PORT by us-value instead of gap size (bench/agents/SPEC_V3.md:68-69)","verdict":"mixed"},{"rule":"Pass the state-copy check before any send","since":"2026-10-22","enforced_by":"bench/modal/statecopy.py; judge.py strict mode","evidence_for":"Explained all 5 MoE judge FAILs at bs=128 (AGENTS.md:126-129)","evidence_against":null,"verdict":"helped"},{"rule":"Promote on judge-calibrated evidence, never on raw clone or box x","since":"2026-10-23","enforced_by":"bench/tools/board.py j/c column; bench/tools/judge_cal.py; bench/tools/internal_judge.py","evidence_for":"Box read decode 3.6% above the judge; box rank vs judge rank Spearman 0.28, 0.63 after calibration (bench/tools/judge_cal.py:4-5; commit 26ce4af52eb)","evidence_against":"One judge read moves ~1% run to run; the board row 2.3408 was +1.7% over its bytes' median, so from 11-01 internal paired truth decides and the judge only checks (bench/agents/SPEC_V3.md:104-110; PROTOCOL.md rule 9: judge sd 0.60%)","verdict":"mixed"},{"rule":"Start every candidate as a <= 50-line diff of the current head (ROUTE / PORT / COMPOSE / PIN / TWEAK)","since":"2026-11-01","enforced_by":"bench/agents/SPEC_V3.md rules 1-2; bench/tools/chain.py","evidence_for":"All 44 decode hits since 10-31 (7.1% of 619 trials) were those five kinds; from-scratch designs ~0 hits, 20+ dead on 11-01 (bench/agents/SPEC_V3.md:9-24); 0 of 92 one-shot > 40-line edits kept vs 6 of 319 knob edits (bench/research/fork_flush_1101.md:191)","evidence_against":"Gain per landing fell +1.08 -> +0.12% on 10-31 (bench/SYSTEMS_THINKING.md:12-13); hit rate 0.5% at row 2.2-2.3 and 0% above 2.3 (bench/agents/SPEC_V3.md:73-78)","verdict":"mixed","ledger_evidence":{"value":"full-hit rate by mechanism group: compose / combine 14.9% of 786 tried ideas (90% CI 0.129-0.171); port / route 10.4% of 356 tried ideas (90% CI 0.080-0.134); knob / pin / tune 2.7% of 368 tried ideas (90% CI 0.016-0.045); kernel structure 6.0% of 662 tried ideas (90% CI 0.047-0.077); latency chain / launch 8.0% of 465 tried ideas (90% CI 0.061-0.103); bytes / precision 6.2% of 276 tried ideas (90% CI 0.042-0.090)","kind":"measured","src":"bench/sci/events.jsonl via build_science.py (this build)"}},{"rule":"Keep sub-1% gains, race them, confirm the finalists paired and compose them (PROTOCOL rule 13)","since":"2026-10-26","enforced_by":"bench/PROTOCOL.md rule 13; bench/tools/race.py; bench/tools/keeps.py; bench/tools/chain.py","evidence_for":"+0.25% on decode = 2 board places (AGENTS.md:269); 115 decode chain landings 10-30..11-05 (bench/chain/muse_decode_layer.jsonl); PF0RH + STG stacked into cand63 (bench/agents/SPEC_V3.md:42)","evidence_against":"24 h to 10-29: lanes claimed +4,084 bps, the board moved +57 bps, conversion 1.4% (bench/books/POSTMORTEM_1029.md:21)","verdict":"mixed","ledger_evidence":{"value":"2,897 small-gain (0..+1%) trial labels; paired confirmations: 130 CONFIRMED of 667 decided (19.5%), 788 ABORTED of 1612 rows","kind":"measured","src":"bench/sci/events.jsonl via build_science.py (this build)"}},{"rule":"Put the best public rival code beside ours and take the faster route per case","since":"2026-10-27","enforced_by":"bench/external/<team>/USE.md; AGENTS.md 'What worked on 10-27'","evidence_for":"One afternoon: prefill 1.737 -> 1.799, attention 2.597 -> 2.620, bf16 1.466 -> 1.5375 (#171 -> #46) (AGENTS.md:260); later ports bf16 +2.43% strict, prefill fmha7 +6.2% hybrid (bench/books/POSTMORTEM_1029.md:27)","evidence_against":"Decode 'port' ideas hit 2 of 26 (8%) since 10-31 (bench/agents/SPEC_V3.md:10-11)","verdict":"helped","ledger_evidence":{"value":"rival / peer code 13.2% of 76 tried ideas (90% CI 0.080-0.208) vs papers (arXiv) 18.2% of 22 tried ideas (90% CI 0.084-0.349) vs lane's own work (origin not stated) 7.5% of 1535 tried ideas (90% CI 0.065-0.087)","kind":"measured","src":"bench/sci/events.jsonl via build_science.py (this build)"}},{"rule":"Give a paper idea GPU time only with a receipt that its target stage binds on our base","since":"2026-10-24","enforced_by":"bench/tools/sci.py add --binds (paper-class sources); bench/research/PAPER_AUDIT.md","evidence_for":"5 of 20 paper ideas caught before GPU; >= 6 of 12 GPU misses could have been stopped on CPU (bench/research/PAPER_AUDIT.md:52-56)","evidence_against":"Only 153 of 14,667 ideas carry a binds field (bench/sci/events.jsonl count); the 11-01 sweep of 41 papers gave 3 ideas, all killed (bench/agents/FAMILY_LEARNINGS.md:107-108)","verdict":"helped","ledger_evidence":{"value":"papers-class ideas (sci.py source_class): {\"ideas\": 144, \"with_trial\": 39, \"with_full_hit\": 1, \"with_binds_field\": 16}","kind":"measured","src":"bench/sci/events.jsonl via build_science.py (this build)"}},{"rule":"One sender, one send per problem at a time, scored runs paced through the day","since":"2026-10-23","enforced_by":"bench/tools/okc_run.py (send lock, pacing, row bar)","evidence_for":"Before pacing, decode spent 10 of 10 scored runs by 9:50 PM 10-22 and its row froze 19 h, rank 2 -> 8 (AGENTS.md:88-89); two main sessions sent one problem in parallel at 00:03 UTC (bench/tools/okc_run.py:777-779)","evidence_against":null,"verdict":"unknown"}],"idea_to_trial_min":7},"episodes":[{"id":"argon-bf16-v2-20261102","model":"Gemini 4 Argon High","harness":"Antigravity 1.4.12","family":"dense_proj_bf16","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":51.14314098497682},"T2":{"x":1.4463634627802266,"pct":67.24688226589855},"T3":{"x":1.8927269255604537,"pct":88.0},"T4":{"x":1.9357434465959187,"pct":90.0}},"steps":[{"step":1,"x":0.9204010713015859,"lo":0.9168907826490328,"hi":0.9238674159665393,"pct":42.99131716964966,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.49,"ours_us":37.44,"base_us":18.36,"floor_pct":26.4},{"load":"decode 1","shape":"m=1 o","x":0.827,"ours_us":11.17,"base_us":9.23,"floor_pct":49.2},{"load":"decode 1","shape":"m=1 pooler","x":0.85,"ours_us":32.56,"base_us":27.66,"floor_pct":49.1},{"load":"decode 1","shape":"m=1 unpooler","x":0.728,"ours_us":11.9,"base_us":8.66,"floor_pct":31.5},{"load":"decode 1","shape":"m=1 router","x":0.954,"ours_us":12.64,"base_us":12.06,"floor_pct":29.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.488,"ours_us":37.7,"base_us":18.41,"floor_pct":26.4},{"load":"decode 16","shape":"m=16 o","x":0.818,"ours_us":11.58,"base_us":9.47,"floor_pct":48.0},{"load":"decode 16","shape":"m=16 pooler","x":0.856,"ours_us":32.65,"base_us":27.96,"floor_pct":49.2},{"load":"decode 16","shape":"m=16 unpooler","x":0.77,"ours_us":32.88,"base_us":25.33,"floor_pct":48.8},{"load":"decode 16","shape":"m=16 router","x":2.232,"ours_us":12.82,"base_us":28.62,"floor_pct":29.7},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.467,"ours_us":40.46,"base_us":18.91,"floor_pct":24.7},{"load":"decode 32","shape":"m=32 o","x":0.739,"ours_us":12.79,"base_us":9.45,"floor_pct":43.9},{"load":"decode 32","shape":"m=32 pooler","x":0.827,"ours_us":34.65,"base_us":28.64,"floor_pct":46.6},{"load":"decode 32","shape":"m=32 unpooler","x":0.725,"ours_us":35.21,"base_us":25.52,"floor_pct":45.8},{"load":"decode 32","shape":"m=32 router","x":1.568,"ours_us":13.88,"base_us":21.76,"floor_pct":27.8},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.494,"ours_us":37.28,"base_us":18.4,"floor_pct":26.6},{"load":"verify 1","shape":"m=8 o","x":0.822,"ours_us":11.24,"base_us":9.24,"floor_pct":49.2},{"load":"verify 1","shape":"m=8 pooler","x":0.856,"ours_us":32.69,"base_us":27.98,"floor_pct":49.0},{"load":"verify 1","shape":"m=8 unpooler","x":0.747,"ours_us":11.86,"base_us":8.86,"floor_pct":31.8},{"load":"verify 1","shape":"m=8 router","x":2.204,"ours_us":12.85,"base_us":28.32,"floor_pct":29.4},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.001,"ours_us":22.11,"base_us":22.13,"floor_pct":47.1},{"load":"verify 16","shape":"m=128 o","x":0.861,"ours_us":5.16,"base_us":4.45,"floor_pct":115.9},{"load":"verify 16","shape":"m=128 pooler","x":0.974,"ours_us":29.38,"base_us":28.62,"floor_pct":56.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.979,"ours_us":27.32,"base_us":26.75,"floor_pct":60.9},{"load":"verify 16","shape":"m=128 router","x":2.212,"ours_us":14.33,"base_us":31.7,"floor_pct":29.4},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.994,"ours_us":26.02,"base_us":25.87,"floor_pct":48.9},{"load":"verify 32","shape":"m=256 o","x":0.982,"ours_us":11.35,"base_us":11.14,"floor_pct":59.7},{"load":"verify 32","shape":"m=256 pooler","x":0.977,"ours_us":35.98,"base_us":35.16,"floor_pct":58.6},{"load":"verify 32","shape":"m=256 unpooler","x":0.991,"ours_us":30.54,"base_us":30.28,"floor_pct":69.0},{"load":"verify 32","shape":"m=256 router","x":3.305,"ours_us":14.82,"base_us":48.97,"floor_pct":31.6}]},{"step":2,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":2,"x":1.149978737245633,"lo":1.146202242086218,"hi":1.1531204083676896,"pct":52.85559716396586,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.03,"ours_us":18.09,"base_us":18.64,"floor_pct":54.5},{"load":"decode 1","shape":"m=1 o","x":0.853,"ours_us":10.91,"base_us":9.31,"floor_pct":50.4},{"load":"decode 1","shape":"m=1 pooler","x":1.038,"ours_us":26.83,"base_us":27.85,"floor_pct":59.6},{"load":"decode 1","shape":"m=1 unpooler","x":1.081,"ours_us":8.25,"base_us":8.92,"floor_pct":45.4},{"load":"decode 1","shape":"m=1 router","x":1.315,"ours_us":9.32,"base_us":12.26,"floor_pct":40.2},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.005,"ours_us":18.63,"base_us":18.72,"floor_pct":53.3},{"load":"decode 16","shape":"m=16 o","x":0.854,"ours_us":11.11,"base_us":9.49,"floor_pct":50.0},{"load":"decode 16","shape":"m=16 pooler","x":1.035,"ours_us":26.96,"base_us":27.9,"floor_pct":59.6},{"load":"decode 16","shape":"m=16 unpooler","x":0.907,"ours_us":27.84,"base_us":25.25,"floor_pct":57.7},{"load":"decode 16","shape":"m=16 router","x":2.975,"ours_us":9.61,"base_us":28.59,"floor_pct":39.6},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.024,"ours_us":18.54,"base_us":18.98,"floor_pct":53.9},{"load":"decode 32","shape":"m=32 o","x":0.794,"ours_us":11.95,"base_us":9.49,"floor_pct":47.0},{"load":"decode 32","shape":"m=32 pooler","x":1.051,"ours_us":27.37,"base_us":28.77,"floor_pct":59.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.887,"ours_us":28.76,"base_us":25.52,"floor_pct":56.1},{"load":"decode 32","shape":"m=32 router","x":1.773,"ours_us":12.31,"base_us":21.83,"floor_pct":31.4},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.013,"ours_us":18.45,"base_us":18.68,"floor_pct":53.7},{"load":"verify 1","shape":"m=8 o","x":0.825,"ours_us":11.21,"base_us":9.25,"floor_pct":49.3},{"load":"verify 1","shape":"m=8 pooler","x":1.035,"ours_us":26.79,"base_us":27.74,"floor_pct":59.8},{"load":"verify 1","shape":"m=8 unpooler","x":1.076,"ours_us":8.29,"base_us":8.93,"floor_pct":45.4},{"load":"verify 1","shape":"m=8 router","x":2.949,"ours_us":9.7,"base_us":28.61,"floor_pct":38.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.158,"ours_us":19.5,"base_us":22.58,"floor_pct":53.5},{"load":"verify 16","shape":"m=128 o","x":0.968,"ours_us":10.37,"base_us":10.03,"floor_pct":57.7},{"load":"verify 16","shape":"m=128 pooler","x":0.994,"ours_us":29.98,"base_us":29.81,"floor_pct":55.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.867,"ours_us":31.31,"base_us":27.14,"floor_pct":53.1},{"load":"verify 16","shape":"m=128 router","x":2.351,"ours_us":13.4,"base_us":31.51,"floor_pct":31.4},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.227,"ours_us":21.78,"base_us":26.74,"floor_pct":58.5},{"load":"verify 32","shape":"m=256 o","x":0.621,"ours_us":18.45,"base_us":11.46,"floor_pct":36.7},{"load":"verify 32","shape":"m=256 pooler","x":1.065,"ours_us":33.4,"base_us":35.58,"floor_pct":63.1},{"load":"verify 32","shape":"m=256 unpooler","x":0.836,"ours_us":36.18,"base_us":30.24,"floor_pct":58.3},{"load":"verify 32","shape":"m=256 router","x":3.39,"ours_us":14.5,"base_us":49.14,"floor_pct":32.3}]},{"step":3,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":3,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":3,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":3,"x":0.992052897634296,"lo":0.987008635553778,"hi":0.9955320365273845,"pct":46.028001426669825,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.511,"ours_us":35.85,"base_us":18.31,"floor_pct":27.5},{"load":"decode 1","shape":"m=1 o","x":1.067,"ours_us":8.64,"base_us":9.22,"floor_pct":63.7},{"load":"decode 1","shape":"m=1 pooler","x":0.697,"ours_us":40.05,"base_us":27.91,"floor_pct":39.9},{"load":"decode 1","shape":"m=1 unpooler","x":0.512,"ours_us":16.61,"base_us":8.5,"floor_pct":22.6},{"load":"decode 1","shape":"m=1 router","x":1.657,"ours_us":7.22,"base_us":11.97,"floor_pct":51.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.337,"ours_us":53.85,"base_us":18.14,"floor_pct":18.4},{"load":"decode 16","shape":"m=16 o","x":0.8,"ours_us":11.77,"base_us":9.42,"floor_pct":47.2},{"load":"decode 16","shape":"m=16 pooler","x":0.477,"ours_us":59.17,"base_us":28.25,"floor_pct":27.1},{"load":"decode 16","shape":"m=16 unpooler","x":0.62,"ours_us":40.63,"base_us":25.2,"floor_pct":39.5},{"load":"decode 16","shape":"m=16 router","x":2.956,"ours_us":9.55,"base_us":28.23,"floor_pct":39.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.01,"ours_us":18.19,"base_us":18.38,"floor_pct":55.0},{"load":"decode 32","shape":"m=32 o","x":0.997,"ours_us":9.38,"base_us":9.34,"floor_pct":59.9},{"load":"decode 32","shape":"m=32 pooler","x":1.046,"ours_us":27.22,"base_us":28.46,"floor_pct":59.3},{"load":"decode 32","shape":"m=32 unpooler","x":0.941,"ours_us":27.58,"base_us":25.96,"floor_pct":58.5},{"load":"decode 32","shape":"m=32 router","x":1.73,"ours_us":12.49,"base_us":21.6,"floor_pct":30.9},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.004,"ours_us":18.23,"base_us":18.31,"floor_pct":54.3},{"load":"verify 1","shape":"m=8 o","x":0.987,"ours_us":9.4,"base_us":9.27,"floor_pct":58.8},{"load":"verify 1","shape":"m=8 pooler","x":1.028,"ours_us":27.09,"base_us":27.85,"floor_pct":59.1},{"load":"verify 1","shape":"m=8 unpooler","x":1.019,"ours_us":8.46,"base_us":8.62,"floor_pct":44.5},{"load":"verify 1","shape":"m=8 router","x":2.919,"ours_us":9.84,"base_us":28.73,"floor_pct":38.4},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.093,"ours_us":19.81,"base_us":21.66,"floor_pct":52.6},{"load":"verify 16","shape":"m=128 o","x":0.974,"ours_us":10.18,"base_us":9.91,"floor_pct":58.8},{"load":"verify 16","shape":"m=128 pooler","x":1.007,"ours_us":29.84,"base_us":30.03,"floor_pct":55.8},{"load":"verify 16","shape":"m=128 unpooler","x":0.897,"ours_us":29.79,"base_us":26.73,"floor_pct":55.9},{"load":"verify 16","shape":"m=128 router","x":2.332,"ours_us":13.3,"base_us":31.02,"floor_pct":31.7},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.169,"ours_us":21.96,"base_us":25.68,"floor_pct":58.0},{"load":"verify 32","shape":"m=256 o","x":0.987,"ours_us":11.27,"base_us":11.13,"floor_pct":60.1},{"load":"verify 32","shape":"m=256 pooler","x":1.064,"ours_us":33.46,"base_us":35.61,"floor_pct":63.0},{"load":"verify 32","shape":"m=256 unpooler","x":0.98,"ours_us":31.16,"base_us":30.55,"floor_pct":67.7},{"load":"verify 32","shape":"m=256 router","x":3.396,"ours_us":14.38,"base_us":48.84,"floor_pct":32.5}]},{"step":4,"x":1.2222079051187993,"lo":1.2164036465044776,"hi":1.2299398970769224,"pct":56.18046785535075,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.018,"ours_us":18.38,"base_us":18.71,"floor_pct":53.7},{"load":"decode 1","shape":"m=1 o","x":1.043,"ours_us":8.99,"base_us":9.38,"floor_pct":61.1},{"load":"decode 1","shape":"m=1 pooler","x":1.008,"ours_us":27.59,"base_us":27.82,"floor_pct":57.9},{"load":"decode 1","shape":"m=1 unpooler","x":1.049,"ours_us":8.38,"base_us":8.79,"floor_pct":44.7},{"load":"decode 1","shape":"m=1 router","x":2.168,"ours_us":5.75,"base_us":12.46,"floor_pct":65.3},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.008,"ours_us":18.41,"base_us":18.55,"floor_pct":54.0},{"load":"decode 16","shape":"m=16 o","x":1.033,"ours_us":9.3,"base_us":9.61,"floor_pct":59.7},{"load":"decode 16","shape":"m=16 pooler","x":0.995,"ours_us":28.55,"base_us":28.4,"floor_pct":56.3},{"load":"decode 16","shape":"m=16 unpooler","x":0.998,"ours_us":25.54,"base_us":25.48,"floor_pct":62.9},{"load":"decode 16","shape":"m=16 router","x":3.133,"ours_us":8.95,"base_us":28.05,"floor_pct":42.5},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.994,"ours_us":19.08,"base_us":18.97,"floor_pct":52.4},{"load":"decode 32","shape":"m=32 o","x":1.05,"ours_us":9.27,"base_us":9.73,"floor_pct":60.6},{"load":"decode 32","shape":"m=32 pooler","x":1.021,"ours_us":28.21,"base_us":28.8,"floor_pct":57.2},{"load":"decode 32","shape":"m=32 unpooler","x":1.012,"ours_us":25.55,"base_us":25.85,"floor_pct":63.2},{"load":"decode 32","shape":"m=32 router","x":2.218,"ours_us":9.68,"base_us":21.47,"floor_pct":39.9},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.022,"ours_us":18.34,"base_us":18.74,"floor_pct":54.0},{"load":"verify 1","shape":"m=8 o","x":1.032,"ours_us":9.5,"base_us":9.8,"floor_pct":58.2},{"load":"verify 1","shape":"m=8 pooler","x":1.011,"ours_us":27.49,"base_us":27.8,"floor_pct":58.3},{"load":"verify 1","shape":"m=8 unpooler","x":1.077,"ours_us":8.44,"base_us":9.09,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 router","x":3.156,"ours_us":9.0,"base_us":28.4,"floor_pct":42.0},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.055,"ours_us":20.75,"base_us":21.89,"floor_pct":50.2},{"load":"verify 16","shape":"m=128 o","x":1.024,"ours_us":9.97,"base_us":10.21,"floor_pct":60.1},{"load":"verify 16","shape":"m=128 pooler","x":0.994,"ours_us":29.77,"base_us":29.6,"floor_pct":55.9},{"load":"verify 16","shape":"m=128 unpooler","x":1.015,"ours_us":26.79,"base_us":27.19,"floor_pct":62.1},{"load":"verify 16","shape":"m=128 router","x":2.812,"ours_us":11.1,"base_us":31.21,"floor_pct":38.0},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.187,"ours_us":21.83,"base_us":25.9,"floor_pct":58.4},{"load":"verify 32","shape":"m=256 o","x":0.991,"ours_us":11.43,"base_us":11.33,"floor_pct":59.2},{"load":"verify 32","shape":"m=256 pooler","x":1.064,"ours_us":33.73,"base_us":35.89,"floor_pct":62.5},{"load":"verify 32","shape":"m=256 unpooler","x":1.019,"ours_us":29.98,"base_us":30.54,"floor_pct":70.3},{"load":"verify 32","shape":"m=256 router","x":3.387,"ours_us":14.4,"base_us":48.78,"floor_pct":32.5}]},{"step":5,"x":1.2473285278319683,"lo":1.2433581254375645,"hi":1.252906872212432,"pct":57.765131220258354,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.176,"ours_us":15.57,"base_us":18.32,"floor_pct":63.4},{"load":"decode 1","shape":"m=1 o","x":1.085,"ours_us":8.72,"base_us":9.46,"floor_pct":63.0},{"load":"decode 1","shape":"m=1 pooler","x":1.057,"ours_us":26.27,"base_us":27.78,"floor_pct":60.9},{"load":"decode 1","shape":"m=1 unpooler","x":1.317,"ours_us":6.83,"base_us":9.0,"floor_pct":54.9},{"load":"decode 1","shape":"m=1 router","x":2.055,"ours_us":5.96,"base_us":12.25,"floor_pct":62.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.025,"ours_us":17.95,"base_us":18.4,"floor_pct":55.3},{"load":"decode 16","shape":"m=16 o","x":0.974,"ours_us":9.75,"base_us":9.5,"floor_pct":57.0},{"load":"decode 16","shape":"m=16 pooler","x":1.045,"ours_us":26.32,"base_us":27.5,"floor_pct":61.0},{"load":"decode 16","shape":"m=16 unpooler","x":0.992,"ours_us":25.19,"base_us":25.0,"floor_pct":63.8},{"load":"decode 16","shape":"m=16 router","x":3.132,"ours_us":9.18,"base_us":28.75,"floor_pct":41.5},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.036,"ours_us":18.05,"base_us":18.71,"floor_pct":55.4},{"load":"decode 32","shape":"m=32 o","x":1.001,"ours_us":9.58,"base_us":9.59,"floor_pct":58.6},{"load":"decode 32","shape":"m=32 pooler","x":1.052,"ours_us":27.01,"base_us":28.42,"floor_pct":59.8},{"load":"decode 32","shape":"m=32 unpooler","x":1.001,"ours_us":25.49,"base_us":25.52,"floor_pct":63.3},{"load":"decode 32","shape":"m=32 router","x":2.214,"ours_us":9.83,"base_us":21.78,"floor_pct":39.3},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.01,"ours_us":18.11,"base_us":18.3,"floor_pct":54.7},{"load":"verify 1","shape":"m=8 o","x":0.997,"ours_us":9.52,"base_us":9.5,"floor_pct":58.0},{"load":"verify 1","shape":"m=8 pooler","x":1.045,"ours_us":26.51,"base_us":27.7,"floor_pct":60.4},{"load":"verify 1","shape":"m=8 unpooler","x":1.088,"ours_us":8.4,"base_us":9.14,"floor_pct":44.8},{"load":"verify 1","shape":"m=8 router","x":3.144,"ours_us":9.08,"base_us":28.54,"floor_pct":41.6},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.112,"ours_us":19.67,"base_us":21.86,"floor_pct":53.0},{"load":"verify 16","shape":"m=128 o","x":0.997,"ours_us":10.05,"base_us":10.02,"floor_pct":59.6},{"load":"verify 16","shape":"m=128 pooler","x":0.974,"ours_us":29.37,"base_us":28.6,"floor_pct":56.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.988,"ours_us":26.31,"base_us":25.98,"floor_pct":63.2},{"load":"verify 16","shape":"m=128 router","x":2.757,"ours_us":11.46,"base_us":31.58,"floor_pct":36.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.199,"ours_us":21.6,"base_us":25.91,"floor_pct":59.0},{"load":"verify 32","shape":"m=256 o","x":1.012,"ours_us":11.1,"base_us":11.23,"floor_pct":61.0},{"load":"verify 32","shape":"m=256 pooler","x":1.068,"ours_us":33.33,"base_us":35.61,"floor_pct":63.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.997,"ours_us":30.48,"base_us":30.38,"floor_pct":69.2},{"load":"verify 32","shape":"m=256 router","x":3.385,"ours_us":14.57,"base_us":49.31,"floor_pct":32.1}]},{"step":6,"x":1.2655598166140394,"lo":1.2557470956679733,"hi":1.273046112129908,"pct":61.202200050705095,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.183,"ours_us":15.76,"base_us":18.65,"floor_pct":62.6},{"load":"decode 1","shape":"m=1 o","x":1.091,"ours_us":8.62,"base_us":9.41,"floor_pct":63.8},{"load":"decode 1","shape":"m=1 pooler","x":1.053,"ours_us":26.42,"base_us":27.82,"floor_pct":60.5},{"load":"decode 1","shape":"m=1 unpooler","x":1.296,"ours_us":6.68,"base_us":8.66,"floor_pct":56.1},{"load":"decode 1","shape":"m=1 router","x":2.068,"ours_us":5.78,"base_us":11.96,"floor_pct":64.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.013,"ours_us":18.21,"base_us":18.44,"floor_pct":54.6},{"load":"decode 16","shape":"m=16 o","x":1.009,"ours_us":9.31,"base_us":9.39,"floor_pct":59.7},{"load":"decode 16","shape":"m=16 pooler","x":1.032,"ours_us":26.76,"base_us":27.62,"floor_pct":60.0},{"load":"decode 16","shape":"m=16 unpooler","x":0.998,"ours_us":25.26,"base_us":25.22,"floor_pct":63.6},{"load":"decode 16","shape":"m=16 router","x":3.121,"ours_us":9.08,"base_us":28.35,"floor_pct":41.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.033,"ours_us":18.01,"base_us":18.61,"floor_pct":55.5},{"load":"decode 32","shape":"m=32 o","x":0.977,"ours_us":9.59,"base_us":9.38,"floor_pct":58.6},{"load":"decode 32","shape":"m=32 pooler","x":1.053,"ours_us":27.19,"base_us":28.62,"floor_pct":59.4},{"load":"decode 32","shape":"m=32 unpooler","x":1.004,"ours_us":25.33,"base_us":25.42,"floor_pct":63.7},{"load":"decode 32","shape":"m=32 router","x":2.214,"ours_us":9.82,"base_us":21.75,"floor_pct":39.3},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.029,"ours_us":17.88,"base_us":18.39,"floor_pct":55.4},{"load":"verify 1","shape":"m=8 o","x":1.004,"ours_us":9.32,"base_us":9.35,"floor_pct":59.3},{"load":"verify 1","shape":"m=8 pooler","x":1.08,"ours_us":15.71,"base_us":16.97,"floor_pct":102.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.147,"ours_us":8.25,"base_us":9.46,"floor_pct":45.6},{"load":"verify 1","shape":"m=8 router","x":3.16,"ours_us":8.96,"base_us":28.3,"floor_pct":42.2},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.089,"ours_us":19.08,"base_us":20.78,"floor_pct":54.6},{"load":"verify 16","shape":"m=128 o","x":1.053,"ours_us":7.61,"base_us":8.02,"floor_pct":78.6},{"load":"verify 16","shape":"m=128 pooler","x":0.966,"ours_us":29.98,"base_us":28.97,"floor_pct":55.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.985,"ours_us":26.82,"base_us":26.42,"floor_pct":62.0},{"load":"verify 16","shape":"m=128 router","x":2.907,"ours_us":10.91,"base_us":31.72,"floor_pct":38.6},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.18,"ours_us":21.69,"base_us":25.59,"floor_pct":58.7},{"load":"verify 32","shape":"m=256 o","x":0.981,"ours_us":11.36,"base_us":11.14,"floor_pct":59.6},{"load":"verify 32","shape":"m=256 pooler","x":1.072,"ours_us":26.4,"base_us":28.3,"floor_pct":79.9},{"load":"verify 32","shape":"m=256 unpooler","x":1.0,"ours_us":30.27,"base_us":30.28,"floor_pct":69.7},{"load":"verify 32","shape":"m=256 router","x":6.645,"ours_us":6.17,"base_us":40.99,"floor_pct":75.8}]}],"confirm":{"step":null,"x":1.2463126771790565,"lo":1.2394879065403317,"hi":1.2511950920311545,"pct":57.94576814575671,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.2,"ours_us":15.4,"base_us":18.49,"floor_pct":64.1},{"load":"decode 1","shape":"m=1 o","x":1.066,"ours_us":8.83,"base_us":9.41,"floor_pct":62.3},{"load":"decode 1","shape":"m=1 pooler","x":1.04,"ours_us":26.4,"base_us":27.45,"floor_pct":60.5},{"load":"decode 1","shape":"m=1 unpooler","x":1.307,"ours_us":6.61,"base_us":8.64,"floor_pct":56.7},{"load":"decode 1","shape":"m=1 router","x":2.0,"ours_us":6.03,"base_us":12.06,"floor_pct":62.2},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.998,"ours_us":18.3,"base_us":18.27,"floor_pct":54.3},{"load":"decode 16","shape":"m=16 o","x":0.997,"ours_us":9.38,"base_us":9.35,"floor_pct":59.2},{"load":"decode 16","shape":"m=16 pooler","x":1.043,"ours_us":26.25,"base_us":27.38,"floor_pct":61.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.01,"ours_us":24.62,"base_us":24.87,"floor_pct":65.2},{"load":"decode 16","shape":"m=16 router","x":3.126,"ours_us":9.22,"base_us":28.81,"floor_pct":41.3},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.034,"ours_us":18.1,"base_us":18.72,"floor_pct":55.3},{"load":"decode 32","shape":"m=32 o","x":0.988,"ours_us":9.51,"base_us":9.39,"floor_pct":59.1},{"load":"decode 32","shape":"m=32 pooler","x":1.036,"ours_us":27.1,"base_us":28.08,"floor_pct":59.6},{"load":"decode 32","shape":"m=32 unpooler","x":1.008,"ours_us":25.24,"base_us":25.44,"floor_pct":64.0},{"load":"decode 32","shape":"m=32 router","x":2.191,"ours_us":9.9,"base_us":21.7,"floor_pct":39.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.033,"ours_us":18.12,"base_us":18.71,"floor_pct":54.6},{"load":"verify 1","shape":"m=8 o","x":0.995,"ours_us":9.38,"base_us":9.33,"floor_pct":58.9},{"load":"verify 1","shape":"m=8 pooler","x":1.061,"ours_us":26.25,"base_us":27.86,"floor_pct":61.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.075,"ours_us":8.2,"base_us":8.82,"floor_pct":45.9},{"load":"verify 1","shape":"m=8 router","x":3.146,"ours_us":9.07,"base_us":28.54,"floor_pct":41.6},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.106,"ours_us":19.56,"base_us":21.64,"floor_pct":53.3},{"load":"verify 16","shape":"m=128 o","x":0.988,"ours_us":10.14,"base_us":10.02,"floor_pct":59.0},{"load":"verify 16","shape":"m=128 pooler","x":0.963,"ours_us":29.91,"base_us":28.79,"floor_pct":55.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.968,"ours_us":26.98,"base_us":26.13,"floor_pct":61.7},{"load":"verify 16","shape":"m=128 router","x":2.829,"ours_us":11.15,"base_us":31.54,"floor_pct":37.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.217,"ours_us":21.5,"base_us":26.17,"floor_pct":59.2},{"load":"verify 32","shape":"m=256 o","x":1.002,"ours_us":11.09,"base_us":11.11,"floor_pct":61.1},{"load":"verify 32","shape":"m=256 pooler","x":1.069,"ours_us":33.07,"base_us":35.34,"floor_pct":63.8},{"load":"verify 32","shape":"m=256 unpooler","x":0.996,"ours_us":30.19,"base_us":30.09,"floor_pct":69.8},{"load":"verify 32","shape":"m=256 router","x":3.466,"ours_us":14.17,"base_us":49.12,"floor_pct":33.0}]},"src":"bench/product/runs/learnability/argon-bf16-v2-20261102.jsonl"},{"id":"bf16-A","model":"Claude Opus 5.5","harness":"Claude Code","family":"dense_proj_bf16","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":50.75850516569654},"T2":{"x":1.4535347788908006,"pct":67.07204780258934},"T3":{"x":1.9070695577816008,"pct":88.0},"T4":{"x":1.9504120477311826,"pct":90.0}},"steps":[{"step":1,"x":1.0039254597515677,"lo":1.0004475550768672,"hi":1.0067555416800575,"pct":46.137904419574646,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.969,"ours_us":19.1,"base_us":18.51,"floor_pct":51.7},{"load":"decode 1","shape":"m=1 o","x":0.79,"ours_us":12.01,"base_us":9.49,"floor_pct":45.8},{"load":"decode 1","shape":"m=1 pooler","x":1.009,"ours_us":27.79,"base_us":28.04,"floor_pct":57.5},{"load":"decode 1","shape":"m=1 unpooler","x":0.972,"ours_us":9.09,"base_us":8.83,"floor_pct":41.3},{"load":"decode 1","shape":"m=1 router","x":1.607,"ours_us":7.63,"base_us":12.26,"floor_pct":49.1},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.969,"ours_us":19.48,"base_us":18.88,"floor_pct":51.0},{"load":"decode 16","shape":"m=16 o","x":0.775,"ours_us":12.2,"base_us":9.46,"floor_pct":45.5},{"load":"decode 16","shape":"m=16 pooler","x":0.996,"ours_us":27.79,"base_us":27.68,"floor_pct":57.8},{"load":"decode 16","shape":"m=16 unpooler","x":0.898,"ours_us":28.3,"base_us":25.4,"floor_pct":56.8},{"load":"decode 16","shape":"m=16 router","x":3.736,"ours_us":7.68,"base_us":28.69,"floor_pct":49.6},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.669,"ours_us":28.22,"base_us":18.86,"floor_pct":35.5},{"load":"decode 32","shape":"m=32 o","x":0.625,"ours_us":15.69,"base_us":9.81,"floor_pct":35.8},{"load":"decode 32","shape":"m=32 pooler","x":0.751,"ours_us":38.36,"base_us":28.8,"floor_pct":42.1},{"load":"decode 32","shape":"m=32 unpooler","x":0.634,"ours_us":39.98,"base_us":25.36,"floor_pct":40.4},{"load":"decode 32","shape":"m=32 router","x":2.497,"ours_us":8.78,"base_us":21.91,"floor_pct":44.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.968,"ours_us":19.32,"base_us":18.7,"floor_pct":51.3},{"load":"verify 1","shape":"m=8 o","x":0.806,"ours_us":11.83,"base_us":9.54,"floor_pct":46.7},{"load":"verify 1","shape":"m=8 pooler","x":0.999,"ours_us":27.87,"base_us":27.82,"floor_pct":57.5},{"load":"verify 1","shape":"m=8 unpooler","x":0.99,"ours_us":9.09,"base_us":9.0,"floor_pct":41.4},{"load":"verify 1","shape":"m=8 router","x":3.736,"ours_us":7.69,"base_us":28.74,"floor_pct":49.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.805,"ours_us":27.61,"base_us":22.23,"floor_pct":37.8},{"load":"verify 16","shape":"m=128 o","x":0.596,"ours_us":16.91,"base_us":10.07,"floor_pct":35.4},{"load":"verify 16","shape":"m=128 pooler","x":0.875,"ours_us":33.37,"base_us":29.21,"floor_pct":49.9},{"load":"verify 16","shape":"m=128 unpooler","x":0.783,"ours_us":33.91,"base_us":26.55,"floor_pct":49.1},{"load":"verify 16","shape":"m=128 router","x":2.409,"ours_us":13.17,"base_us":31.72,"floor_pct":32.0},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.697,"ours_us":37.79,"base_us":26.32,"floor_pct":33.7},{"load":"verify 32","shape":"m=256 o","x":0.49,"ours_us":23.13,"base_us":11.33,"floor_pct":29.3},{"load":"verify 32","shape":"m=256 pooler","x":0.857,"ours_us":42.24,"base_us":36.19,"floor_pct":49.9},{"load":"verify 32","shape":"m=256 unpooler","x":0.654,"ours_us":46.25,"base_us":30.26,"floor_pct":45.6},{"load":"verify 32","shape":"m=256 router","x":2.802,"ours_us":17.52,"base_us":49.09,"floor_pct":26.7}]},{"step":2,"x":1.032741648436409,"lo":1.0294165218077636,"hi":1.036834282108699,"pct":47.95129471260756,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.628,"ours_us":29.23,"base_us":18.35,"floor_pct":33.8},{"load":"decode 1","shape":"m=1 o","x":0.665,"ours_us":14.04,"base_us":9.34,"floor_pct":39.2},{"load":"decode 1","shape":"m=1 pooler","x":0.694,"ours_us":40.13,"base_us":27.85,"floor_pct":39.8},{"load":"decode 1","shape":"m=1 unpooler","x":0.708,"ours_us":12.12,"base_us":8.58,"floor_pct":30.9},{"load":"decode 1","shape":"m=1 router","x":1.577,"ours_us":7.75,"base_us":12.22,"floor_pct":48.4},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.75,"ours_us":24.43,"base_us":18.33,"floor_pct":40.7},{"load":"decode 16","shape":"m=16 o","x":0.889,"ours_us":10.64,"base_us":9.46,"floor_pct":52.2},{"load":"decode 16","shape":"m=16 pooler","x":1.026,"ours_us":27.17,"base_us":27.86,"floor_pct":59.1},{"load":"decode 16","shape":"m=16 unpooler","x":0.938,"ours_us":26.96,"base_us":25.3,"floor_pct":59.6},{"load":"decode 16","shape":"m=16 router","x":3.673,"ours_us":7.74,"base_us":28.42,"floor_pct":49.2},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.71,"ours_us":26.08,"base_us":18.53,"floor_pct":38.4},{"load":"decode 32","shape":"m=32 o","x":0.744,"ours_us":12.88,"base_us":9.58,"floor_pct":43.6},{"load":"decode 32","shape":"m=32 pooler","x":0.998,"ours_us":28.61,"base_us":28.54,"floor_pct":56.4},{"load":"decode 32","shape":"m=32 unpooler","x":0.71,"ours_us":35.7,"base_us":25.35,"floor_pct":45.2},{"load":"decode 32","shape":"m=32 router","x":2.416,"ours_us":8.91,"base_us":21.54,"floor_pct":43.3},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.735,"ours_us":25.11,"base_us":18.45,"floor_pct":39.4},{"load":"verify 1","shape":"m=8 o","x":0.834,"ours_us":11.28,"base_us":9.4,"floor_pct":49.0},{"load":"verify 1","shape":"m=8 pooler","x":0.869,"ours_us":32.05,"base_us":27.84,"floor_pct":50.0},{"load":"verify 1","shape":"m=8 unpooler","x":0.786,"ours_us":11.08,"base_us":8.7,"floor_pct":34.0},{"load":"verify 1","shape":"m=8 router","x":3.633,"ours_us":7.83,"base_us":28.44,"floor_pct":48.2},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.961,"ours_us":22.38,"base_us":21.51,"floor_pct":46.6},{"load":"verify 16","shape":"m=128 o","x":0.991,"ours_us":10.1,"base_us":10.0,"floor_pct":59.3},{"load":"verify 16","shape":"m=128 pooler","x":0.933,"ours_us":30.03,"base_us":28.02,"floor_pct":55.4},{"load":"verify 16","shape":"m=128 unpooler","x":1.008,"ours_us":26.76,"base_us":26.97,"floor_pct":62.2},{"load":"verify 16","shape":"m=128 router","x":2.384,"ours_us":13.12,"base_us":31.28,"floor_pct":32.1},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.995,"ours_us":25.91,"base_us":25.78,"floor_pct":49.2},{"load":"verify 32","shape":"m=256 o","x":1.003,"ours_us":11.12,"base_us":11.15,"floor_pct":60.9},{"load":"verify 32","shape":"m=256 pooler","x":1.002,"ours_us":35.78,"base_us":35.84,"floor_pct":58.9},{"load":"verify 32","shape":"m=256 unpooler","x":0.998,"ours_us":30.54,"base_us":30.48,"floor_pct":69.0},{"load":"verify 32","shape":"m=256 router","x":3.321,"ours_us":14.78,"base_us":49.09,"floor_pct":31.6}]},{"step":3,"x":1.1026278366906546,"lo":1.0994195915613383,"hi":1.1069443931236551,"pct":50.864453670131084,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.011,"ours_us":18.26,"base_us":18.46,"floor_pct":54.0},{"load":"decode 1","shape":"m=1 o","x":0.781,"ours_us":12.02,"base_us":9.39,"floor_pct":45.7},{"load":"decode 1","shape":"m=1 pooler","x":0.959,"ours_us":29.01,"base_us":27.81,"floor_pct":55.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.028,"ours_us":8.58,"base_us":8.82,"floor_pct":43.7},{"load":"decode 1","shape":"m=1 router","x":1.525,"ours_us":7.98,"base_us":12.17,"floor_pct":47.0},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.997,"ours_us":18.7,"base_us":18.64,"floor_pct":53.1},{"load":"decode 16","shape":"m=16 o","x":0.821,"ours_us":11.65,"base_us":9.56,"floor_pct":47.7},{"load":"decode 16","shape":"m=16 pooler","x":0.96,"ours_us":28.77,"base_us":27.62,"floor_pct":55.8},{"load":"decode 16","shape":"m=16 unpooler","x":0.884,"ours_us":28.39,"base_us":25.09,"floor_pct":56.6},{"load":"decode 16","shape":"m=16 router","x":3.368,"ours_us":8.47,"base_us":28.54,"floor_pct":44.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.856,"ours_us":22.11,"base_us":18.93,"floor_pct":45.3},{"load":"decode 32","shape":"m=32 o","x":0.751,"ours_us":12.64,"base_us":9.5,"floor_pct":44.4},{"load":"decode 32","shape":"m=32 pooler","x":0.855,"ours_us":33.52,"base_us":28.67,"floor_pct":48.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.804,"ours_us":31.64,"base_us":25.43,"floor_pct":51.0},{"load":"decode 32","shape":"m=32 router","x":2.32,"ours_us":9.38,"base_us":21.75,"floor_pct":41.2},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.003,"ours_us":18.42,"base_us":18.47,"floor_pct":53.7},{"load":"verify 1","shape":"m=8 o","x":0.816,"ours_us":11.54,"base_us":9.42,"floor_pct":47.9},{"load":"verify 1","shape":"m=8 pooler","x":0.95,"ours_us":29.13,"base_us":27.66,"floor_pct":55.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.05,"ours_us":8.5,"base_us":8.92,"floor_pct":44.3},{"load":"verify 1","shape":"m=8 router","x":3.378,"ours_us":8.42,"base_us":28.42,"floor_pct":44.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.003,"ours_us":22.47,"base_us":22.54,"floor_pct":46.4},{"load":"verify 16","shape":"m=128 o","x":0.995,"ours_us":10.14,"base_us":10.09,"floor_pct":59.1},{"load":"verify 16","shape":"m=128 pooler","x":0.98,"ours_us":29.76,"base_us":29.18,"floor_pct":55.9},{"load":"verify 16","shape":"m=128 unpooler","x":0.989,"ours_us":27.36,"base_us":27.05,"floor_pct":60.8},{"load":"verify 16","shape":"m=128 router","x":1.518,"ours_us":20.89,"base_us":31.71,"floor_pct":20.2},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.005,"ours_us":26.15,"base_us":26.3,"floor_pct":48.7},{"load":"verify 32","shape":"m=256 o","x":1.01,"ours_us":11.05,"base_us":11.17,"floor_pct":61.3},{"load":"verify 32","shape":"m=256 pooler","x":1.01,"ours_us":35.46,"base_us":35.82,"floor_pct":59.5},{"load":"verify 32","shape":"m=256 unpooler","x":1.008,"ours_us":30.34,"base_us":30.58,"floor_pct":69.5},{"load":"verify 32","shape":"m=256 router","x":2.273,"ours_us":21.53,"base_us":48.92,"floor_pct":21.7}]},{"step":4,"x":1.189094753555012,"lo":1.1836846440175015,"hi":1.1941468249856573,"pct":54.8355568534774,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.992,"ours_us":18.78,"base_us":18.62,"floor_pct":52.5},{"load":"decode 1","shape":"m=1 o","x":0.989,"ours_us":9.54,"base_us":9.44,"floor_pct":57.6},{"load":"decode 1","shape":"m=1 pooler","x":0.996,"ours_us":27.97,"base_us":27.85,"floor_pct":57.2},{"load":"decode 1","shape":"m=1 unpooler","x":1.001,"ours_us":8.87,"base_us":8.88,"floor_pct":42.3},{"load":"decode 1","shape":"m=1 router","x":1.548,"ours_us":7.86,"base_us":12.17,"floor_pct":47.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.993,"ours_us":18.58,"base_us":18.45,"floor_pct":53.5},{"load":"decode 16","shape":"m=16 o","x":0.974,"ours_us":9.75,"base_us":9.5,"floor_pct":57.0},{"load":"decode 16","shape":"m=16 pooler","x":0.977,"ours_us":28.6,"base_us":27.95,"floor_pct":56.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.009,"ours_us":25.0,"base_us":25.22,"floor_pct":64.3},{"load":"decode 16","shape":"m=16 router","x":3.587,"ours_us":7.97,"base_us":28.58,"floor_pct":47.8},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.006,"ours_us":18.81,"base_us":18.93,"floor_pct":53.2},{"load":"decode 32","shape":"m=32 o","x":1.005,"ours_us":9.54,"base_us":9.59,"floor_pct":58.9},{"load":"decode 32","shape":"m=32 pooler","x":0.999,"ours_us":28.73,"base_us":28.7,"floor_pct":56.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.996,"ours_us":25.25,"base_us":25.15,"floor_pct":63.9},{"load":"decode 32","shape":"m=32 router","x":2.467,"ours_us":8.84,"base_us":21.81,"floor_pct":43.7},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.99,"ours_us":18.61,"base_us":18.42,"floor_pct":53.2},{"load":"verify 1","shape":"m=8 o","x":0.974,"ours_us":9.7,"base_us":9.45,"floor_pct":57.0},{"load":"verify 1","shape":"m=8 pooler","x":0.995,"ours_us":28.06,"base_us":27.92,"floor_pct":57.1},{"load":"verify 1","shape":"m=8 unpooler","x":0.999,"ours_us":8.97,"base_us":8.96,"floor_pct":42.0},{"load":"verify 1","shape":"m=8 router","x":3.607,"ours_us":7.85,"base_us":28.31,"floor_pct":48.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.0,"ours_us":22.3,"base_us":22.3,"floor_pct":46.8},{"load":"verify 16","shape":"m=128 o","x":0.976,"ours_us":10.32,"base_us":10.06,"floor_pct":58.0},{"load":"verify 16","shape":"m=128 pooler","x":1.005,"ours_us":29.48,"base_us":29.64,"floor_pct":56.4},{"load":"verify 16","shape":"m=128 unpooler","x":0.978,"ours_us":27.24,"base_us":26.65,"floor_pct":61.1},{"load":"verify 16","shape":"m=128 router","x":2.392,"ours_us":13.27,"base_us":31.74,"floor_pct":31.7},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.996,"ours_us":25.99,"base_us":25.9,"floor_pct":49.0},{"load":"verify 32","shape":"m=256 o","x":1.02,"ours_us":11.07,"base_us":11.3,"floor_pct":61.2},{"load":"verify 32","shape":"m=256 pooler","x":0.988,"ours_us":36.0,"base_us":35.57,"floor_pct":58.6},{"load":"verify 32","shape":"m=256 unpooler","x":1.004,"ours_us":30.19,"base_us":30.31,"floor_pct":69.8},{"load":"verify 32","shape":"m=256 router","x":3.315,"ours_us":14.82,"base_us":49.14,"floor_pct":31.5}]},{"step":5,"x":1.1281870899630662,"lo":1.1245361878382996,"hi":1.1322792636447223,"pct":51.98746847442128,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.0,"ours_us":18.63,"base_us":18.62,"floor_pct":53.0},{"load":"decode 1","shape":"m=1 o","x":0.738,"ours_us":12.71,"base_us":9.38,"floor_pct":43.3},{"load":"decode 1","shape":"m=1 pooler","x":0.973,"ours_us":28.54,"base_us":27.77,"floor_pct":56.0},{"load":"decode 1","shape":"m=1 unpooler","x":1.016,"ours_us":8.76,"base_us":8.9,"floor_pct":42.8},{"load":"decode 1","shape":"m=1 router","x":1.605,"ours_us":7.55,"base_us":12.11,"floor_pct":49.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.001,"ours_us":18.42,"base_us":18.44,"floor_pct":53.9},{"load":"decode 16","shape":"m=16 o","x":0.825,"ours_us":11.71,"base_us":9.66,"floor_pct":47.5},{"load":"decode 16","shape":"m=16 pooler","x":0.976,"ours_us":28.5,"base_us":27.82,"floor_pct":56.4},{"load":"decode 16","shape":"m=16 unpooler","x":0.914,"ours_us":27.7,"base_us":25.3,"floor_pct":58.0},{"load":"decode 16","shape":"m=16 router","x":3.702,"ours_us":7.82,"base_us":28.96,"floor_pct":48.6},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.84,"ours_us":22.38,"base_us":18.79,"floor_pct":44.7},{"load":"decode 32","shape":"m=32 o","x":0.597,"ours_us":16.16,"base_us":9.64,"floor_pct":34.8},{"load":"decode 32","shape":"m=32 pooler","x":0.855,"ours_us":33.6,"base_us":28.74,"floor_pct":48.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.785,"ours_us":32.19,"base_us":25.26,"floor_pct":50.2},{"load":"decode 32","shape":"m=32 router","x":2.561,"ours_us":8.48,"base_us":21.7,"floor_pct":45.6},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.001,"ours_us":18.82,"base_us":18.84,"floor_pct":52.6},{"load":"verify 1","shape":"m=8 o","x":0.812,"ours_us":11.67,"base_us":9.48,"floor_pct":47.3},{"load":"verify 1","shape":"m=8 pooler","x":0.973,"ours_us":28.56,"base_us":27.8,"floor_pct":56.1},{"load":"verify 1","shape":"m=8 unpooler","x":1.042,"ours_us":8.61,"base_us":8.98,"floor_pct":43.7},{"load":"verify 1","shape":"m=8 router","x":3.7,"ours_us":7.7,"base_us":28.48,"floor_pct":49.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.006,"ours_us":22.14,"base_us":22.29,"floor_pct":47.1},{"load":"verify 16","shape":"m=128 o","x":1.036,"ours_us":9.78,"base_us":10.14,"floor_pct":61.2},{"load":"verify 16","shape":"m=128 pooler","x":0.969,"ours_us":29.92,"base_us":28.98,"floor_pct":55.6},{"load":"verify 16","shape":"m=128 unpooler","x":1.0,"ours_us":26.54,"base_us":26.54,"floor_pct":62.7},{"load":"verify 16","shape":"m=128 router","x":2.38,"ours_us":13.24,"base_us":31.51,"floor_pct":31.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.999,"ours_us":26.25,"base_us":26.22,"floor_pct":48.5},{"load":"verify 32","shape":"m=256 o","x":1.016,"ours_us":11.05,"base_us":11.22,"floor_pct":61.3},{"load":"verify 32","shape":"m=256 pooler","x":1.022,"ours_us":35.56,"base_us":36.36,"floor_pct":59.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.997,"ours_us":30.4,"base_us":30.32,"floor_pct":69.4},{"load":"verify 32","shape":"m=256 router","x":3.33,"ours_us":14.67,"base_us":48.85,"floor_pct":31.9}]},{"step":6,"x":1.186158435555303,"lo":1.181476293787015,"hi":1.1909497740034678,"pct":54.901735792568424,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.988,"ours_us":18.69,"base_us":18.46,"floor_pct":52.8},{"load":"decode 1","shape":"m=1 o","x":0.977,"ours_us":9.55,"base_us":9.34,"floor_pct":57.6},{"load":"decode 1","shape":"m=1 pooler","x":1.002,"ours_us":27.61,"base_us":27.67,"floor_pct":57.9},{"load":"decode 1","shape":"m=1 unpooler","x":0.99,"ours_us":8.85,"base_us":8.76,"floor_pct":42.4},{"load":"decode 1","shape":"m=1 router","x":1.568,"ours_us":7.67,"base_us":12.03,"floor_pct":48.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.983,"ours_us":18.75,"base_us":18.42,"floor_pct":53.0},{"load":"decode 16","shape":"m=16 o","x":0.989,"ours_us":9.6,"base_us":9.5,"floor_pct":57.9},{"load":"decode 16","shape":"m=16 pooler","x":0.991,"ours_us":27.95,"base_us":27.69,"floor_pct":57.5},{"load":"decode 16","shape":"m=16 unpooler","x":1.0,"ours_us":25.18,"base_us":25.18,"floor_pct":63.8},{"load":"decode 16","shape":"m=16 router","x":3.647,"ours_us":7.8,"base_us":28.46,"floor_pct":48.8},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.995,"ours_us":18.91,"base_us":18.82,"floor_pct":52.9},{"load":"decode 32","shape":"m=32 o","x":0.989,"ours_us":9.57,"base_us":9.46,"floor_pct":58.7},{"load":"decode 32","shape":"m=32 pooler","x":0.993,"ours_us":28.75,"base_us":28.55,"floor_pct":56.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.992,"ours_us":25.51,"base_us":25.3,"floor_pct":63.3},{"load":"decode 32","shape":"m=32 router","x":2.526,"ours_us":8.67,"base_us":21.9,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.997,"ours_us":18.66,"base_us":18.6,"floor_pct":53.0},{"load":"verify 1","shape":"m=8 o","x":0.975,"ours_us":9.63,"base_us":9.38,"floor_pct":57.4},{"load":"verify 1","shape":"m=8 pooler","x":0.991,"ours_us":27.84,"base_us":27.6,"floor_pct":57.5},{"load":"verify 1","shape":"m=8 unpooler","x":1.012,"ours_us":8.86,"base_us":8.97,"floor_pct":42.5},{"load":"verify 1","shape":"m=8 router","x":3.588,"ours_us":7.88,"base_us":28.28,"floor_pct":47.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.993,"ours_us":22.39,"base_us":22.25,"floor_pct":46.5},{"load":"verify 16","shape":"m=128 o","x":1.002,"ours_us":10.07,"base_us":10.09,"floor_pct":59.4},{"load":"verify 16","shape":"m=128 pooler","x":0.979,"ours_us":29.93,"base_us":29.31,"floor_pct":55.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.982,"ours_us":27.02,"base_us":26.54,"floor_pct":61.6},{"load":"verify 16","shape":"m=128 router","x":2.374,"ours_us":13.31,"base_us":31.59,"floor_pct":31.6},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.981,"ours_us":26.65,"base_us":26.14,"floor_pct":47.8},{"load":"verify 32","shape":"m=256 o","x":0.996,"ours_us":11.1,"base_us":11.06,"floor_pct":61.0},{"load":"verify 32","shape":"m=256 pooler","x":0.979,"ours_us":36.09,"base_us":35.34,"floor_pct":58.4},{"load":"verify 32","shape":"m=256 unpooler","x":1.007,"ours_us":30.4,"base_us":30.61,"floor_pct":69.4},{"load":"verify 32","shape":"m=256 router","x":3.295,"ours_us":14.9,"base_us":49.1,"floor_pct":31.4}]}],"confirm":{"step":null,"x":1.1960171254275809,"lo":1.19129136361941,"hi":1.2000155886855388,"pct":55.18912858116126,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.966,"ours_us":19.03,"base_us":18.38,"floor_pct":51.9},{"load":"decode 1","shape":"m=1 o","x":1.016,"ours_us":9.24,"base_us":9.39,"floor_pct":59.5},{"load":"decode 1","shape":"m=1 pooler","x":1.0,"ours_us":28.02,"base_us":28.02,"floor_pct":57.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.008,"ours_us":8.82,"base_us":8.89,"floor_pct":42.5},{"load":"decode 1","shape":"m=1 router","x":1.595,"ours_us":7.57,"base_us":12.08,"floor_pct":49.5},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.991,"ours_us":18.66,"base_us":18.49,"floor_pct":53.2},{"load":"decode 16","shape":"m=16 o","x":1.003,"ours_us":9.51,"base_us":9.54,"floor_pct":58.4},{"load":"decode 16","shape":"m=16 pooler","x":0.977,"ours_us":28.41,"base_us":27.77,"floor_pct":56.5},{"load":"decode 16","shape":"m=16 unpooler","x":1.003,"ours_us":25.2,"base_us":25.28,"floor_pct":63.7},{"load":"decode 16","shape":"m=16 router","x":3.664,"ours_us":7.77,"base_us":28.46,"floor_pct":49.0},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.013,"ours_us":18.51,"base_us":18.74,"floor_pct":54.1},{"load":"decode 32","shape":"m=32 o","x":1.007,"ours_us":9.52,"base_us":9.58,"floor_pct":59.0},{"load":"decode 32","shape":"m=32 pooler","x":1.012,"ours_us":28.15,"base_us":28.5,"floor_pct":57.4},{"load":"decode 32","shape":"m=32 unpooler","x":1.006,"ours_us":25.25,"base_us":25.39,"floor_pct":63.9},{"load":"decode 32","shape":"m=32 router","x":2.523,"ours_us":8.65,"base_us":21.84,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.988,"ours_us":18.81,"base_us":18.58,"floor_pct":52.6},{"load":"verify 1","shape":"m=8 o","x":1.025,"ours_us":9.32,"base_us":9.56,"floor_pct":59.3},{"load":"verify 1","shape":"m=8 pooler","x":1.003,"ours_us":27.63,"base_us":27.73,"floor_pct":58.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.048,"ours_us":8.63,"base_us":9.05,"floor_pct":43.6},{"load":"verify 1","shape":"m=8 router","x":3.603,"ours_us":7.86,"base_us":28.3,"floor_pct":48.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.014,"ours_us":22.01,"base_us":22.33,"floor_pct":47.4},{"load":"verify 16","shape":"m=128 o","x":0.994,"ours_us":10.12,"base_us":10.06,"floor_pct":59.1},{"load":"verify 16","shape":"m=128 pooler","x":1.003,"ours_us":29.44,"base_us":29.54,"floor_pct":56.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.995,"ours_us":26.82,"base_us":26.7,"floor_pct":62.0},{"load":"verify 16","shape":"m=128 router","x":2.396,"ours_us":13.24,"base_us":31.73,"floor_pct":31.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.996,"ours_us":26.24,"base_us":26.14,"floor_pct":48.5},{"load":"verify 32","shape":"m=256 o","x":1.006,"ours_us":11.15,"base_us":11.22,"floor_pct":60.7},{"load":"verify 32","shape":"m=256 pooler","x":0.982,"ours_us":36.17,"base_us":35.51,"floor_pct":58.3},{"load":"verify 32","shape":"m=256 unpooler","x":1.0,"ours_us":30.33,"base_us":30.34,"floor_pct":69.5},{"load":"verify 32","shape":"m=256 router","x":3.336,"ours_us":14.69,"base_us":49.02,"floor_pct":31.8}]},"src":"bench/product/runs/learnability/bf16-A.jsonl"},{"id":"bf16-B","model":"Claude Opus 5.5","harness":"Claude Code","family":"dense_proj_bf16","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":50.745616920078966},"T2":{"x":1.4537769552831892,"pct":67.0661895091268},"T3":{"x":1.9075539105663786,"pct":88.0},"T4":{"x":1.9509074085337963,"pct":90.0}},"steps":[{"step":1,"x":1.030659253796212,"lo":1.0265650718073422,"hi":1.0343415945974217,"pct":47.389640936441815,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.879,"ours_us":21.6,"base_us":18.98,"floor_pct":45.7},{"load":"decode 1","shape":"m=1 o","x":0.931,"ours_us":10.51,"base_us":9.79,"floor_pct":52.3},{"load":"decode 1","shape":"m=1 pooler","x":0.934,"ours_us":30.03,"base_us":28.04,"floor_pct":53.2},{"load":"decode 1","shape":"m=1 unpooler","x":0.99,"ours_us":8.87,"base_us":8.78,"floor_pct":42.3},{"load":"decode 1","shape":"m=1 router","x":1.245,"ours_us":9.79,"base_us":12.19,"floor_pct":38.3},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.839,"ours_us":22.06,"base_us":18.5,"floor_pct":45.0},{"load":"decode 16","shape":"m=16 o","x":1.032,"ours_us":9.39,"base_us":9.69,"floor_pct":59.2},{"load":"decode 16","shape":"m=16 pooler","x":0.925,"ours_us":29.98,"base_us":27.74,"floor_pct":53.6},{"load":"decode 16","shape":"m=16 unpooler","x":0.855,"ours_us":29.41,"base_us":25.15,"floor_pct":54.6},{"load":"decode 16","shape":"m=16 router","x":2.8,"ours_us":10.23,"base_us":28.66,"floor_pct":37.2},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.889,"ours_us":21.03,"base_us":18.69,"floor_pct":47.6},{"load":"decode 32","shape":"m=32 o","x":0.787,"ours_us":12.14,"base_us":9.56,"floor_pct":46.3},{"load":"decode 32","shape":"m=32 pooler","x":0.986,"ours_us":28.96,"base_us":28.55,"floor_pct":55.8},{"load":"decode 32","shape":"m=32 unpooler","x":0.891,"ours_us":28.44,"base_us":25.34,"floor_pct":56.8},{"load":"decode 32","shape":"m=32 router","x":2.022,"ours_us":10.9,"base_us":22.04,"floor_pct":35.4},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.853,"ours_us":21.91,"base_us":18.68,"floor_pct":45.2},{"load":"verify 1","shape":"m=8 o","x":0.976,"ours_us":9.7,"base_us":9.46,"floor_pct":57.0},{"load":"verify 1","shape":"m=8 pooler","x":0.941,"ours_us":29.9,"base_us":28.14,"floor_pct":53.6},{"load":"verify 1","shape":"m=8 unpooler","x":1.015,"ours_us":8.84,"base_us":8.97,"floor_pct":42.6},{"load":"verify 1","shape":"m=8 router","x":2.792,"ours_us":10.18,"base_us":28.43,"floor_pct":37.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.825,"ours_us":27.47,"base_us":22.65,"floor_pct":38.0},{"load":"verify 16","shape":"m=128 o","x":0.707,"ours_us":14.28,"base_us":10.09,"floor_pct":41.9},{"load":"verify 16","shape":"m=128 pooler","x":0.9,"ours_us":32.43,"base_us":29.18,"floor_pct":51.3},{"load":"verify 16","shape":"m=128 unpooler","x":0.83,"ours_us":32.29,"base_us":26.81,"floor_pct":51.5},{"load":"verify 16","shape":"m=128 router","x":2.341,"ours_us":13.62,"base_us":31.89,"floor_pct":30.9},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.683,"ours_us":37.91,"base_us":25.91,"floor_pct":33.6},{"load":"verify 32","shape":"m=256 o","x":0.572,"ours_us":19.74,"base_us":11.28,"floor_pct":34.3},{"load":"verify 32","shape":"m=256 pooler","x":0.839,"ours_us":42.3,"base_us":35.5,"floor_pct":49.9},{"load":"verify 32","shape":"m=256 unpooler","x":0.705,"ours_us":42.77,"base_us":30.15,"floor_pct":49.3},{"load":"verify 32","shape":"m=256 router","x":3.027,"ours_us":16.21,"base_us":49.08,"floor_pct":28.8}]},{"step":2,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"GRADER_BUG_REFUNDED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.889,"ours_us":21.2,"base_us":18.85,"floor_pct":46.6},{"load":"decode 1","shape":"m=1 o","x":0.933,"ours_us":9.96,"base_us":9.3,"floor_pct":55.2},{"load":"decode 1","shape":"m=1 pooler","x":1.019,"ours_us":27.37,"base_us":27.9,"floor_pct":58.4},{"load":"decode 1","shape":"m=1 unpooler","x":1.039,"ours_us":8.5,"base_us":8.83,"floor_pct":44.1},{"load":"decode 1","shape":"m=1 router","x":1.054,"ours_us":11.47,"base_us":12.09,"floor_pct":32.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.869,"ours_us":21.34,"base_us":18.54,"floor_pct":46.5},{"load":"decode 16","shape":"m=16 o","x":1.005,"ours_us":9.45,"base_us":9.5,"floor_pct":58.8},{"load":"decode 16","shape":"m=16 pooler","x":1.042,"ours_us":27.06,"base_us":28.21,"floor_pct":59.4},{"load":"decode 16","shape":"m=16 unpooler","x":0.999,"ours_us":25.7,"base_us":25.67,"floor_pct":62.5},{"load":"decode 16","shape":"m=16 router","x":2.359,"ours_us":12.17,"base_us":28.71,"floor_pct":31.3},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.855,"ours_us":21.76,"base_us":18.61,"floor_pct":46.0},{"load":"decode 32","shape":"m=32 o","x":0.935,"ours_us":10.14,"base_us":9.48,"floor_pct":55.4},{"load":"decode 32","shape":"m=32 pooler","x":1.023,"ours_us":27.84,"base_us":28.5,"floor_pct":58.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.962,"ours_us":26.24,"base_us":25.23,"floor_pct":61.5},{"load":"decode 32","shape":"m=32 router","x":1.737,"ours_us":12.6,"base_us":21.9,"floor_pct":30.7},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.872,"ours_us":21.53,"base_us":18.77,"floor_pct":46.0},{"load":"verify 1","shape":"m=8 o","x":0.978,"ours_us":9.66,"base_us":9.45,"floor_pct":57.2},{"load":"verify 1","shape":"m=8 pooler","x":1.017,"ours_us":27.44,"base_us":27.9,"floor_pct":58.4},{"load":"verify 1","shape":"m=8 unpooler","x":1.028,"ours_us":8.65,"base_us":8.9,"floor_pct":43.5},{"load":"verify 1","shape":"m=8 router","x":2.356,"ours_us":12.1,"base_us":28.52,"floor_pct":31.2},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.868,"ours_us":25.39,"base_us":22.02,"floor_pct":41.1},{"load":"verify 16","shape":"m=128 o","x":0.899,"ours_us":11.23,"base_us":10.1,"floor_pct":53.3},{"load":"verify 16","shape":"m=128 pooler","x":0.811,"ours_us":35.3,"base_us":28.62,"floor_pct":47.1},{"load":"verify 16","shape":"m=128 unpooler","x":0.804,"ours_us":33.21,"base_us":26.7,"floor_pct":50.1},{"load":"verify 16","shape":"m=128 router","x":0.511,"ours_us":62.32,"base_us":31.85,"floor_pct":6.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.766,"ours_us":33.63,"base_us":25.76,"floor_pct":37.9},{"load":"verify 32","shape":"m=256 o","x":0.822,"ours_us":13.51,"base_us":11.1,"floor_pct":50.1},{"load":"verify 32","shape":"m=256 pooler","x":0.753,"ours_us":46.62,"base_us":35.09,"floor_pct":45.2}]},{"step":3,"x":1.0743784622576087,"lo":1.069118162973861,"hi":1.0777371380506438,"pct":49.73949983798279,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.806,"ours_us":23.03,"base_us":18.55,"floor_pct":42.9},{"load":"decode 1","shape":"m=1 o","x":0.864,"ours_us":10.76,"base_us":9.3,"floor_pct":51.1},{"load":"decode 1","shape":"m=1 pooler","x":1.013,"ours_us":27.84,"base_us":28.21,"floor_pct":57.4},{"load":"decode 1","shape":"m=1 unpooler","x":0.916,"ours_us":9.53,"base_us":8.74,"floor_pct":39.3},{"load":"decode 1","shape":"m=1 router","x":1.276,"ours_us":9.45,"base_us":12.06,"floor_pct":39.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.997,"ours_us":18.5,"base_us":18.44,"floor_pct":53.7},{"load":"decode 16","shape":"m=16 o","x":0.88,"ours_us":10.76,"base_us":9.47,"floor_pct":51.7},{"load":"decode 16","shape":"m=16 pooler","x":1.015,"ours_us":27.34,"base_us":27.74,"floor_pct":58.8},{"load":"decode 16","shape":"m=16 unpooler","x":0.94,"ours_us":26.84,"base_us":25.24,"floor_pct":59.9},{"load":"decode 16","shape":"m=16 router","x":2.912,"ours_us":9.78,"base_us":28.48,"floor_pct":38.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.967,"ours_us":19.12,"base_us":18.5,"floor_pct":52.3},{"load":"decode 32","shape":"m=32 o","x":0.813,"ours_us":11.69,"base_us":9.5,"floor_pct":48.1},{"load":"decode 32","shape":"m=32 pooler","x":1.017,"ours_us":27.76,"base_us":28.23,"floor_pct":58.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.914,"ours_us":27.95,"base_us":25.55,"floor_pct":57.8},{"load":"decode 32","shape":"m=32 router","x":2.042,"ours_us":10.63,"base_us":21.7,"floor_pct":36.4},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.016,"ours_us":18.14,"base_us":18.43,"floor_pct":54.6},{"load":"verify 1","shape":"m=8 o","x":0.859,"ours_us":10.82,"base_us":9.3,"floor_pct":51.1},{"load":"verify 1","shape":"m=8 pooler","x":1.017,"ours_us":27.31,"base_us":27.79,"floor_pct":58.7},{"load":"verify 1","shape":"m=8 unpooler","x":0.922,"ours_us":9.47,"base_us":8.73,"floor_pct":39.8},{"load":"verify 1","shape":"m=8 router","x":2.907,"ours_us":9.74,"base_us":28.32,"floor_pct":38.8},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.99,"ours_us":22.46,"base_us":22.24,"floor_pct":46.4},{"load":"verify 16","shape":"m=128 o","x":0.852,"ours_us":11.65,"base_us":9.93,"floor_pct":51.4},{"load":"verify 16","shape":"m=128 pooler","x":0.738,"ours_us":39.81,"base_us":29.39,"floor_pct":41.8},{"load":"verify 16","shape":"m=128 unpooler","x":0.742,"ours_us":36.47,"base_us":27.05,"floor_pct":45.6},{"load":"verify 16","shape":"m=128 router","x":2.484,"ours_us":12.78,"base_us":31.74,"floor_pct":33.0},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.833,"ours_us":31.1,"base_us":25.9,"floor_pct":41.0},{"load":"verify 32","shape":"m=256 o","x":0.855,"ours_us":12.83,"base_us":10.98,"floor_pct":52.8},{"load":"verify 32","shape":"m=256 pooler","x":0.825,"ours_us":42.81,"base_us":35.33,"floor_pct":49.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.788,"ours_us":38.12,"base_us":30.04,"floor_pct":55.3},{"load":"verify 32","shape":"m=256 router","x":3.16,"ours_us":15.49,"base_us":48.94,"floor_pct":30.2}]},{"step":3,"x":1.1825872259387007,"lo":1.1783998448096016,"hi":1.1860017601733677,"pct":54.903853032670725,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.002,"ours_us":18.42,"base_us":18.46,"floor_pct":53.6},{"load":"decode 1","shape":"m=1 o","x":0.982,"ours_us":9.39,"base_us":9.22,"floor_pct":58.5},{"load":"decode 1","shape":"m=1 pooler","x":1.031,"ours_us":27.06,"base_us":27.9,"floor_pct":59.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.039,"ours_us":8.34,"base_us":8.67,"floor_pct":44.9},{"load":"decode 1","shape":"m=1 router","x":1.52,"ours_us":8.03,"base_us":12.21,"floor_pct":46.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.994,"ours_us":18.43,"base_us":18.32,"floor_pct":53.9},{"load":"decode 16","shape":"m=16 o","x":0.992,"ours_us":9.43,"base_us":9.36,"floor_pct":58.9},{"load":"decode 16","shape":"m=16 pooler","x":1.026,"ours_us":27.19,"base_us":27.89,"floor_pct":59.1},{"load":"decode 16","shape":"m=16 unpooler","x":1.006,"ours_us":24.96,"base_us":25.12,"floor_pct":64.4},{"load":"decode 16","shape":"m=16 router","x":3.446,"ours_us":8.23,"base_us":28.36,"floor_pct":46.2},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.992,"ours_us":18.81,"base_us":18.66,"floor_pct":53.2},{"load":"decode 32","shape":"m=32 o","x":1.003,"ours_us":9.38,"base_us":9.42,"floor_pct":59.9},{"load":"decode 32","shape":"m=32 pooler","x":0.991,"ours_us":28.35,"base_us":28.09,"floor_pct":56.9},{"load":"decode 32","shape":"m=32 unpooler","x":1.001,"ours_us":25.25,"base_us":25.26,"floor_pct":63.9},{"load":"decode 32","shape":"m=32 router","x":2.339,"ours_us":9.24,"base_us":21.62,"floor_pct":41.8},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.993,"ours_us":18.66,"base_us":18.52,"floor_pct":53.1},{"load":"verify 1","shape":"m=8 o","x":0.979,"ours_us":9.47,"base_us":9.27,"floor_pct":58.3},{"load":"verify 1","shape":"m=8 pooler","x":1.026,"ours_us":26.95,"base_us":27.66,"floor_pct":59.4},{"load":"verify 1","shape":"m=8 unpooler","x":1.025,"ours_us":8.48,"base_us":8.69,"floor_pct":44.4},{"load":"verify 1","shape":"m=8 router","x":3.493,"ours_us":8.16,"base_us":28.5,"floor_pct":46.3},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.009,"ours_us":22.1,"base_us":22.3,"floor_pct":47.2},{"load":"verify 16","shape":"m=128 o","x":0.962,"ours_us":10.15,"base_us":9.77,"floor_pct":59.0},{"load":"verify 16","shape":"m=128 pooler","x":1.027,"ours_us":28.65,"base_us":29.42,"floor_pct":58.1},{"load":"verify 16","shape":"m=128 unpooler","x":0.99,"ours_us":26.76,"base_us":26.5,"floor_pct":62.2},{"load":"verify 16","shape":"m=128 router","x":1.75,"ours_us":17.84,"base_us":31.22,"floor_pct":23.6},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.016,"ours_us":26.26,"base_us":26.67,"floor_pct":48.5},{"load":"verify 32","shape":"m=256 o","x":0.993,"ours_us":11.34,"base_us":11.26,"floor_pct":59.7},{"load":"verify 32","shape":"m=256 pooler","x":0.993,"ours_us":35.58,"base_us":35.34,"floor_pct":59.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.989,"ours_us":30.18,"base_us":29.86,"floor_pct":69.9},{"load":"verify 32","shape":"m=256 router","x":2.668,"ours_us":18.25,"base_us":48.7,"floor_pct":25.6}]},{"step":4,"x":1.0607173969012014,"lo":1.0574001903515557,"hi":1.0645411422858684,"pct":49.134311414936235,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.694,"ours_us":26.75,"base_us":18.58,"floor_pct":36.9},{"load":"decode 1","shape":"m=1 o","x":0.972,"ours_us":9.69,"base_us":9.42,"floor_pct":56.7},{"load":"decode 1","shape":"m=1 pooler","x":1.015,"ours_us":27.17,"base_us":27.58,"floor_pct":58.8},{"load":"decode 1","shape":"m=1 unpooler","x":1.045,"ours_us":8.25,"base_us":8.62,"floor_pct":45.5},{"load":"decode 1","shape":"m=1 router","x":0.888,"ours_us":13.79,"base_us":12.24,"floor_pct":27.2},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.979,"ours_us":18.78,"base_us":18.38,"floor_pct":52.9},{"load":"decode 16","shape":"m=16 o","x":0.992,"ours_us":9.42,"base_us":9.34,"floor_pct":59.0},{"load":"decode 16","shape":"m=16 pooler","x":1.033,"ours_us":27.13,"base_us":28.02,"floor_pct":59.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.008,"ours_us":25.02,"base_us":25.23,"floor_pct":64.2},{"load":"decode 16","shape":"m=16 router","x":3.021,"ours_us":9.42,"base_us":28.46,"floor_pct":40.4},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.891,"ours_us":21.06,"base_us":18.76,"floor_pct":47.5},{"load":"decode 32","shape":"m=32 o","x":0.942,"ours_us":10.04,"base_us":9.46,"floor_pct":55.9},{"load":"decode 32","shape":"m=32 pooler","x":0.999,"ours_us":28.31,"base_us":28.27,"floor_pct":57.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.973,"ours_us":26.5,"base_us":25.78,"floor_pct":60.9},{"load":"decode 32","shape":"m=32 router","x":1.338,"ours_us":16.09,"base_us":21.53,"floor_pct":24.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.995,"ours_us":18.58,"base_us":18.49,"floor_pct":53.3},{"load":"verify 1","shape":"m=8 o","x":0.986,"ours_us":9.51,"base_us":9.38,"floor_pct":58.1},{"load":"verify 1","shape":"m=8 pooler","x":1.036,"ours_us":27.02,"base_us":27.99,"floor_pct":59.3},{"load":"verify 1","shape":"m=8 unpooler","x":1.039,"ours_us":8.45,"base_us":8.78,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 router","x":3.391,"ours_us":8.47,"base_us":28.72,"floor_pct":44.6},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.0,"ours_us":22.3,"base_us":22.3,"floor_pct":46.8},{"load":"verify 16","shape":"m=128 o","x":0.986,"ours_us":10.11,"base_us":9.98,"floor_pct":59.2},{"load":"verify 16","shape":"m=128 pooler","x":0.955,"ours_us":29.67,"base_us":28.34,"floor_pct":56.1},{"load":"verify 16","shape":"m=128 unpooler","x":0.995,"ours_us":26.47,"base_us":26.34,"floor_pct":62.8},{"load":"verify 16","shape":"m=128 router","x":2.47,"ours_us":12.79,"base_us":31.59,"floor_pct":32.9},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.015,"ours_us":26.15,"base_us":26.54,"floor_pct":48.7},{"load":"verify 32","shape":"m=256 o","x":1.01,"ours_us":11.05,"base_us":11.16,"floor_pct":61.3},{"load":"verify 32","shape":"m=256 pooler","x":0.979,"ours_us":35.96,"base_us":35.21,"floor_pct":58.6},{"load":"verify 32","shape":"m=256 unpooler","x":0.997,"ours_us":30.42,"base_us":30.32,"floor_pct":69.3},{"load":"verify 32","shape":"m=256 router","x":0.646,"ours_us":75.5,"base_us":48.76,"floor_pct":6.2}]},{"step":5,"x":1.2052085469648746,"lo":1.2007971961070867,"hi":1.2089999867474306,"pct":55.47049897364123,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.001,"ours_us":18.7,"base_us":18.72,"floor_pct":52.8},{"load":"decode 1","shape":"m=1 o","x":1.014,"ours_us":9.35,"base_us":9.48,"floor_pct":58.8},{"load":"decode 1","shape":"m=1 pooler","x":1.018,"ours_us":27.61,"base_us":28.11,"floor_pct":57.9},{"load":"decode 1","shape":"m=1 unpooler","x":1.055,"ours_us":8.35,"base_us":8.81,"floor_pct":44.9},{"load":"decode 1","shape":"m=1 router","x":1.511,"ours_us":8.15,"base_us":12.32,"floor_pct":46.0},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.009,"ours_us":18.4,"base_us":18.57,"floor_pct":54.0},{"load":"decode 16","shape":"m=16 o","x":0.977,"ours_us":9.68,"base_us":9.46,"floor_pct":57.4},{"load":"decode 16","shape":"m=16 pooler","x":1.017,"ours_us":27.48,"base_us":27.94,"floor_pct":58.5},{"load":"decode 16","shape":"m=16 unpooler","x":1.011,"ours_us":25.04,"base_us":25.31,"floor_pct":64.2},{"load":"decode 16","shape":"m=16 router","x":3.347,"ours_us":8.55,"base_us":28.61,"floor_pct":44.5},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.998,"ours_us":18.9,"base_us":18.86,"floor_pct":52.9},{"load":"decode 32","shape":"m=32 o","x":1.014,"ours_us":9.44,"base_us":9.58,"floor_pct":59.5},{"load":"decode 32","shape":"m=32 pooler","x":0.987,"ours_us":28.6,"base_us":28.23,"floor_pct":56.5},{"load":"decode 32","shape":"m=32 unpooler","x":1.012,"ours_us":25.4,"base_us":25.7,"floor_pct":63.6},{"load":"decode 32","shape":"m=32 router","x":2.301,"ours_us":9.43,"base_us":21.69,"floor_pct":41.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.987,"ours_us":19.01,"base_us":18.75,"floor_pct":52.1},{"load":"verify 1","shape":"m=8 o","x":1.014,"ours_us":9.2,"base_us":9.33,"floor_pct":60.1},{"load":"verify 1","shape":"m=8 pooler","x":1.019,"ours_us":27.27,"base_us":27.78,"floor_pct":58.7},{"load":"verify 1","shape":"m=8 unpooler","x":1.038,"ours_us":8.59,"base_us":8.92,"floor_pct":43.8},{"load":"verify 1","shape":"m=8 router","x":3.398,"ours_us":8.45,"base_us":28.71,"floor_pct":44.7},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.013,"ours_us":22.19,"base_us":22.49,"floor_pct":47.0},{"load":"verify 16","shape":"m=128 o","x":1.0,"ours_us":10.03,"base_us":10.03,"floor_pct":59.7},{"load":"verify 16","shape":"m=128 pooler","x":1.041,"ours_us":28.38,"base_us":29.54,"floor_pct":58.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.971,"ours_us":27.3,"base_us":26.5,"floor_pct":60.9},{"load":"verify 16","shape":"m=128 router","x":2.672,"ours_us":11.76,"base_us":31.41,"floor_pct":35.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.01,"ours_us":26.35,"base_us":26.62,"floor_pct":48.3},{"load":"verify 32","shape":"m=256 o","x":0.996,"ours_us":11.38,"base_us":11.34,"floor_pct":59.5},{"load":"verify 32","shape":"m=256 pooler","x":1.009,"ours_us":35.39,"base_us":35.72,"floor_pct":59.6},{"load":"verify 32","shape":"m=256 unpooler","x":1.006,"ours_us":29.96,"base_us":30.14,"floor_pct":70.4},{"load":"verify 32","shape":"m=256 router","x":3.777,"ours_us":12.97,"base_us":48.98,"floor_pct":36.1}]},{"step":6,"x":1.2029951274902166,"lo":1.1989752184798623,"hi":1.20686099458205,"pct":55.82624488702238,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.996,"ours_us":18.45,"base_us":18.38,"floor_pct":53.5},{"load":"decode 1","shape":"m=1 o","x":1.003,"ours_us":9.27,"base_us":9.3,"floor_pct":59.3},{"load":"decode 1","shape":"m=1 pooler","x":0.993,"ours_us":27.76,"base_us":27.58,"floor_pct":57.6},{"load":"decode 1","shape":"m=1 unpooler","x":1.03,"ours_us":8.41,"base_us":8.66,"floor_pct":44.6},{"load":"decode 1","shape":"m=1 router","x":1.526,"ours_us":7.99,"base_us":12.2,"floor_pct":46.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.999,"ours_us":18.46,"base_us":18.45,"floor_pct":53.8},{"load":"decode 16","shape":"m=16 o","x":0.987,"ours_us":9.4,"base_us":9.27,"floor_pct":59.1},{"load":"decode 16","shape":"m=16 pooler","x":1.011,"ours_us":27.61,"base_us":27.91,"floor_pct":58.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.0,"ours_us":25.04,"base_us":25.04,"floor_pct":64.1},{"load":"decode 16","shape":"m=16 router","x":3.441,"ours_us":8.28,"base_us":28.47,"floor_pct":46.0},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.992,"ours_us":18.66,"base_us":18.51,"floor_pct":53.6},{"load":"decode 32","shape":"m=32 o","x":0.991,"ours_us":9.47,"base_us":9.38,"floor_pct":59.4},{"load":"decode 32","shape":"m=32 pooler","x":0.998,"ours_us":28.46,"base_us":28.39,"floor_pct":56.7},{"load":"decode 32","shape":"m=32 unpooler","x":1.005,"ours_us":25.03,"base_us":25.15,"floor_pct":64.5},{"load":"decode 32","shape":"m=32 router","x":2.272,"ours_us":9.46,"base_us":21.5,"floor_pct":40.8},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.992,"ours_us":18.76,"base_us":18.6,"floor_pct":52.8},{"load":"verify 1","shape":"m=8 o","x":0.998,"ours_us":9.27,"base_us":9.26,"floor_pct":59.6},{"load":"verify 1","shape":"m=8 pooler","x":1.008,"ours_us":27.14,"base_us":27.37,"floor_pct":59.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.041,"ours_us":8.42,"base_us":8.76,"floor_pct":44.7},{"load":"verify 1","shape":"m=8 router","x":3.541,"ours_us":8.08,"base_us":28.62,"floor_pct":46.7},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.033,"ours_us":21.84,"base_us":22.58,"floor_pct":47.7},{"load":"verify 16","shape":"m=128 o","x":0.989,"ours_us":9.91,"base_us":9.81,"floor_pct":60.4},{"load":"verify 16","shape":"m=128 pooler","x":1.042,"ours_us":28.21,"base_us":29.41,"floor_pct":59.0},{"load":"verify 16","shape":"m=128 unpooler","x":1.001,"ours_us":26.93,"base_us":26.97,"floor_pct":61.8},{"load":"verify 16","shape":"m=128 router","x":2.695,"ours_us":11.58,"base_us":31.22,"floor_pct":36.4},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.024,"ours_us":26.07,"base_us":26.7,"floor_pct":48.9},{"load":"verify 32","shape":"m=256 o","x":1.012,"ours_us":11.02,"base_us":11.16,"floor_pct":61.4},{"load":"verify 32","shape":"m=256 pooler","x":0.986,"ours_us":35.5,"base_us":35.01,"floor_pct":59.4},{"load":"verify 32","shape":"m=256 unpooler","x":0.999,"ours_us":30.49,"base_us":30.47,"floor_pct":69.2},{"load":"verify 32","shape":"m=256 router","x":3.726,"ours_us":13.08,"base_us":48.74,"floor_pct":35.7}]}],"confirm":{"step":null,"x":1.2071974363501476,"lo":1.2026095788030222,"hi":1.2117470589752926,"pct":55.69088968356908,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.0,"ours_us":18.61,"base_us":18.61,"floor_pct":53.0},{"load":"decode 1","shape":"m=1 o","x":1.012,"ours_us":9.33,"base_us":9.44,"floor_pct":59.0},{"load":"decode 1","shape":"m=1 pooler","x":1.013,"ours_us":27.52,"base_us":27.87,"floor_pct":58.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.024,"ours_us":8.36,"base_us":8.56,"floor_pct":44.9},{"load":"decode 1","shape":"m=1 router","x":1.521,"ours_us":8.01,"base_us":12.18,"floor_pct":46.8},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.006,"ours_us":18.51,"base_us":18.62,"floor_pct":53.7},{"load":"decode 16","shape":"m=16 o","x":0.989,"ours_us":9.47,"base_us":9.36,"floor_pct":58.7},{"load":"decode 16","shape":"m=16 pooler","x":1.023,"ours_us":27.33,"base_us":27.97,"floor_pct":58.8},{"load":"decode 16","shape":"m=16 unpooler","x":1.015,"ours_us":24.99,"base_us":25.35,"floor_pct":64.3},{"load":"decode 16","shape":"m=16 router","x":3.451,"ours_us":8.28,"base_us":28.58,"floor_pct":46.0},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.011,"ours_us":18.71,"base_us":18.92,"floor_pct":53.5},{"load":"decode 32","shape":"m=32 o","x":0.999,"ours_us":9.55,"base_us":9.54,"floor_pct":58.8},{"load":"decode 32","shape":"m=32 pooler","x":0.998,"ours_us":28.52,"base_us":28.46,"floor_pct":56.6},{"load":"decode 32","shape":"m=32 unpooler","x":0.998,"ours_us":25.37,"base_us":25.31,"floor_pct":63.6},{"load":"decode 32","shape":"m=32 router","x":2.286,"ours_us":9.45,"base_us":21.62,"floor_pct":40.9},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.994,"ours_us":18.77,"base_us":18.66,"floor_pct":52.7},{"load":"verify 1","shape":"m=8 o","x":0.991,"ours_us":9.51,"base_us":9.42,"floor_pct":58.1},{"load":"verify 1","shape":"m=8 pooler","x":1.017,"ours_us":27.37,"base_us":27.82,"floor_pct":58.5},{"load":"verify 1","shape":"m=8 unpooler","x":1.023,"ours_us":8.52,"base_us":8.71,"floor_pct":44.2},{"load":"verify 1","shape":"m=8 router","x":3.541,"ours_us":8.06,"base_us":28.53,"floor_pct":46.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.012,"ours_us":22.37,"base_us":22.64,"floor_pct":46.6},{"load":"verify 16","shape":"m=128 o","x":1.001,"ours_us":9.99,"base_us":10.0,"floor_pct":59.9},{"load":"verify 16","shape":"m=128 pooler","x":1.037,"ours_us":28.42,"base_us":29.46,"floor_pct":58.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.996,"ours_us":27.03,"base_us":26.94,"floor_pct":61.5},{"load":"verify 16","shape":"m=128 router","x":2.662,"ours_us":11.76,"base_us":31.31,"floor_pct":35.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.023,"ours_us":25.96,"base_us":26.55,"floor_pct":49.1},{"load":"verify 32","shape":"m=256 o","x":0.996,"ours_us":11.2,"base_us":11.16,"floor_pct":60.5},{"load":"verify 32","shape":"m=256 pooler","x":1.025,"ours_us":34.94,"base_us":35.81,"floor_pct":60.4},{"load":"verify 32","shape":"m=256 unpooler","x":1.002,"ours_us":30.38,"base_us":30.45,"floor_pct":69.4},{"load":"verify 32","shape":"m=256 router","x":3.769,"ours_us":12.91,"base_us":48.66,"floor_pct":36.2}]},"src":"bench/product/runs/learnability/bf16-B.jsonl"},{"id":"moe-A","model":"Claude Opus 5.5","harness":"Claude Code","family":"moe_nvfp4_ep8","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":0.29684302004018265},"T2":{"x":137.60950654824418,"pct":37.13492864547281},"T3":{"x":274.21901309648837,"pct":74.0},"T4":{"x":333.50961052275613,"pct":90.0}},"steps":[{"step":1,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"GRADER_BUG_REFUNDED","note":null,"cases":[]},{"step":2,"x":67.20141620713015,"lo":67.06128338942725,"hi":67.48985670257125,"pct":18.232731168108717,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":124.403,"ours_us":461.34,"base_us":57392.36,"floor_pct":10.7},{"load":"decode 16","shape":"bs=16 dims=full","x":74.609,"ours_us":774.12,"base_us":57756.93,"floor_pct":24.6},{"load":"decode 32","shape":"bs=32 dims=full","x":107.934,"ours_us":535.36,"base_us":57783.98,"floor_pct":35.6},{"load":"verify 1","shape":"bs=8 dims=full","x":116.33,"ours_us":495.17,"base_us":57602.92,"floor_pct":38.5},{"load":"verify 16","shape":"bs=128 dims=full","x":33.584,"ours_us":1733.92,"base_us":58232.01,"floor_pct":11.0},{"load":"verify 32","shape":"bs=256 dims=full","x":23.532,"ours_us":2503.0,"base_us":58901.0,"floor_pct":9.2}]},{"step":2,"x":161.3485860462596,"lo":160.99558063679635,"hi":161.8743353641995,"pct":43.6128398435558,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":372.101,"ours_us":154.94,"base_us":57651.95,"floor_pct":31.7},{"load":"decode 16","shape":"bs=16 dims=full","x":229.732,"ours_us":252.26,"base_us":57952.75,"floor_pct":75.6},{"load":"decode 32","shape":"bs=32 dims=full","x":144.704,"ours_us":400.92,"base_us":58014.67,"floor_pct":47.6},{"load":"verify 1","shape":"bs=8 dims=full","x":305.034,"ours_us":189.45,"base_us":57787.7,"floor_pct":100.7},{"load":"verify 16","shape":"bs=128 dims=full","x":84.191,"ours_us":693.98,"base_us":58427.19,"floor_pct":27.5},{"load":"verify 32","shape":"bs=256 dims=full","x":55.541,"ours_us":1064.76,"base_us":59137.48,"floor_pct":21.7}]},{"step":3,"x":214.51761831642062,"lo":213.7071907877432,"hi":215.23277653404043,"pct":58.086168285899596,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":505.878,"ours_us":113.69,"base_us":57513.71,"floor_pct":43.3},{"load":"decode 16","shape":"bs=16 dims=full","x":258.589,"ours_us":223.81,"base_us":57874.9,"floor_pct":85.2},{"load":"decode 32","shape":"bs=32 dims=full","x":229.013,"ours_us":252.69,"base_us":57869.07,"floor_pct":75.5},{"load":"verify 1","shape":"bs=8 dims=full","x":308.305,"ours_us":187.31,"base_us":57748.39,"floor_pct":101.8},{"load":"verify 16","shape":"bs=128 dims=full","x":135.203,"ours_us":431.37,"base_us":58321.9,"floor_pct":44.3},{"load":"verify 32","shape":"bs=256 dims=full","x":78.036,"ours_us":756.48,"base_us":59033.16,"floor_pct":30.5}]},{"step":4,"x":136.15686017688427,"lo":135.7559798764701,"hi":136.59773580956085,"pct":36.84080841188893,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":245.713,"ours_us":234.29,"base_us":57568.64,"floor_pct":21.0},{"load":"decode 16","shape":"bs=16 dims=full","x":344.766,"ours_us":168.1,"base_us":57955.54,"floor_pct":113.5},{"load":"decode 32","shape":"bs=32 dims=full","x":314.529,"ours_us":184.38,"base_us":57992.71,"floor_pct":103.5},{"load":"verify 1","shape":"bs=8 dims=full","x":96.098,"ours_us":601.01,"base_us":57755.79,"floor_pct":31.7},{"load":"verify 16","shape":"bs=128 dims=full","x":61.24,"ours_us":952.06,"base_us":58303.67,"floor_pct":20.1},{"load":"verify 32","shape":"bs=256 dims=full","x":40.633,"ours_us":1453.03,"base_us":59040.61,"floor_pct":15.9}]},{"step":5,"x":241.71260644977312,"lo":241.1160306541624,"hi":242.02965497100246,"pct":65.38104479693294,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":610.88,"ours_us":94.24,"base_us":57570.01,"floor_pct":52.2},{"load":"decode 16","shape":"bs=16 dims=full","x":328.124,"ours_us":176.57,"base_us":57937.7,"floor_pct":108.1},{"load":"decode 32","shape":"bs=32 dims=full","x":248.595,"ours_us":233.38,"base_us":58017.32,"floor_pct":81.8},{"load":"verify 1","shape":"bs=8 dims=full","x":287.544,"ours_us":200.9,"base_us":57768.96,"floor_pct":95.0},{"load":"verify 16","shape":"bs=128 dims=full","x":124.986,"ours_us":466.99,"base_us":58367.55,"floor_pct":40.9},{"load":"verify 32","shape":"bs=256 dims=full","x":111.364,"ours_us":530.39,"base_us":59066.03,"floor_pct":43.6}]},{"step":6,"x":219.4748725792469,"lo":219.1459090654345,"hi":219.9396601234624,"pct":59.547828890403764,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":608.37,"ours_us":94.39,"base_us":57422.23,"floor_pct":52.1},{"load":"decode 16","shape":"bs=16 dims=full","x":327.213,"ours_us":176.49,"base_us":57748.7,"floor_pct":108.1},{"load":"decode 32","shape":"bs=32 dims=full","x":204.84,"ours_us":282.26,"base_us":57817.22,"floor_pct":67.6},{"load":"verify 1","shape":"bs=8 dims=full","x":305.151,"ours_us":188.78,"base_us":57605.1,"floor_pct":101.1},{"load":"verify 16","shape":"bs=128 dims=full","x":107.854,"ours_us":539.54,"base_us":58191.3,"floor_pct":35.4},{"load":"verify 32","shape":"bs=256 dims=full","x":83.281,"ours_us":706.98,"base_us":58877.46,"floor_pct":32.7}]}],"confirm":{"step":null,"x":218.54292306429204,"lo":217.63000920020775,"hi":219.03909564984795,"pct":58.97540117346703,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":584.723,"ours_us":98.67,"base_us":57692.47,"floor_pct":49.9},{"load":"decode 16","shape":"bs=16 dims=full","x":321.644,"ours_us":180.66,"base_us":58109.44,"floor_pct":105.6},{"load":"decode 32","shape":"bs=32 dims=full","x":253.523,"ours_us":229.32,"base_us":58137.83,"floor_pct":83.2},{"load":"verify 1","shape":"bs=8 dims=full","x":287.833,"ours_us":201.35,"base_us":57956.6,"floor_pct":94.7},{"load":"verify 16","shape":"bs=128 dims=full","x":103.807,"ours_us":563.41,"base_us":58485.58,"floor_pct":33.9},{"load":"verify 32","shape":"bs=256 dims=full","x":76.473,"ours_us":773.65,"base_us":59163.38,"floor_pct":29.9}]},"src":"bench/product/runs/learnability/moe-A.jsonl"}],"layers":[{"id":"generate","name":"Generate","desc":"Where ideas and code come from."},{"id":"measure","name":"Measure","desc":"Turn a file into a trusted number."},{"id":"decide","name":"Decide","desc":"Keep, kill, confirm, combine, send."},{"id":"remember","name":"Remember","desc":"Ledgers that outlive any session."},{"id":"coordinate","name":"Coordinate","desc":"Agents and the owner talk."},{"id":"guard","name":"Guard","desc":"Stop the known ways to fail."},{"id":"infra","name":"Infra","desc":"GPUs, hosts, sync."}],"headline":{"ideas":14936,"full_hits":340,"proven_hits":236,"trials":13319,"teams":2014,"days":14,"problems":5,"gpu_runs":58618,"gpu_runs_passed":57235,"judge_runs":533,"judge_scored":181,"unique_files":7610,"run_hours":2710.4,"box_full":55401,"peak_rank":1,"peak_hours":17.6,"peak_from":"2026-10-24T00:20:23Z","final_overall":25,"usd_total":41864.77,"usd_gpu":25776.02,"usd_llm":16088.75,"usd_sub_equiv":29204.06,"usd_work":27891.67,"usd_gpu_work":18615.92,"usd_llm_work":9275.75,"commits":8021,"commits_raw":36127,"peak_lanes":62,"rented_gpu_hours":3329.2,"records":209,"lane_runs":73533,"models":15,"mistakes":322,"landings":135},"components":[{"name":"Modal judge clone","layer":"measure","file":"bench/modal/judge.py","introduced":"2026-10-22T16:09:10-07:00","introduced_src":"git first-add 68e45985b25","date_note":null,"what":"Re-runs the organizers' scoring method (cold L2, CUDA graphs, paired SGLang ratios) on a Modal H200 for any file.","measured_effect":{"value":"Judge / clone ratio for the same bytes: decode x0.969, MoE x0.967, attention x0.959, bf16 x1.00, prefill x1.03; so a clone gain is not a judge gain","kind":"measured","src":"AGENTS.md:74-76"},"file_exists":true,"git_first_add":"2026-10-22T16:09:10-07:00 68e45985b25 bf16: E6 (L2 prefetch)"},{"name":"Clone and random-draw inputs","layer":"measure","file":"bench/tools/clone_inputs.py","introduced":"2026-10-23T00:44:52-07:00","introduced_src":"git first-add 8878520c43e","date_note":null,"what":"Gives labs the clone's exact per-case inputs plus >= 20 random routing draws, so a lab predicts the judge.","measured_effect":{"value":"Clone's 4 fixed draws read w2_wg48sl +0.83%; the random-draw predictor read +0.45%. Random-routing labs had misjudged wide MoE kernels by ~4%","kind":"measured","src":"bench/tools/clone_inputs.py:3-8; AGENTS.md:141"},"file_exists":true,"git_first_add":"2026-10-23T00:44:52-07:00 8878520c43e bench/tools/clone_inpu"},{"name":"Paired ABBA confirmation","layer":"measure","file":"bench/tools/paired_confirm.py","introduced":"2026-10-27T02:55:44-07:00","introduced_src":"git first-add 2e91ee068b1","date_note":"A one-off confirm script (confirm-decode.md, first-add 2026-10-26T16:18 2aad6607c4f) preceded the tool.","what":"Predeclared same-box A-B-B-A blocks with a 95% CI decide whether a candidate beats the current base.","measured_effect":{"value":"72 h to 10-29: 1,011 paired verdicts, 81 CONFIRMED, 544 (54%) ABORTED for infrastructure. Prototype run 10-26: decode composite C screened +0.289%, confirmed +0.222% (95% CI 0.180..0.264)","kind":"measured","src":"bench/books/POSTMORTEM_1029.md:14-19; bench/reports/evolution/2026-10-26-01a17a50/confirm-decode.md:3-5"},"file_exists":true,"git_first_add":"2026-10-27T02:55:44-07:00 2e91ee068b1 paired_confirm: predec"},{"name":"Scoped judge-exact A/B (jexact_scoped)","layer":"measure","file":"bench/tools/jexact_scoped.py","introduced":"2026-10-31T21:54:58-07:00","introduced_src":"git first-add 70b7db71f17","date_note":"First-add is a bulk 'keep: this disk's work, 5819 files' commit; the file may have existed on disk earlier. SPEC_V3 rule 9 made it the verdict tool on 11-01.","what":"Times only the judge cases a candidate touches; route_identity proves the other cases identical, so they count zero.","measured_effect":{"value":"Resolution ~0.02% vs whole-x noise 0.3-0.6%. Whole-x verdicts only saw wins >= ~1%: hit rate 31.8% below row 2.0, 0.5% at 2.2-2.3, 0% at >= 2.3","kind":"measured","src":"bench/agents/SPEC_V3.md:73-85"},"file_exists":true,"git_first_add":"2026-10-31T22:29:13-07:00 edb7c27b03a sync: main's content a"},{"name":"Internal judge","layer":"measure","file":"bench/tools/internal_judge.py","introduced":"2026-11-01T15:53:28-08:00","introduced_src":"git first-add ad3b38bf87d","date_note":"First-add is a push_sync squash commit ('main's content as one commit'). Docstring dates it 11-01.","what":"Scores decode chain candidates with the judge's own timing sequence; real judge reads become an external check only.","measured_effect":{"value":"Timing protocol vs 37 judge runs: bias -0.34%, MAE 0.66%, 86% of judge-ordered pairs ordered right","kind":"measured","src":"bench/tools/internal_judge.py:9-10 (cites judge_gap/LEARNINGS.md)"},"owner_quote":{"text":"we must have an internal source of truth and an external one, which we trust a bit less","src":"bench/agents/SPEC_V3.md:102"},"file_exists":true,"git_first_add":"2026-11-01T15:53:28-08:00 ad3b38bf87d sync: main's content a"},{"name":"Judge-calibrated screen","layer":"measure","file":"bench/tools/judge_cal.py","introduced":"2026-10-30T17:58:32-07:00","introduced_src":"git first-add 26ce4af52eb","date_note":null,"what":"Fits per-case box-to-judge factors and code-feature offsets so box screens rank files the way the judge does.","measured_effect":{"value":"Box vs judge rank agreement Spearman 0.28 -> 0.63 (leave-one-out); bias +3.6% -> -0.2%","kind":"measured","src":"git commit 26ce4af52eb message; bench/tools/judge_cal.py:4-5"},"owner_quote":{"text":"why is our decode layer so low? ... we are #36","src":"bench/tools/judge_cal.py:2"},"file_exists":true,"git_first_add":"2026-10-30T17:59:12-07:00 cd1141c0df1 sync: main's content a"},{"name":"Instruction budget (ISA)","layer":"measure","file":"bench/tools/isa_budget.py","introduced":"2026-10-23T13:47:50-07:00","introduced_src":"git first-add 4639753c54c","date_note":"First-add is a batch commit ('main 20:55: sub-goal workflow done').","what":"Derives the minimal instructions per pipe for a kernel and compares them with SASS counts to name the waste.","measured_effect":{"value":"First use: attention verify-full ran INT at 9x the needed count, latency-bound at 69% of the HBM floor","kind":"measured","src":"AGENTS.md:145"},"owner_quote":{"text":"decompose into the minimal set of operations and derive what instructions to run","src":"AGENTS.md:145 (owner's advisor, 10-23)"},"file_exists":true,"git_first_add":"2026-10-23T13:47:50-07:00 4639753c54c main 20:55: sub-goal w"},{"name":"Scientist ledger and kill rule","layer":"decide","file":"bench/tools/sci.py","introduced":"2026-10-22T17:35:59-07:00","introduced_src":"git first-add ff8c7643af5","date_note":null,"what":"Append-only ledger of every idea (predicted gain, p) and trial; replays state and retires ideas after 3 misses or 2 BAD.","measured_effect":{"value":"To 11-05: 14,667 ideas, 13,638 trials: 1,201 hit (8.8%), 6,179 miss, 2,936 flat, 2,316 error, 1,006 bad; 1,018 kill events, 9 revives","kind":"measured","src":"bench/sci/events.jsonl (read-only count of kind=add/trial/kill/revive, 2026-11-05)"},"owner_quote":{"text":"after misses kill ideas, move on; keep pushing the ones that start to work","src":"bench/tools/sci.py:7"},"file_exists":true,"git_first_add":"2026-10-22T17:35:59-07:00 ff8c7643af5 bench: auto-scientist "},{"name":"Racing screen and gene bank (race + keeps)","layer":"decide","file":"bench/tools/race.py","introduced":"2026-10-26T16:10:26-07:00","introduced_src":"git first-add 271ceb360eb (keeps.py; race.py 7ea0a869726 at 16:23)","date_note":null,"what":"Ranks small unpaired candidates for paired confirmation and keeps near-best files as genes to stack; no % cutoff.","measured_effect":{"value":"Gene bank at work: PF0RH (+0.009%) and STG (+1.1..1.4 us) were stacked into decode cand63","kind":"measured","src":"bench/agents/SPEC_V3.md:40-42"},"owner_quote":{"text":"below 1% actually starts to matter and we win by sheer scale of experiments ... weighted by the impact","src":"AGENTS.md:266-267"},"file_exists":true,"git_first_add":"2026-10-26T16:23:48-07:00 7ea0a869726 Racing replaces % thre"},{"name":"Gap to physical floor","layer":"decide","file":"bench/tools/gap.py","introduced":"2026-10-22T17:26:21-07:00","introduced_src":"git first-add 654e0e351cf","date_note":null,"what":"Prices every judge case against its physical floor (bytes / 4.8 TB/s, FLOPs / peak) and ranks work by score gain to floor.","measured_effect":{"value":"10-24 read: attention 58%, prefill 61%, decode 68%, MoE 74%, bf16 88% of physical ceiling. Recipe 'floor-first' won 11 of 157 clean tests (7%) vs 'compose' 130 of 812 (16%)","kind":"measured","src":"AGENTS.md:62; bench/WORKED.md:7,13"},"file_exists":true,"git_first_add":"2026-10-22T17:26:21-07:00 654e0e351cf bench: GAP.md ledger ("},{"name":"Candidate chain (one merge queue)","layer":"decide","file":"bench/tools/chain.py","introduced":"2026-10-30T14:41:18-07:00","introduced_src":"git first-add 022c5af0fdf","date_note":"First-add is a push_sync squash commit; the ledger's first row is 2026-10-30T22:37:12Z, consistent.","what":"One linear chain per problem: each landing is a diff confirmed paired against the CURRENT head, written by one merge queue.","measured_effect":{"value":"Decode: 115 merged landings 10-30 22:37Z to 11-05 05:03Z (cand0 x 1.168 -> cand126). Gain per landing on 10-31 fell +1.08 -> +0.59 -> +0.72 -> +0.31 -> +0.18 -> +0.12%","kind":"measured","src":"bench/chain/muse_decode_layer.jsonl (read-only count); bench/SYSTEMS_THINKING.md:12-13"},"owner_quote":{"text":"one source of truth, one main branch ... every idea is a branch you can bisect","src":"bench/tools/chain.py:4"},"file_exists":true,"git_first_add":"2026-10-30T14:41:18-07:00 022c5af0fdf sync: main's content a"},{"name":"Win loss audit","layer":"decide","file":"bench/tools/win_audit.py","introduced":"2026-10-30T15:35:03-07:00","introduced_src":"git first-add 70577bb5a29","date_note":"First-add is a cloud-box bulk commit ('swarm box: agents' work 22:35 UTC').","what":"Follows every CONFIRMED paired win and names the stage where it stopped before the scored row.","measured_effect":null,"why_null":"It prints its audit on demand; no stored audit output was found to count losses before and after it shipped.","owner_quote":{"text":"12 of 13 changes to the best file had no confirmed test; a confirmed +1.34% win was dropped","src":"bench/tools/win_audit.py:4-5"},"file_exists":true,"git_first_add":"2026-10-30T15:36:40-07:00 18acbeadf18 sync: main's content a"},{"name":"Hunch loop (learning loop)","layer":"decide","file":"bench/tools/hunch_loop.py","introduced":"2026-10-23T20:52:57-07:00","introduced_src":"git first-add ac3cd9e74b9","date_note":null,"what":"Every 30 min rebuilds FLOOR, WORKED and HUNCHES: hit rate per recipe, calibrated hunch odds, routes the top unowned hunch.","measured_effect":{"value":"Its calibration table: ideas registered at p 0.8-1.0 won 35%, p 0.6-0.8 won 29%, p 0.4-0.6 won 7%, p < 0.4 won 2%","kind":"measured","src":"bench/WORKED.md:18-24"},"owner_quote":{"text":"create a HUNCH list ... and then a list with what actually worked ... INVENT learning loops like this","src":"bench/tools/hunch_loop.py:5-6"},"file_exists":true,"git_first_add":"2026-10-23T20:52:57-07:00 ac3cd9e74b9 learning loop: hunch_l"},{"name":"Idea-basin detector","layer":"decide","file":"bench/tools/basin.py","introduced":"2026-11-01T11:21:06-08:00","introduced_src":"git first-add 4ac3a98e8d1","date_note":"First-add is a bulk 'keep: this disk's work' commit.","what":"CPU-only check that flags strike lanes re-touching the same knobs with no gain (concentration, novelty, re-touch).","measured_effect":{"value":"11-05 scan: 8 of 31 decode strike lanes in ALARM (e.g. d32s_norms novelty 0.00, re-touch 0.92). Basis: 72% of 444 attempts were 1-3 knob edits, 47% re-touched a measured knob","kind":"measured","src":"okc/muse_decode_layer/strike/lanes/BASINS.md:1-14; bench/research/fork_flush_1101.md:17-18"},"file_exists":true,"git_first_add":"2026-11-01T19:49:32Z 0c686b8cb02 agent box pi:b26:pi-b26-1=r"},{"name":"Codex islands","layer":"generate","file":"bench/tools/codex_island.py","introduced":"2026-10-22T17:56:14-07:00","introduced_src":"git first-add 05e714d3a92","date_note":"islands.py supervisor first-add 2026-10-22T18:53:34 (86a07516d36, an auto 'growth:' commit).","what":"GPT-6 Codex agents as rotating islands per problem (explore, alpha, combination lanes), context flushed each round.","measured_effect":{"value":"At decode row >= 2.2, 22% of whole-x trials were errors, mostly old Codex lanes. Off from 10-23 07:25 UTC","kind":"measured","src":"bench/agents/SPEC_V3.md:82; AGENTS.md:110"},"owner_quote":{"text":"i dont have anymore codex compute","src":"AGENTS.md:110"},"file_exists":true,"git_first_add":"2026-10-22T17:56:14-07:00 05e714d3a92 bench: Codex island la"},{"name":"pi agent launcher and brain switch","layer":"generate","file":"bench/tools/pi_agent.py","introduced":"2026-10-23T02:29:05-07:00","introduced_src":"git first-add e081eeb0cdb","date_note":"brain.py first-add 2026-10-23T15:48:05-07:00 (f760b4f65d2).","what":"Launches pi-harness agents with hard $ and time caps; brain.py and llm.env pick the model (Argon, Sol, DeepSeek).","measured_effect":{"value":"Same 12 case lanes: Argon GPU-off wave ~$96 -> 1 possible candidate; Sol round $2.83 -> 14 candidate files. pi had shown $506 while the real Argon bill was ~$2k","kind":"measured","src":"bench/SYSTEMS.md:129,132; bench/agents/FAMILY_LEARNINGS.md:115"},"owner_quote":{"text":"use gpt 6.1 sol for many cheap workflows ... no argon, we are spending way too much","src":"AGENTS.md:96"},"file_exists":true,"git_first_add":"2026-10-23T02:29:05-07:00 e081eeb0cdb main: pi_agent.py (pi "},{"name":"pi + Argon islands","layer":"generate","file":"bench/tools/pi_island.py","introduced":"2026-10-23T14:31:46-07:00","introduced_src":"git first-add 0f199e4f263","date_note":"AGENTS.md:90 dates the owner order 10-23 21:55 UTC (14:55 PDT), 24 min after the git first-add.","what":"Codex-island briefs re-run on pi with Gemini 4 Argon; 15 islands live from 10-23 22:00 UTC, $25 cap each.","measured_effect":{"value":"Argon arm: 2 clean gains >= 1% in 136 shots. Argon bill $505 (10-23 UTC), $1,098 (10-24 UTC). All 15 islands hit their $25 totals by 00:00 UTC","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:124-125; bench/SYSTEMS.md:107,131"},"owner_quote":{"text":"replace all codex research with Pi and Argon harness","src":"AGENTS.md:90"},"file_exists":true,"git_first_add":"2026-10-23T14:31:46-07:00 0f199e4f263 owner: replace Codex w"},{"name":"Manager fleet","layer":"generate","file":"bench/tools/manager.py","introduced":"2026-10-24T10:49:51-07:00","introduced_src":"git first-add 1cc290b34e6","date_note":"bench/SYSTEMS.md:134 logs the MANAGER HARNESS decision at 10-24 08:55 UTC (01:55 PDT), ~9 h before the git first-add.","what":"Keeps case lanes, a hypothesis generator and an adversary live in rounds; later holds 30-40 pi agents at all times.","measured_effect":{"value":"11-01 00:01-16:01 UTC: 715 of 1,386 Sonnet runner runs ended in under 1 min; 380 ended BLOCKED. Earlier Sol lanes stopped after 1-9 min of their 75","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:76-77,121"},"owner_quote":{"text":"at least 6 threads always running ... in sync ... never stop","src":"bench/SYSTEMS.md:134"},"file_exists":true,"git_first_add":"2026-10-24T10:49:51-07:00 1cc290b34e6 manager.py: the manage"},{"name":"Deep research and Exa corpus","layer":"generate","file":"bench/tools/deep_research.py","introduced":"2026-10-23T16:09:26-07:00","introduced_src":"git first-add 6ad3a36bf1e","date_note":"exa.py first-add 2026-10-22T18:53:34-07:00 (86a07516d36, an auto 'growth:' commit).","what":"Cited answers to one specific question (Exa deep, Gemini Deep Research) plus a 220-report, ~3,800-page corpus.","measured_effect":{"value":"Gemini Deep Research: 47 reports, 0 of 81 specific claims survived premise checks. Exa: $4.44 for 477 calls, 10-24..10-29","kind":"measured","src":"bench/SYSTEMS.md:117; bench/agents/FAMILY_LEARNINGS.md:99-100"},"file_exists":true,"git_first_add":"2026-10-23T16:09:26-07:00 6ad3a36bf1e deep_research.py (Gemi"},{"name":"Paper library and premise check","layer":"generate","file":"bench/tools/papers.py","introduced":"2026-10-23T14:59:05-07:00","introduced_src":"git first-add 0874a0884ac","date_note":"The --binds gate in sci.py was added 2026-10-24T12:40:33-07:00 (602c4f64f18).","what":"Greppable paper library; from 10-24 a paper-class idea needs a --binds receipt that its target stage binds.","measured_effect":{"value":"20 paper ideas tested: 12 missed (another limit bound), 5 caught by the premise check before GPU, 1 shipped (attn-002: v32 full 481 -> 366 us)","kind":"measured","src":"bench/research/PAPER_AUDIT.md:11-18,52-56"},"file_exists":true,"git_first_add":"2026-10-23T14:59:05-07:00 0874a0884ac Agents set up end to e"},{"name":"Fork-and-Flush lineages","layer":"generate","file":"bench/research/fork_flush_1101.md","introduced":"2026-11-01T10:46:14-08:00","introduced_src":"git first-add d45640e10a9","date_note":"Runs inside bench/tools/strike_lanes.py ('ff'), first-add 2026-10-31. The paper's 0.71 vs 0.45 (arXiv 2610.07447) is the paper's number, not ours.","what":"Every 4 leader runs, fork a structural lane 4 ways from the leader's files; 3 forks start with no chat.","measured_effect":{"value":"ON for only 2 lanes on 11-01. First d32s round (stale base cand59): branches -0.08 / +0.08 / -0.25%, two voided; no confirmed FF win found","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:61; bench/manager/NEXT_ACTIONS.md:249"},"owner_quote":{"text":"our hit rate btw is pretty low now ... rebuild spec","src":"bench/agents/SPEC_V3.md:3-5"},"file_exists":true,"git_first_add":"2026-11-01T11:58:52-08:00 3ed1741de8b sync: main's content a"},{"name":"MAP-Elites map-reduce","layer":"generate","file":"bench/tools/mapreduce.py","introduced":"2026-10-29T13:02:02-07:00","introduced_src":"git first-add 2016cf4bde9","date_note":null,"what":"Finds the best file per judge case, merges case winners into one dispatch file, writes AlphaEvolve-style briefs.","measured_effect":{"value":"68 REDUCE files built; 55 were over the 1,000,000-byte file cap; 0 reached a paired verdict","kind":"measured","src":"bench/mapreduce/reduce.jsonl (read-only count); bench/manager/paired.jsonl (no /mapreduce/ rows)"},"owner_quote":{"text":"do map reduce and do alpha generation concepts in here asap","src":"bench/tools/mapreduce.py:4"},"file_exists":true,"git_first_add":"2026-10-29T13:05:35-07:00 8a020474434 sync: main's content a"},{"name":"Mailbox and ledger notes","layer":"coordinate","file":"bench/tools/mailbox.py","introduced":"2026-10-22T21:12:43-07:00","introduced_src":"git first-add 1b70661a631","date_note":"First-add is a batch commit (message is about statecopy.py).","what":"Per-agent read receipts ('blue dot') on shared files and notes; sci.py notes, todos and triggers carry the messages.","measured_effect":null,"why_null":"Activity only: 78,280 notes, 6,203 closed, 3,885 triggers fired (bench/sci/events.jsonl count). Counter-case: 5 todo notes to 'meta-0c' sat unseen 2.5 h (bench/agents/FAMILY_LEARNINGS.md:40). No outcome link measured.","owner_quote":{"text":"each agent has their own READ mailbox ... kinda like a blue dot","src":"bench/tools/mailbox.py:4"},"file_exists":true,"git_first_add":"2026-10-22T21:12:43-07:00 1b70661a631 tools: statecopy.py an"},{"name":"Agent groups and forum","layer":"coordinate","file":"bench/tools/groups.py","introduced":"2026-10-22T23:45:21-07:00","introduced_src":"git first-add c9385d6a36f","date_note":"forum.py first-add 2026-10-27T16:11:23-07:00 (1278663045d).","what":"Agent-run group chats (open or private) and, from 10-27, a deployed forum with channels, mentions and receipts.","measured_effect":null,"why_null":"Activity only: 23 group folders in bench/groups, 13,676 forum_sent events in the ledger; no link from a post to a gain was measured.","owner_quote":{"text":"let agents organise groups ... private chats for agents","src":"bench/tools/groups.py:4-5"},"file_exists":true,"git_first_add":"2026-10-22T23:45:21-07:00 c9385d6a36f bench/tools/groups.py:"},{"name":"Coordinator, work queue and unblock sweep","layer":"coordinate","file":"bench/manager/COORDINATOR.md","introduced":"2026-10-27T14:43:22-07:00","introduced_src":"git first-add f1366359246 (unblock.py)","date_note":"WORK_QUEUE.json first-add 14:52 (79883886d23) and COORDINATOR.md 14:57 (e3cb176b9d1), same afternoon.","what":"Start-of-session steps and tempo targets; idle lanes pull the next ranked task; unblock.py turns each stall into a flag.","measured_effect":{"value":"Built on: GPU busy 41% with 28 jobs queued; claude5 3/5 and claude6 3/4 lanes idle for hours. Later failure: 92 stale open tasks, one idea queued 25 times","kind":"measured","src":"bench/manager/COORDINATOR.md:11-12; bench/agents/FAMILY_LEARNINGS.md:73"},"owner_quote":{"text":"we are underutilised on resources and are not at our pace or tempo","src":"bench/manager/COORDINATOR.md:3-5"},"file_exists":true,"git_first_add":"2026-10-27T14:57:00-07:00 e3cb176b9d1 COORDINATOR.md: start "},{"name":"Mistake log","layer":"remember","file":"bench/MISTAKES.md","introduced":"2026-10-22T20:55:48-07:00","introduced_src":"git first-add 49e12a79980","date_note":"First-add is a commit about the SASS tool (batch).","what":"One row per judge FAIL, voided trial or retracted claim: the mistake class, its cost, and the gate that now catches it.","measured_effect":null,"why_null":"57,127 lines (38 MB) by 11-05, mostly appended on-policy rows; the repeat rate of mistake classes was never measured.","file_exists":true,"git_first_add":"2026-10-22T20:55:48-07:00 49e12a79980 sass tool: compile wit"},{"name":"Sleep cycle and CORE.md","layer":"remember","file":"bench/tools/sleep.py","introduced":"2026-10-27T15:17:41-07:00","introduced_src":"git first-add 3a72a8913cc","date_note":null,"what":"Every 3 h: consolidate lessons, rebuild a <= 150-line CORE.md, check lane transcripts against rules, dream cross-problem tasks.","measured_effect":{"value":"On-policy checks: 963 rule applications vs 159 violations (86% applied, 158 lane checks). Dreams: 6 hits of 285 resolved (2.1%), 402 dreams queued","kind":"measured","src":"bench/manager/sleep/onpolicy/*.json and bench/manager/sleep/dreams.jsonl (read-only count)"},"owner_quote":{"text":"the student has the knowledge but does not use it (paper line the design targets)","src":"bench/manager/SLEEP.md:14"},"file_exists":true,"git_first_add":"2026-10-27T15:17:41-07:00 3a72a8913cc sleep cycle (arXiv 260"},{"name":"Agent traces database","layer":"remember","file":"bench/tools/traces.py","introduced":"2026-10-24T19:45:17-07:00","introduced_src":"git first-add c20ccf2512c","date_note":null,"what":"Every agent's turns, tool calls, ledgers and briefs into one SQLite db; feeds trajectory and brief-quality models.","measured_effect":{"value":"Brief model (brief text -> P(beat prior)) grouped CV AUC 0.877. Logs showed 0 of 7,914 lane runs did any research","kind":"measured","src":"bench/SYSTEMS.md:152; bench/manager/AB_RESEARCH.md:4"},"file_exists":true,"git_first_add":"2026-10-24T19:45:17-07:00 c20ccf2512c fleet to Sol (owner 10"},{"name":"Send pacing and per-problem lock","layer":"guard","file":"bench/tools/okc_run.py","introduced":"2026-10-23T11:50:29-07:00","introduced_src":"git pickaxe 'def released(kind' 2994d96b770","date_note":"okc_run.py itself first-add 2026-10-22T15:12:16 (96c9d26cc4c); the per-problem lock came 2026-10-23T17:16:22-07:00 (b8bd3ca5d70) after two main sessions sent the same problem at 00:03 UTC.","what":"The only sender: releases 10 scored runs through the day, one send per problem at a time, row bar on scored runs.","measured_effect":{"value":"Built after decode spent 10 of 10 scored runs by 9:50 PM PDT 10-22 and its row froze 19 h (rank 2 -> 8). Later the row bar refused cand59 (2.3044 < 2.3408)","kind":"measured","src":"AGENTS.md:88-89; bench/agents/FAMILY_LEARNINGS.md:138"},"owner_quote":{"text":"be careful and distribute submissions carefully","src":"AGENTS.md:79"},"file_exists":true,"git_first_add":"2026-10-22T15:12:16-07:00 96c9d26cc4c Kernel challenge: okc "},{"name":"State-copy gate","layer":"guard","file":"bench/modal/statecopy.py","introduced":"2026-10-22T20:50:47-07:00","introduced_src":"git first-add d758bcde074","date_note":"First-add is a batch commit (message about OWNER_TODO.md); AGENTS.md dates the finding N1300 10-23 03:55 UTC, the same evening.","what":"Re-runs a kernel on tensor-cloned copies of its prepared state, as the judge does, to catch baked raw addresses.","measured_effect":{"value":"Explained all 5 MoE judge FAILs at bs=128 (rel L2 1.414: the kernel read the previous call's buffers from a state copy)","kind":"measured","src":"bench/modal/statecopy.py:4-9; AGENTS.md:126-129"},"file_exists":true,"git_first_add":"2026-10-22T20:50:47-07:00 d758bcde074 main: bench/OWNER_TODO"},{"name":"PDL load-order gate","layer":"guard","file":"bench/audit/gates/pdl_order_gate.py","introduced":"2026-10-23T15:49:03-07:00","introduced_src":"git first-add 8208c46bfe1","date_note":"First-add is a batch commit ('main tick 22:49'); AGENTS.md credits claude-dec-prewait 10-24, under 1 day later.","what":"Reads the SASS of our modules and fails a kernel whose loads of producer output move above griddepcontrol.wait.","measured_effect":{"value":"[sealed until Nov 19]","kind":"measured","src":"AGENTS.md:144"},"file_exists":true,"git_first_add":"2026-10-23T15:49:03-07:00 8208c46bfe1 main tick 22:49: ledge"},{"name":"Paired-plan guard","layer":"guard","file":"bench/tools/paired_guard.py","introduced":"2026-10-30T15:27:35-07:00","introduced_src":"git first-add 20422a89802","date_note":null,"what":"Never spends a paired plan on a base or box that cannot run: base registry, base-first preflight, fault attribution.","measured_effect":{"value":"ABORTED share of paired verdicts: 271 of 564 (48%) from 10-28 22:00 to 10-30 22:00 -> 27 of 221 (12%) from 10-30 22:00 to 11-03 01:32. Docstring pre-count: 380 of 606 plans (63%)","kind":"measured","src":"bench/manager/paired.jsonl (read-only count by 't' field, timezone unmarked); bench/tools/paired_guard.py:5-8"},"owner_quote":{"text":"we are wasting compute. figure out a way to prevent this from happening","src":"bench/tools/paired_guard.py:2-3"},"file_exists":true,"git_first_add":"2026-10-30T15:36:40-07:00 18acbeadf18 sync: main's content a"},{"name":"Idle-box reaper and spend guard","layer":"guard","file":"bench/tools/boxwatch.py","introduced":"2026-10-22T21:06:22-07:00","introduced_src":"git first-add 4dbe3c50c69","date_note":"spend_guard.py first-add 2026-10-23T11:51:43-07:00 (03b44effc5b). boxwatch's first-add commit is a batch commit.","what":"Samples every H200 box; stops boxes idle 15-30 min and orphan apps; spend_guard caps boxes and alarms on burn.","measured_effect":{"value":"Before: $777 for 171 box-hours, 62% idle (10-23 audit). Reaper 10-23..11-05: 403 idle boxes stopped, est. $3,656.53 refunded to the daily cap","kind":"measured","src":"bench/tools/spend_guard.py:3; bench/modal/reaped.jsonl (sum of saved_usd_est, rc=0 rows)"},"owner_quote":{"text":"have a thing which tracks the idle stuff, we are not $$$ rich","src":"bench/tools/boxwatch.py:14"},"file_exists":true,"git_first_add":"2026-10-22T21:06:22-07:00 4dbe3c50c69 main: MoE scored 2.022"},{"name":"Writer provenance gate","layer":"guard","file":"bench/tools/provenance.py","introduced":"2026-10-27T19:35:06-07:00","introduced_src":"git first-add 24d574ec473","date_note":"writer_registry.jsonl first-add 2026-10-27T19:41:12-07:00 (bf310e8bebc). bench/agents/FAMILY_LEARNINGS.md:25 notes the written 'Opus only' rule differs from the running Sonnet writer tier.","what":"Only registered writer models may write source; every file goes into the writer registry and a compile gate before GPU.","measured_effect":{"value":"Registry 3,862 files (Sonnet 2,046, subscription seats 1,504, Opus 312). First-try gate pass: Opus 99.1% of 651, Sonnet 94.6% of 1,404, Haiku 89.1% of 404","kind":"measured","src":"bench/agents/FAMILY_LEARNINGS.md:68,83"},"owner_quote":{"text":"DeepSeek is the default brain; code is written only by Opus 5.5","src":"AGENTS.md:109"},"file_exists":true,"git_first_add":"2026-10-27T19:35:06-07:00 24d574ec473 pi writer: Claude Sonn"},{"name":"Readable-source send refusal","layer":"guard","file":"bench/tools/okc_run.py","introduced":"2026-11-03T13:49:55-08:00","introduced_src":"git pickaxe 'def encoded_source_reason' 085e86cb1f1","date_note":null,"what":"okc_run.py refuses any file with compressed or encoded source (lzma, zlib, base64, decoded exec); no override.","measured_effect":null,"why_null":"Added the day 46 of our runs were disqualified (11-03), one day before the kernel close; no later refusal count was found.","owner_quote":{"text":"submissions must be readable (organizers' rule, 11-02 08:00 PT)","src":"bench/tools/okc_run.py:628"},"file_exists":true,"git_first_add":"2026-10-22T15:12:16-07:00 96c9d26cc4c Kernel challenge: okc "},{"name":"Stop hooks (limit check, turn end)","layer":"guard","file":"bench/tools/turn_end.py","introduced":"2026-10-24T19:25:49Z","introduced_src":"git first-add 273d7d334c6 (limit_check.py)","date_note":"limit_check.py first-add is a cloud-box bulk commit; bench/SYSTEMS.md:136 logs the hook at 10-24 12:15 UTC (7 h earlier). turn_end.py first-add 2026-11-01T17:19:07Z (00a37e36507).","what":"Blocks a session's stop while a problem is below its ceiling or the goal is unmet; the reason becomes the next prompt.","measured_effect":{"value":"Against: about 12 consecutive turns answered the stop hook with a status line and ran no experiment","kind":"measured","src":"bench/books/POSTMORTEM_1029.md:33"},"owner_quote":{"text":"on every single hook ... check if you have actually reached the physical limit ... never stop","src":"bench/tools/limit_check.py:2-3"},"file_exists":true,"git_first_add":"2026-11-01T08:16:15-08:00 d5234baa503 sync: main's content a"},{"name":"H200 boxes","layer":"infra","file":"bench/modal/box.py","introduced":"2026-10-22T17:43:04-07:00","introduced_src":"git first-add 20770817c93","date_note":null,"what":"Rented Modal H200 sandboxes with the judge's image, a FIFO GPU queue, late binding and an in-box judge.","measured_effect":{"value":"1 h rentals died with their queues (10-24: 83 RL runs in 7 h) -> 4 h default. 1 CPU core per box starved nvcc -> 16 cores per box","kind":"measured","src":"bench/SYSTEMS.md:72-73,98"},"owner_quote":{"text":"rent one H200 for exploration ... do it","src":"bench/modal/box.py:3"},"file_exists":true,"git_first_add":"2026-10-22T17:43:04-07:00 20770817c93 bench: H200 box (modal"},{"name":"Object-store sync (push_sync, safe_pull)","layer":"infra","file":"bench/tools/push_sync.py","introduced":"2026-10-23T15:10:45-07:00","introduced_src":"git first-add 424f4a48913","date_note":"safe_pull.py first-add 2026-10-23T15:12:47-07:00 (3ad17dee710).","what":"Push and pull with merge-tree and commit-tree only; ledgers are unioned, so 40 writing agents never lose an append.","measured_effect":{"value":"Built after a git hard reset 10-23 05:09 UTC zeroed island files and lost ~10 ledger writes. Later: 2.27 GB per push try lost every race for > 1 h","kind":"measured","src":"AGENTS.md:214-216; bench/agents/FAMILY_LEARNINGS.md:39"},"file_exists":true,"git_first_add":"2026-10-23T15:10:45-07:00 424f4a48913 push_sync.py: push whe"},{"name":"Error alarm hook","layer":"infra","file":"bench/tools/alarm.py","introduced":"2026-10-23T19:21:41-07:00","introduced_src":"git first-add e9ab9270eef","date_note":null,"what":"Every 2 min checks dead boxes, crashed launches, dead lanes, judge fails, loop tracebacks; Telegram plus inbox.","measured_effect":{"value":"First night: boxes 5 and 8 found dead with live reservations; box.py down refunded $31.42","kind":"measured","src":"bench/SYSTEMS.md:115-116"},"owner_quote":{"text":"is there a HOOK for error or smth u know??","src":"bench/tools/alarm.py:2"},"file_exists":true,"git_first_add":"2026-10-23T19:21:41-07:00 e9ab9270eef alarm.py: the error ho"},{"name":"GPU idle filler","layer":"infra","file":"bench/tools/gpu_fill.py","introduced":"2026-10-24T17:51:19-07:00","introduced_src":"git first-add 8637cdf92b5","date_note":null,"what":"Gives each idle box a strict run of a near-best file, so idle billed GPU time adds evidence.","measured_effect":{"value":"Built when 11 of 13 boxes sat at 0% GPU. Coordinator later reports GPU utilisation 17% -> 31%, and calls it an input, not the objective","kind":"measured","src":"bench/tools/gpu_fill.py:5; bench/books/POSTMORTEM_1029.md:38"},"owner_quote":{"text":"max out so nothing is waiting, faster, but keep pushing","src":"bench/tools/gpu_fill.py:2"},"file_exists":true,"git_first_add":"2026-10-24T17:51:19-07:00 8637cdf92b5 fleet throughput: GPU "}],"prose":{"task_note":["The ceiling is a model, not a bound. No final entry exceeds one. One prefill leader read 2.328x, above our 2.27x model, on 11-02, before it left the top of the board."],"machine":["The swarm was not one agent. It was many lanes. A lane is one agent session with one brief, aimed at one case of one problem.","A lane wrote a candidate file and ran it on an H200 box against a clone of the judge. A paired run on the same box, against the current best file, decided whether to keep it. Only the best files went to the organizers' judge, which allowed a few scored runs per day.","Every idea, trial, note and mistake went into append-only ledgers that outlived every session. The harness grew as problems appeared: 40 pieces over the 14 days, 21 of them in the first two days."],"climb":["Decode moved from 1.168x on the first evening to 2.451x at the close, 110% faster. Our overall rank equalled our decode rank in 402 of 402 snapshots that recorded both, so the decode rank stands in for the overall rank here (a proxy; see the Data page).","Board reads show us #1 on the decode board from 00:20 to 07:06 UTC on Oct 24: 6.8 hours and 38 reads. The next read, at 17:57, showed #3. After our first #1 read our decode row rose 16.7% and the leader's 24.4%. The board grew to 1,300 entries, and the top bunched. At the close, a 6.7% gap to #1 meant #25.","One own goal cost more than any rival. On Nov 2 the organizers banned compressed source. No tool watched their rule page, and lanes kept packing files. On Nov 3 (Pacific time), 46 of our scored runs were disqualified, by the count in our mistake log (believed; other notes give 40 and 48). Decode fell from 2.373x (#34) to 2.250x (#215) until a clean file rebuilt it. MoE fell 4.2% (#32 to #94), attention 2.5% (#19 to #27) and bf16 2.1% (#176 to #224). The three rows did not come back before the close."],"scaling":["The swarm ran 58,618 GPU runs in 13.4 days on 7,610 distinct files. 55,401 were full runs on rented H200 boxes. 533 went to the organizers' judge.","On each problem we fit the gap closed, (x \u2212 1) / (ceiling \u2212 1), against log10 of the GPU runs spent so far. Pooled over the five problems, each tenfold increase went with 17.3 percentage points more of the gap (90% interval 16.8 to 17.7). The scored rows give 17.2. Per problem, log experiments and gap closed correlate at r = 0.91 to 0.97. A best-so-far curve only rises, so it correlates with any clock; the slope per tenfold is the number to plan with. It also mixes compute with new ideas, new tools and new harnesses, so read it as an association.","The steps per decade are uneven. Decode closed 23% of the gap from 100 to 1,000 runs, 6% from 1,000 to 10,000, and 18% per decade after that. In their last decade, prefill closed 34% per decade and bf16 21%.","We also counted 24-hour records: a valid all-shape run more than 0.2% faster than the best valid run of the previous 24 hours on that problem. The swarm set 209. The bar sits inside the noise of one run (0.3% to 0.6% on the clone, believed from our notes): 130 records gained less than 1%, and from Oct 27, 62 of the 72 records with a code hash re-ran bytes that had already run. Records per hour rose with runs per hour to the power 0.50 (0.54 with glitch reads removed) and fell with the runs already spent, to the power -0.52. With a bar this low, more runs produce more records by chance alone.","The first three days, Oct 22 to Oct 24, set 107 of the 209 records with 4% of the GPU runs. On Oct 22 alone, 48 runs set 13 records against near-baseline first files."],"models":["15 model-and-harness pairs ran lanes. A lane run is one agent session on one brief. The chart compares them per lane-hour, which is summed session time. Claude Code, Codex and Antigravity keep no session ledger, so their sessions are rebuilt from GPU activity; that favours them against pi lanes, which also log idle time. Read the chart as a rough ranking.","Who wrote the bytes mattered more than who re-ran them. The writer registry covers few files: 4 credits are measured and 205 are inferred from the first lane that ran the same bytes. On that inference, Claude Opus 5.5 in Claude Code wrote about 123 of 211 record files, GPT-6 Astra in the Codex CLI 25 before its quota ran out, Gemini 4 Argon 27 across two harnesses, and Sonnet, Haiku, Kimi K3 and Terra lanes 2 between them.","Those cheaper lanes ran 66,629 lane runs, mostly reruns and narrow edits. Without a strong writer, run volume added re-measurements, not new code.","Family-level odds ratios are in the data file. The largest, pi lanes with Gemini 4 Argon briefed by the main session (4.34 times the odds of a record per run), is the top of 34 estimates, so it carries a winner's curse. Treat it as a lead, not a result.","A text model on 4,277 lane briefs separated lanes that set a record with a cross-validated AUC of 0.913 (believed: copied from a mid-competition note). Its top features are mostly lane and role names, so it largely learned that confirmation lanes cannot set records. Its prospective test never ran. A pre-registered A/B of mechanism briefs against generic briefs ended 0 records to 0.","A second pre-registered A/B gave some lanes a required paper or rival-code step. Neither arm set a record (11 and 13 lanes), so it says nothing yet."],"worked":["We sorted every move of a scored row, and every landing on the late candidate chain, by the kind of change. The kinds are hand-tagged from run notes, so they are believed, not measured.","Early, new designs paid. From day one to day three, decode went from 1.168x to 2.101x, mostly through new kernels and ports of our own kernels between problems. Structural rewrites hold 31% of the log gain in scored-row moves.","Late, only small edits paid. In the window our notes measured (Oct 31 to Nov 1, 19:45 UTC), all 44 decode hits were routes, ports, stacks, pins or small tweaks of the incumbent (believed). Through the close, full decode trials from Oct 31 hit 2.3%. On the chain, tweaks and stacks of near-misses gave 88% of the gain, at a median of 0.18% per landing.","Borrowed code moved fast. Our notes record that on Oct 27 one afternoon of ports from rivals' open-source entries moved more rows than a day of lane trials (believed, from AGENTS.md). Library heuristics lost to configs found by an exhaustive offline search and frozen per shape.","Ideas priced against the floor did better than average. Ideas the ledger tags as first principles passed a controlled check 29 times in 198 full-problem tries (14.6%), against 6.3% for all sources."],"sources_caption":"Rival or peer code: 10 of 76 tried. Papers: 0 of 22 tried; one-sided Fisher p = 0.068. Source labels come from a keyword map over the idea text, so they are believed.","cost":["The work cost $27,892: $18,616 of GPU time and $9,276 of API-billed model calls. That is total spend minus measured waste: GPU boxes that sat rented with an empty queue, and API model lanes that never produced a result. E2B sandboxes and the subscription fees are not included. The Claude and Codex seats ran on subscriptions; at API list prices their use would add about $29,204 (estimated).","That is $0.48 per GPU run and $133 per 24-hour record. Wins got more expensive as the problems were mined: the period table below shows the trend.","Most late records re-ran old bytes, so new code cost far more. Counting only records whose file had not run before: $511 each from Oct 25 to 27 (12 records), $9,302 from Oct 28 to 31 (2 records) and $3,086 from Nov 1 to 4 (4 records). Counts this small give only an order of magnitude. Records before Oct 25 carry no code hash."],"spend_caption":"UTC days. GPU: Modal bills where they exist, the internal ledger elsewhere, plus Daytona, Prime and Nebius (estimated). Models: provider-reported or calibrated per-run bills from the lane ledgers. Not included: E2B sandboxes, subscription fees.","wrong":["The mistake ledger holds 322 dated rows. Each names the cost and, where possible, the rule or gate that followed.","A written rule did not stop a habit. For the three most frequent classes (sleep / poll wait loops in agent turns, skipped start-of-turn reads (CORE.md / mailbox / first moves), off-brief work), 76% to 98% of their rows were logged after their first rule. The ledger has no comparable count for gates built into tools.","The worst hours were infrastructure, and scale multiplied each single point of failure:"],"outages":["hourly review (HOURLY.md) frozen by one union-merged JSONL line: 26.0 hours","landing loop stopped on 4 problems, nothing alarmed: 13 candidates waited 3.6-25.4 h unjudged: 25.4 hours","Mac disk full: three 22 GB worktrees (67 GB) on a 78 GB .git; confirm loop could not write, 380 plans aborted NO_RESULT, nothing shipped: 16.0 hours","decode writer pipeline stalled: claude6 session limit + claude3/4/5 weekly limits; 37 writer tasks waited up to 6.5 h, pi runners idle (16/24 live): 6.5 hours"],"lost":["On the day each row last improved our places were Decode layer #22 of 1,291, MoE W4A16 #7 of 1,004, Attention decode #10 of 1,151, Prefill layer #19 of 684, BF16 projections #49 of 1,500. Then 4 of the five rows stopped moving: the push went to decode alone from Nov 1, and the other problems kept the row they had on Oct 28 to Oct 30. The field did not stop. Their leaders gained 5.9%, 3.6%, 5.6%, 4.7% after our last gain.","The disqualification then removed the best rows we had left on three of those boards.","Decode, the one board we kept working, finished close to where it was at its last gain. The four we left finished far below it."],"spend_wrong":["Money bought less every day. When the swarm had spent its first $5,000, by 10-25, the geomean of our five best internal rows was 1.97x. The remaining $37,253 moved it to 2.23x.","A large part of the GPU bill was not GPU work. In a measured window from Oct 26 to Oct 28, CPU time on the GPU boxes and builders cost $4,158 against $2,809 for the H200s themselves (58% CPU): kernels compiled on rented GPU machines.","Rented boxes sat idle with an empty queue 38% of the sampled time, about $7,160. An idle reaper stopped hundreds of boxes, but only after they had idled.","On the model side, 55% of the API spend of lanes that run GPU jobs went to rounds that ran no GPU job at all, about $5,222. API lanes that never set a record cost $6,813: Claude Opus 5.5 via pi $3,617, Claude Sonnet 5.5 via Claude Managed Agents $1,664, GPT-6.1 Terra via pi $1,250, GPT-6.1 Terra via OpenAI Agents $219, Claude Haiku 5.5 via pi $63.","The subscriptions did most of the useful work. Claude Opus 5.5 in Claude Code, on flat-fee seats, wrote most of the record files; the API swarm of cheaper models wrote 2."],"plan5k":["Keep every board moving: one writer lane per problem until the close, never a single-problem push.","Compile on CPU machines, time on the GPU: no H200 waits for a compiler.","Rent GPUs only while a queue exists, and stop a box the minute its queue empties.","Give the writer seat to the strongest model on a subscription. Cut API lanes that write no new bytes in a day.","Count a win only when the file is new and the gain clears the noise, so no money buys re-measurements.","Watch the organizers' rule page with a tool, and re-send every row in the allowed form the same day a rule changes."],"important_q":"Our best reading: a check sharp enough to see the gain was scarcer than ideas, agents or GPU runs. The data supports this only in part.","important":["Ideas were plentiful: 14,936 of them. Lane runs were plentiful: 73,533. GPU runs were plentiful: 58,618. Records whose file was new were rare: 18 of the 94 records that carry a code hash.","As the rows rose, real gains shrank toward the noise of a whole-problem run, and whole-problem trials stopped seeing them. Each sharper instrument came with new landings: a paired same-box A/B, then a judge-exact replay, then scoped verdicts on one case.","The evidence against is real. Our own notes say the hit-rate collapse by level is \"real, not only resolution\". With the sharp checks in place, full decode trials from Oct 31 to the close still hit only 2.3%, late gains stayed small, and the gap to #1 was still 6.7% at the close. A sharper check finds small wins. It does not invent the different structure the leader likely had.","What we would do differently: build the high-resolution check before scaling the search; require new bytes for a record; give the strongest model the writer seat and cheap models only the reruns; price every idea against the floor before its first run; read every rival's code per case on day one; and watch the organizers' rule page with a tool."],"important_for":["whole-x hit rate 31.8% -> 2.9% -> 0.5% -> 0% as the decode row rose past 2.0 / 2.2 / 2.3 (SPEC_V3.md:73-78)","decode row flat at 2.264 for ~47 h (board: 10-29 22:10Z to 10-31 21:18Z); WHAT_WORKED reports 1,800+ experiments in that window; jexact + scoped verdicts then landed 2.264 -> 2.3726 by 11-02 15:50Z","scoped verdict CI +-0.35% -> +-0.02%; 7.1% hit rate restored (44/619) (in the Oct 31 to Nov 1 window; through the close the rate was 2.3%)"],"important_against":["flat mining: of 435 positive whole-x flats, ~30 re-measured under the judge-like router gave 1 gene; \"the hit-rate collapse by level is real, not only resolution\" (SPEC_V3.md section 6)","after the instrument upgrades, late gains stayed +0.05..+0.3% each and the gap to #1 stayed ~6-7% (CLOSE_1105.md, board_history)"],"important_alt":[{"title":"Building on what exists (incumbent-first, rival ports, frozen configs) beat invention after day 3","claim":"After the first two days nearly every confirmed gain was a route, port, stack, pin or small edit of the head; the biggest single late steps were ports of rival code."},{"title":"An exact grader plus a single-writer chain let thousands of parallel attempts compound","claim":"Because every candidate was a paired diff on one head per problem, small genes from many agents stacked instead of colliding."},{"title":"Scale was not the lever; scale without yield checks and robust infra limited the swarm","claim":"Commits per day, without automated syncs, rose from 388 on Oct 24 (our #1 day) to 2,504 at the peak while the decode rank fell. The hours lost were mostly infrastructure, quota and process failures that scale multiplied."},{"title":"Why the rank fell while the score rose: the field caught up and bunched","claim":"Our decode x rose 16.69% after our #1 read, but the leader rose 24.45% and the board grew to 1300 entries; with the top bunched, a ~6.7% gap meant #22-25."}]},"problems":[{"key":"muse_decode_layer","short":"dec","label":"Decode layer","slot":1,"what":"A whole decoder layer at decode and speculative verify: attention, the MoE, dense projections and norms, about 10 dependent kernels.","ceiling":3.02,"final":{"t":"2026-11-05T14:38:51Z","x":2.4513545799748853,"rank":25,"entries":1300,"leader":"pjmullin3","leader_x":2.614352808274751,"gap_pct":-6.23,"pct_ceiling":81.2},"first":{"t":"2026-10-22T22:55:50Z","x":1.1682},"best_ever_scored":2.4514,"best":[["2026-10-22T22:56:48Z",1.1682],["2026-10-23T00:08:58Z",1.2598],["2026-10-23T01:08:55Z",1.4319],["2026-10-23T02:03:37Z",1.4952],["2026-10-23T02:19:41Z",1.4953],["2026-10-23T03:30:28Z",1.6325],["2026-10-23T03:50:28Z",1.6906],["2026-10-23T04:50:54Z",1.7547],["2026-10-24T00:54:59Z",2.0821],["2026-10-24T01:06:00Z",2.1008],["2026-10-25T00:04:44Z",2.1146],["2026-10-25T06:06:47Z",2.162],["2026-10-27T02:44:41Z",2.204],["2026-10-27T17:29:03Z",2.2134],["2026-10-28T22:36:53Z",2.2496],["2026-10-29T22:10:56Z",2.264],["2026-10-31T21:18:51Z",2.2866],["2026-10-31T23:38:12Z",2.3317],["2026-11-01T12:37:48Z",2.3408],["2026-11-02T05:22:05Z",2.3499],["2026-11-02T15:50:48Z",2.3726],["2026-11-03T21:29:02Z",2.3655],["2026-11-03T21:44:34Z",2.3647],["2026-11-04T00:43:38Z",2.2496],["2026-11-04T01:33:22Z",2.3687],["2026-11-04T03:22:28Z",2.3702],["2026-11-04T05:51:49Z",2.3765],["2026-11-04T06:45:23Z",2.3967],["2026-11-04T13:28:24Z",2.4019],["2026-11-04T20:02:57Z",2.4034],["2026-11-05T00:11:49Z",2.4234],["2026-11-05T03:16:27Z",2.4514]],"leader":[["2026-10-22T22:43:14Z",1.4116],["2026-10-22T23:17:51Z",1.5418],["2026-10-23T00:07:21Z",1.6402],["2026-10-23T00:47:24Z",1.6768],["2026-10-23T00:51:58Z",1.7845],["2026-10-23T01:07:43Z",1.808],["2026-10-23T01:25:26Z",1.811],["2026-10-23T02:10:55Z",1.8269],["2026-10-23T03:30:28Z",1.8642],["2026-10-23T18:13:14Z",1.862],["2026-10-23T18:44:50Z",1.8916],["2026-10-23T23:08:51Z",1.9353],["2026-10-24T00:20:23Z",2.0821],["2026-10-24T01:06:00Z",2.1008],["2026-10-24T17:57:27Z",2.1803],["2026-10-24T19:24:37Z",2.1939],["2026-10-24T19:59:06Z",2.2849],["2026-10-25T22:10:45Z",2.3104],["2026-10-25T22:15:40Z",2.3124],["2026-10-26T19:59:18Z",2.3797],["2026-10-27T10:36:00Z",2.3972],["2026-10-27T11:09:55Z",2.3986],["2026-10-28T10:04:57Z",2.4029],["2026-10-29T16:06:41Z",2.4178],["2026-10-31T20:12:21Z",2.4298],["2026-10-31T20:19:34Z",2.4592],["2026-10-31T22:39:11Z",2.469],["2026-11-01T02:09:14Z",2.4844],["2026-11-02T18:01:17Z",2.4884],["2026-11-02T23:39:21Z",2.4902],["2026-11-03T13:53:25Z",2.5098],["2026-11-03T19:17:24Z",2.5236],["2026-11-03T20:53:46Z",2.5533],["2026-11-04T01:18:51Z",2.5226],["2026-11-04T05:10:02Z",2.5401],["2026-11-04T10:07:18Z",2.551],["2026-11-04T14:12:54Z",2.5725],["2026-11-04T14:18:59Z",2.5762],["2026-11-05T04:25:44Z",2.5772],["2026-11-05T07:11:32Z",2.5876],["2026-11-05T07:41:34Z",2.6144]],"rank":[["2026-10-22T22:56:48Z",27,33],["2026-10-22T23:17:51Z",28,38],["2026-10-22T23:28:21Z",29,41],["2026-10-22T23:30:48Z",30,42],["2026-10-22T23:31:50Z",31,42],["2026-10-22T23:33:30Z",32,43],["2026-10-22T23:34:58Z",33,44],["2026-10-22T23:36:58Z",34,44],["2026-10-22T23:45:50Z",35,47],["2026-10-22T23:59:27Z",36,49],["2026-10-23T00:08:58Z",27,49],["2026-10-23T00:11:58Z",28,50],["2026-10-23T00:47:24Z",29,56],["2026-10-23T00:54:52Z",31,58],["2026-10-23T01:08:55Z",9,59],["2026-10-23T01:17:51Z",10,60],["2026-10-23T02:03:37Z",6,63],["2026-10-23T03:30:28Z",5,82],["2026-10-23T03:50:28Z",2,88],["2026-10-23T04:13:17Z",3,95],["2026-10-23T04:50:54Z",2,106],["2026-10-23T18:13:14Z",8,272],["2026-10-23T18:59:22Z",9,278],["2026-10-23T19:52:08Z",10,287],["2026-10-23T22:12:41Z",11,319],["2026-10-23T22:21:22Z",12,321],["2026-10-23T23:43:42Z",13,339],["2026-10-24T00:20:23Z",1,348],["2026-10-24T17:57:27Z",3,522],["2026-10-24T20:40:02Z",4,545],["2026-10-25T00:18:39Z",6,582],["2026-10-25T02:33:21Z",7,591],["2026-10-25T05:46:39Z",8,609],["2026-10-25T06:06:47Z",5,609],["2026-10-25T07:16:53Z",6,618],["2026-10-25T18:58:51Z",10,670],["2026-10-25T22:29:38Z",11,690],["2026-10-26T19:59:18Z",15,746],["2026-10-26T23:18:10Z",16,754],["2026-10-27T02:44:41Z",11,759],["2026-10-27T05:21:18Z",12,765],["2026-10-27T10:50:09Z",13,782],["2026-10-27T12:39:13Z",14,788],["2026-10-27T17:29:03Z",13,800],["2026-10-27T17:58:38Z",14,802],["2026-10-27T19:59:12Z",15,811],["2026-10-27T22:54:29Z",16,821],["2026-10-28T01:54:38Z",17,823],["2026-10-28T02:39:13Z",18,823],["2026-10-28T03:06:55Z",19,823],["2026-10-28T04:28:22Z",20,828],["2026-10-28T04:54:27Z",21,829],["2026-10-28T05:00:38Z",22,830],["2026-10-28T06:07:17Z",23,835],["2026-10-28T19:02:32Z",24,873],["2026-10-28T22:36:53Z",15,879],["2026-10-29T00:27:39Z",16,882],["2026-10-29T05:29:36Z",17,892],["2026-10-29T05:48:17Z",18,892],["2026-10-29T16:06:41Z",21,952],["2026-10-29T16:57:56Z",22,957],["2026-10-29T18:46:24Z",23,973],["2026-10-29T20:37:19Z",24,988],["2026-10-29T22:10:56Z",18,996],["2026-10-29T22:30:08Z",19,998],["2026-10-30T01:57:58Z",20,1002],["2026-10-30T02:11:39Z",21,1003],["2026-10-30T04:13:13Z",22,1008],["2026-10-30T04:41:19Z",23,1008],["2026-10-30T08:40:35Z",24,1021],["2026-10-30T09:06:55Z",25,1023],["2026-10-30T09:20:10Z",26,1024],["2026-10-30T11:32:49Z",27,1031],["2026-10-30T16:01:47Z",28,1048],["2026-10-30T16:48:53Z",29,1054],["2026-10-30T17:26:22Z",30,1057],["2026-10-30T18:32:21Z",31,1060],["2026-10-30T18:44:28Z",32,1060],["2026-10-30T18:56:59Z",33,1061],["2026-10-30T19:41:27Z",34,1062],["2026-10-30T21:51:14Z",35,1064],["2026-10-30T22:54:20Z",36,1069],["2026-10-31T02:18:53Z",37,1078],["2026-10-31T02:45:25Z",38,1080],["2026-10-31T06:13:11Z",39,1088],["2026-10-31T12:34:41Z",42,1095],["2026-10-31T16:23:54Z",43,1100],["2026-10-31T16:58:34Z",44,1101],["2026-10-31T18:50:56Z",45,1109],["2026-10-31T20:12:21Z",46,1110],["2026-10-31T20:26:32Z",47,1113],["2026-10-31T20:54:37Z",48,1113],["2026-10-31T21:18:51Z",41,1115],["2026-10-31T23:38:12Z",20,1124],["2026-11-01T00:15:48Z",21,1127],["2026-11-01T00:30:02Z",22,1128],["2026-11-01T01:30:09Z",23,1134],["2026-11-01T01:50:13Z",24,1134],["2026-11-01T02:21:26Z",25,1135],["2026-11-01T03:46:41Z",26,1138],["2026-11-01T06:20:24Z",27,1146],["2026-11-01T06:26:18Z",28,1146],["2026-11-01T09:28:12Z",29,1152],["2026-11-01T10:38:19Z",30,1156],["2026-11-01T11:50:36Z",31,1160],["2026-11-01T12:10:49Z",32,1160],["2026-11-01T12:37:48Z",30,1160],["2026-11-01T12:50:48Z",31,1160],["2026-11-01T13:37:55Z",32,1163],["2026-11-02T01:18:05Z",33,1173],["2026-11-02T01:32:03Z",34,1173],["2026-11-02T05:22:05Z",29,1176],["2026-11-02T08:24:50Z",27,1180],["2026-11-02T09:29:05Z",28,1180],["2026-11-02T11:39:36Z",29,1182],["2026-11-02T11:46:39Z",30,1183],["2026-11-02T14:16:33Z",31,1186],["2026-11-02T15:04:13Z",32,1186],["2026-11-02T15:50:48Z",22,1187],["2026-11-02T18:57:50Z",21,1195],["2026-11-02T20:26:40Z",22,1198],["2026-11-02T21:51:49Z",24,1199],["2026-11-02T23:39:21Z",25,1206],["2026-11-03T03:18:41Z",24,1216],["2026-11-03T04:05:11Z",25,1217],["2026-11-03T05:28:01Z",26,1218],["2026-11-03T06:40:08Z",27,1221],["2026-11-03T06:59:38Z",28,1222],["2026-11-03T07:21:06Z",29,1224],["2026-11-03T07:33:08Z",28,1224],["2026-11-03T08:07:43Z",29,1225],["2026-11-03T10:56:22Z",30,1229],["2026-11-03T11:51:07Z",31,1233],["2026-11-03T13:33:26Z",32,1234],["2026-11-03T13:53:25Z",33,1235],["2026-11-03T15:16:49Z",34,1239],["2026-11-03T21:29:02Z",38,1249],["2026-11-03T21:36:38Z",37,1246],["2026-11-03T21:44:34Z",35,1234],["2026-11-03T22:06:08Z",36,1235],["2026-11-03T22:40:09Z",37,1237],["2026-11-03T23:14:23Z",38,1240],["2026-11-04T00:43:38Z",215,1242],["2026-11-04T00:57:35Z",214,1241],["2026-11-04T01:06:38Z",213,1240],["2026-11-04T01:25:22Z",214,1242],["2026-11-04T01:33:22Z",36,1242],["2026-11-04T01:39:23Z",35,1242],["2026-11-04T02:16:42Z",36,1244],["2026-11-04T02:46:29Z",37,1246],["2026-11-04T03:22:28Z",35,1247],["2026-11-04T03:51:38Z",36,1247],["2026-11-04T04:21:26Z",37,1248],["2026-11-04T04:36:19Z",38,1249],["2026-11-04T05:23:24Z",39,1250],["2026-11-04T05:45:40Z",40,1251],["2026-11-04T05:51:49Z",38,1252],["2026-11-04T06:21:25Z",39,1252],["2026-11-04T06:45:23Z",33,1252],["2026-11-04T10:07:18Z",34,1263],["2026-11-04T13:21:30Z",35,1267],["2026-11-04T13:28:24Z",34,1267],["2026-11-04T14:31:49Z",35,1268],["2026-11-04T17:31:22Z",36,1274],["2026-11-04T21:14:19Z",37,1279],["2026-11-04T23:32:22Z",38,1284],["2026-11-04T23:51:25Z",39,1285],["2026-11-05T00:11:49Z",33,1285],["2026-11-05T00:59:59Z",34,1285],["2026-11-05T03:16:27Z",22,1291],["2026-11-05T07:17:25Z",23,1297],["2026-11-05T07:41:34Z",24,1298],["2026-11-05T07:55:05Z",25,1300]],"dq":{"row_before":2.3726,"rank_before":34,"row_after":2.2496,"rank_after":215,"x_change_pct":-5.18,"t_before":"2026-11-03T21:20:35Z","t_after":"2026-11-04T00:43:38Z"},"judge_over_clone":0.9631,"fits":{"internal":{"x_per_decade":0.3061,"x_per_decade_ci90":[0.2974,0.3244],"intercept_x_at_1_run":1.2677,"gap_closed_per_decade":0.146,"gap_closed_per_decade_ci90":[0.1418,0.1547],"fit_range_runs":[24,26959],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":null,"x_to":1.8306,"x_gain":null,"decades":1.0,"gap_closed_per_decade_local":null,"gap_closed_gain":null},{"runs_from":100,"runs_to":1000,"x_from":1.8306,"x_to":2.3047,"x_gain":0.4741,"decades":1.0,"gap_closed_per_decade_local":0.226,"gap_closed_gain":0.226},{"runs_from":1000,"runs_to":10000,"x_from":2.3047,"x_to":2.4245,"x_gain":0.1198,"decades":1.0,"gap_closed_per_decade_local":0.0571,"gap_closed_gain":0.0571},{"runs_from":10000,"runs_to":26959,"x_from":2.4245,"x_to":2.5913,"x_gain":0.1668,"decades":0.431,"gap_closed_per_decade_local":0.1846,"gap_closed_gain":0.0795}]},"scored":{"x_per_decade":0.3295,"x_per_decade_ci90":[0.3024,0.3769],"intercept_x_at_1_run":1.0048,"gap_closed_per_decade":0.1631,"gap_closed_per_decade_ci90":[0.1497,0.1866],"fit_range_runs":[10,26959],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":1.2598,"x_to":1.7547,"x_gain":0.4949,"decades":1.0,"gap_closed_per_decade_local":0.245,"gap_closed_gain":0.245},{"runs_from":100,"runs_to":1000,"x_from":1.7547,"x_to":2.162,"x_gain":0.4073,"decades":1.0,"gap_closed_per_decade_local":0.2016,"gap_closed_gain":0.2016},{"runs_from":1000,"runs_to":10000,"x_from":2.162,"x_to":2.2496,"x_gain":0.0876,"decades":1.0,"gap_closed_per_decade_local":0.0434,"gap_closed_gain":0.0434},{"runs_from":10000,"runs_to":26959,"x_from":2.2496,"x_to":2.4514,"x_gain":0.2018,"decades":0.431,"gap_closed_per_decade_local":0.2319,"gap_closed_gain":0.0999}]}}},{"key":"muse_moe_w4a16","short":"moe","label":"MoE W4A16","slot":2,"what":"The NVFP4 mixture-of-experts block: routing, dequantization, two grouped GEMMs and the combine.","ceiling":3.07,"final":{"t":"2026-11-05T08:14:44Z","x":2.4292289478838875,"rank":122,"entries":1426,"leader":"pjmullin3","leader_x":2.843799916672147,"gap_pct":-14.58,"pct_ceiling":79.1},"first":{"t":"2026-10-22T23:09:46Z","x":1.2712},"best_ever_scored":2.5343,"best":[["2026-10-22T23:17:51Z",1.2712],["2026-10-23T00:47:24Z",1.5188],["2026-10-23T03:30:28Z",1.6724],["2026-10-23T03:59:37Z",1.9607],["2026-10-23T04:05:06Z",2.0217],["2026-10-23T04:24:01Z",2.0784],["2026-10-23T05:36:07Z",2.1235],["2026-10-23T06:08:30Z",2.1551],["2026-10-23T07:39:59Z",2.1801],["2026-10-24T01:59:34Z",2.3353],["2026-10-24T05:46:08Z",2.3494],["2026-10-25T00:04:44Z",2.3776],["2026-10-27T04:49:24Z",2.3859],["2026-10-27T14:18:23Z",2.4292],["2026-10-28T11:46:49Z",2.4535],["2026-10-28T22:41:24Z",2.5343],["2026-11-03T21:44:34Z",2.4292]],"leader":[["2026-10-22T22:43:14Z",1.7532],["2026-10-22T22:44:26Z",1.783],["2026-10-22T23:36:58Z",1.824],["2026-10-22T23:48:40Z",1.9681],["2026-10-23T00:05:51Z",2.0302],["2026-10-23T00:47:24Z",2.1189],["2026-10-23T00:49:58Z",2.1725],["2026-10-23T01:13:53Z",2.1902],["2026-10-23T01:35:30Z",2.1967],["2026-10-23T03:30:28Z",2.3259],["2026-10-24T01:59:34Z",2.3353],["2026-10-24T05:46:08Z",2.3494],["2026-10-24T17:57:27Z",2.5227],["2026-10-25T22:25:41Z",2.5481],["2026-10-25T23:11:41Z",2.5535],["2026-10-26T19:59:18Z",2.5763],["2026-10-26T23:30:26Z",2.6196],["2026-10-28T01:22:59Z",2.6228],["2026-10-28T01:48:12Z",2.6369],["2026-10-28T02:45:48Z",2.6853],["2026-10-29T00:49:19Z",2.6888],["2026-10-29T20:10:11Z",2.6915],["2026-10-30T00:39:54Z",2.6995],["2026-10-30T06:42:44Z",2.7578],["2026-10-31T22:45:37Z",2.7642],["2026-11-04T10:20:53Z",2.7817],["2026-11-04T22:51:08Z",2.7877],["2026-11-05T05:38:04Z",2.8003],["2026-11-05T06:58:12Z",2.8317],["2026-11-05T07:04:25Z",2.8331],["2026-11-05T07:17:25Z",2.8438]],"rank":[["2026-10-22T23:17:51Z",39,54],["2026-10-22T23:24:18Z",40,55],["2026-10-22T23:26:50Z",41,57],["2026-10-22T23:29:40Z",42,58],["2026-10-22T23:34:58Z",43,58],["2026-10-22T23:42:58Z",44,59],["2026-10-22T23:45:50Z",45,60],["2026-10-22T23:59:27Z",49,65],["2026-10-23T00:47:24Z",15,73],["2026-10-23T00:59:18Z",16,75],["2026-10-23T01:11:24Z",17,79],["2026-10-23T01:24:17Z",18,82],["2026-10-23T01:38:58Z",19,85],["2026-10-23T02:02:26Z",20,86],["2026-10-23T02:03:37Z",21,87],["2026-10-23T02:06:27Z",22,87],["2026-10-23T02:20:58Z",23,92],["2026-10-23T03:30:28Z",20,108],["2026-10-23T03:48:58Z",21,118],["2026-10-23T03:59:37Z",2,118],["2026-10-23T10:35:02Z",3,245],["2026-10-23T18:13:14Z",4,359],["2026-10-23T20:23:00Z",5,389],["2026-10-23T21:36:18Z",6,410],["2026-10-23T22:28:51Z",7,419],["2026-10-24T00:17:44Z",8,448],["2026-10-24T01:59:34Z",1,468],["2026-10-24T17:57:27Z",3,614],["2026-10-24T19:23:06Z",4,630],["2026-10-24T22:26:05Z",5,659],["2026-10-24T23:23:26Z",6,668],["2026-10-25T00:04:44Z",3,675],["2026-10-25T04:15:20Z",4,699],["2026-10-25T05:46:39Z",5,705],["2026-10-25T06:02:20Z",6,705],["2026-10-25T07:16:53Z",7,712],["2026-10-25T07:35:37Z",8,713],["2026-10-25T08:16:32Z",9,718],["2026-10-25T08:18:05Z",8,718],["2026-10-25T18:58:51Z",9,764],["2026-10-26T19:59:18Z",20,853],["2026-10-26T20:36:42Z",21,859],["2026-10-26T21:06:35Z",22,861],["2026-10-27T04:49:24Z",21,884],["2026-10-27T09:59:36Z",22,899],["2026-10-27T12:00:50Z",23,904],["2026-10-27T14:18:23Z",17,913],["2026-10-27T16:24:28Z",18,921],["2026-10-27T22:05:36Z",19,934],["2026-10-27T23:49:07Z",20,943],["2026-10-28T07:44:58Z",21,965],["2026-10-28T11:46:49Z",17,977],["2026-10-28T20:20:28Z",18,1001],["2026-10-28T20:33:51Z",19,1002],["2026-10-28T20:38:03Z",20,1003],["2026-10-28T22:41:24Z",7,1004],["2026-10-28T22:43:27Z",8,1004],["2026-10-29T16:06:41Z",10,1087],["2026-10-29T17:17:24Z",11,1096],["2026-10-29T23:23:55Z",12,1131],["2026-10-30T01:51:42Z",13,1140],["2026-10-30T03:07:46Z",14,1144],["2026-10-30T15:29:51Z",15,1190],["2026-10-30T18:08:11Z",16,1200],["2026-10-30T23:19:09Z",17,1207],["2026-10-31T00:15:20Z",18,1207],["2026-10-31T00:27:13Z",19,1208],["2026-10-31T06:13:11Z",20,1221],["2026-10-31T16:23:54Z",22,1244],["2026-10-31T18:24:12Z",23,1249],["2026-11-01T04:51:50Z",24,1274],["2026-11-01T09:28:12Z",25,1282],["2026-11-01T17:09:39Z",26,1298],["2026-11-02T00:57:53Z",27,1301],["2026-11-02T07:57:39Z",28,1308],["2026-11-02T20:18:52Z",29,1320],["2026-11-03T00:06:58Z",30,1331],["2026-11-03T08:13:28Z",31,1349],["2026-11-03T11:30:38Z",32,1355],["2026-11-03T17:19:26Z",31,1364],["2026-11-03T18:39:22Z",32,1369],["2026-11-03T21:44:34Z",94,1369],["2026-11-03T22:06:08Z",95,1370],["2026-11-04T00:35:56Z",97,1377],["2026-11-04T05:23:24Z",98,1382],["2026-11-04T06:14:42Z",99,1382],["2026-11-04T07:06:11Z",100,1386],["2026-11-04T10:00:51Z",101,1391],["2026-11-04T10:20:53Z",102,1393],["2026-11-04T12:08:10Z",103,1395],["2026-11-04T13:02:39Z",105,1396],["2026-11-04T15:53:23Z",106,1401],["2026-11-04T16:33:37Z",107,1401],["2026-11-04T17:31:22Z",108,1404],["2026-11-04T19:07:24Z",109,1404],["2026-11-04T19:36:02Z",110,1404],["2026-11-04T19:50:04Z",111,1404],["2026-11-04T19:53:08Z",112,1405],["2026-11-04T20:02:57Z",113,1405],["2026-11-04T23:58:37Z",114,1412],["2026-11-05T02:49:20Z",115,1416],["2026-11-05T03:22:30Z",116,1419],["2026-11-05T05:44:02Z",117,1423],["2026-11-05T06:44:40Z",118,1424],["2026-11-05T06:51:21Z",119,1424],["2026-11-05T07:23:27Z",120,1425],["2026-11-05T07:55:05Z",122,1427]],"dq":{"row_before":2.5343,"rank_before":32,"row_after":2.4292,"rank_after":94,"x_change_pct":-4.15,"t_before":"2026-11-03T21:36:38Z","t_after":"2026-11-03T21:44:34Z"},"judge_over_clone":0.9633,"fits":{"internal":{"x_per_decade":0.4013,"x_per_decade_ci90":[0.3877,0.4359],"intercept_x_at_1_run":1.2498,"gap_closed_per_decade":0.1867,"gap_closed_per_decade_ci90":[0.1804,0.2029],"fit_range_runs":[22,9297],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":null,"x_to":1.9115,"x_gain":null,"decades":1.0,"gap_closed_per_decade_local":null,"gap_closed_gain":null},{"runs_from":100,"runs_to":1000,"x_from":1.9115,"x_to":2.5172,"x_gain":0.6057,"decades":1.0,"gap_closed_per_decade_local":0.2819,"gap_closed_gain":0.2819},{"runs_from":1000,"runs_to":9297,"x_from":2.5172,"x_to":2.6803,"x_gain":0.1631,"decades":0.968,"gap_closed_per_decade_local":0.0784,"gap_closed_gain":0.0759}]},"scored":{"x_per_decade":0.4414,"x_per_decade_ci90":[0.2985,0.5313],"intercept_x_at_1_run":0.9554,"gap_closed_per_decade":0.2133,"gap_closed_per_decade_ci90":[0.1442,0.2567],"fit_range_runs":[10,9297],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":1.2712,"x_to":1.6724,"x_gain":0.4013,"decades":1.0,"gap_closed_per_decade_local":0.1938,"gap_closed_gain":0.1938},{"runs_from":100,"runs_to":1000,"x_from":1.6724,"x_to":2.3859,"x_gain":0.7135,"decades":1.0,"gap_closed_per_decade_local":0.3447,"gap_closed_gain":0.3447},{"runs_from":1000,"runs_to":9297,"x_from":2.3859,"x_to":2.5343,"x_gain":0.1484,"decades":0.968,"gap_closed_per_decade_local":0.0741,"gap_closed_gain":0.0717}]}}},{"key":"muse_attention_decode","short":"attn","label":"Attention decode","slot":3,"what":"Grouped-query attention over a paged FP8 KV cache at 80K context, one token or an 8-token verify block.","ceiling":3.38,"final":{"t":"2026-11-05T19:10:41Z","x":2.835522644274862,"rank":42,"entries":1380,"leader":"pjmullin3","leader_x":3.3389207412481454,"gap_pct":-15.08,"pct_ceiling":83.9},"first":{"t":"2026-10-22T23:00:46Z","x":1.4804},"best_ever_scored":2.9098,"best":[["2026-10-22T23:17:51Z",1.4804],["2026-10-23T00:47:24Z",1.55],["2026-10-23T01:34:12Z",1.6831],["2026-10-23T02:13:36Z",1.7764],["2026-10-23T03:30:28Z",1.9129],["2026-10-23T04:02:32Z",2.0606],["2026-10-23T04:36:01Z",2.1312],["2026-10-23T05:05:58Z",2.1453],["2026-10-24T01:59:34Z",2.3549],["2026-10-24T05:46:08Z",2.3991],["2026-10-24T22:12:53Z",2.4116],["2026-10-25T01:47:01Z",2.4739],["2026-10-25T07:16:53Z",2.5287],["2026-10-27T01:26:50Z",2.5719],["2026-10-27T16:09:27Z",2.5972],["2026-10-27T20:09:07Z",2.6195],["2026-10-27T23:32:42Z",2.6822],["2026-10-28T09:06:40Z",2.7057],["2026-10-28T18:26:17Z",2.7123],["2026-10-28T21:59:37Z",2.7376],["2026-10-28T22:19:03Z",2.8113],["2026-10-29T22:43:02Z",2.8289],["2026-10-29T23:03:28Z",2.8326],["2026-10-30T00:33:19Z",2.8355],["2026-10-30T23:24:58Z",2.9098],["2026-11-03T21:36:38Z",2.8355]],"leader":[["2026-10-22T22:43:14Z",1.7207],["2026-10-22T23:17:51Z",1.7627],["2026-10-22T23:47:10Z",1.8761],["2026-10-22T23:59:27Z",1.9417],["2026-10-23T00:13:58Z",2.0666],["2026-10-23T00:47:24Z",2.1879],["2026-10-23T00:51:58Z",2.2381],["2026-10-23T01:08:55Z",2.2565],["2026-10-23T01:27:17Z",2.3024],["2026-10-23T01:58:56Z",2.2381],["2026-10-23T02:10:55Z",2.2902],["2026-10-23T03:30:28Z",2.3659],["2026-10-23T20:40:32Z",2.7257],["2026-10-24T17:57:27Z",2.8337],["2026-10-24T20:04:35Z",2.844],["2026-10-24T22:54:47Z",2.9097],["2026-10-26T19:59:18Z",3.0316],["2026-10-28T22:35:00Z",3.0542],["2026-10-30T00:53:23Z",3.2221],["2026-10-31T22:14:56Z",3.3389]],"rank":[["2026-10-22T23:17:51Z",11,41],["2026-10-23T00:47:24Z",7,59],["2026-10-23T00:51:58Z",8,59],["2026-10-23T01:35:30Z",6,66],["2026-10-23T01:55:57Z",7,69],["2026-10-23T01:57:27Z",8,69],["2026-10-23T02:13:36Z",4,73],["2026-10-23T03:30:28Z",3,84],["2026-10-23T04:02:32Z",2,94],["2026-10-23T18:13:14Z",3,312],["2026-10-23T18:50:02Z",4,321],["2026-10-23T20:40:32Z",5,341],["2026-10-23T22:19:53Z",6,358],["2026-10-23T22:45:09Z",7,366],["2026-10-23T22:53:33Z",8,369],["2026-10-23T23:43:42Z",9,388],["2026-10-24T00:16:02Z",10,396],["2026-10-24T01:59:34Z",5,412],["2026-10-24T03:01:39Z",6,423],["2026-10-24T05:46:08Z",5,439],["2026-10-24T07:04:02Z",6,458],["2026-10-24T17:57:27Z",9,570],["2026-10-24T20:04:35Z",10,594],["2026-10-24T21:19:17Z",11,603],["2026-10-24T21:55:52Z",12,612],["2026-10-24T22:14:20Z",11,615],["2026-10-24T22:16:38Z",12,616],["2026-10-24T23:42:24Z",13,632],["2026-10-25T00:14:49Z",14,636],["2026-10-25T01:47:01Z",11,642],["2026-10-25T04:03:34Z",12,653],["2026-10-25T04:55:53Z",11,657],["2026-10-25T05:46:39Z",12,659],["2026-10-25T06:10:23Z",13,661],["2026-10-25T06:30:07Z",14,662],["2026-10-25T07:16:53Z",9,670],["2026-10-25T18:58:51Z",13,718],["2026-10-26T19:59:18Z",23,815],["2026-10-26T21:18:35Z",24,823],["2026-10-27T00:14:04Z",25,829],["2026-10-27T01:26:50Z",16,830],["2026-10-27T03:40:16Z",17,839],["2026-10-27T04:03:25Z",18,840],["2026-10-27T06:51:17Z",19,845],["2026-10-27T09:39:17Z",20,853],["2026-10-27T10:21:22Z",21,856],["2026-10-27T12:21:01Z",22,862],["2026-10-27T12:28:42Z",23,862],["2026-10-27T15:06:48Z",24,869],["2026-10-27T15:28:58Z",25,869],["2026-10-27T16:09:27Z",22,870],["2026-10-27T16:31:40Z",23,872],["2026-10-27T18:13:55Z",24,879],["2026-10-27T18:38:16Z",25,882],["2026-10-27T19:26:42Z",26,885],["2026-10-27T19:28:42Z",27,885],["2026-10-27T19:55:01Z",28,886],["2026-10-27T20:07:08Z",29,887],["2026-10-27T20:09:07Z",21,887],["2026-10-27T20:33:26Z",22,887],["2026-10-27T22:20:02Z",23,894],["2026-10-27T23:32:42Z",9,899],["2026-10-27T23:49:07Z",10,900],["2026-10-28T00:32:53Z",9,900],["2026-10-28T02:10:50Z",10,904],["2026-10-28T04:10:01Z",11,909],["2026-10-28T09:02:22Z",12,924],["2026-10-28T09:06:40Z",9,924],["2026-10-28T10:11:07Z",10,929],["2026-10-28T12:50:50Z",11,939],["2026-10-28T13:00:03Z",12,941],["2026-10-28T18:26:17Z",11,955],["2026-10-28T21:59:37Z",10,961],["2026-10-28T22:19:03Z",6,961],["2026-10-28T22:35:00Z",7,962],["2026-10-29T02:52:46Z",8,973],["2026-10-29T16:06:41Z",9,1042],["2026-10-29T16:18:22Z",10,1044],["2026-10-29T19:04:20Z",11,1064],["2026-10-29T20:55:11Z",12,1077],["2026-10-29T22:43:02Z",11,1085],["2026-10-30T00:19:32Z",12,1089],["2026-10-30T05:27:24Z",13,1105],["2026-10-30T23:24:58Z",10,1151],["2026-10-31T23:11:22Z",11,1202],["2026-11-01T11:24:29Z",12,1233],["2026-11-01T14:10:05Z",13,1235],["2026-11-01T17:38:02Z",14,1242],["2026-11-01T23:47:34Z",15,1245],["2026-11-02T03:11:29Z",16,1246],["2026-11-03T04:18:59Z",17,1285],["2026-11-03T05:11:58Z",18,1286],["2026-11-03T14:51:44Z",19,1308],["2026-11-03T21:36:38Z",27,1319],["2026-11-04T01:12:40Z",28,1321],["2026-11-04T02:16:42Z",29,1324],["2026-11-04T02:52:42Z",30,1325],["2026-11-04T05:02:46Z",31,1328],["2026-11-04T08:07:20Z",32,1333],["2026-11-04T09:21:03Z",33,1335],["2026-11-04T11:08:03Z",34,1337],["2026-11-04T13:02:39Z",35,1341],["2026-11-04T13:41:30Z",36,1342],["2026-11-04T16:42:21Z",37,1346],["2026-11-04T22:10:07Z",38,1354],["2026-11-05T03:29:26Z",37,1363],["2026-11-05T03:37:02Z",38,1364],["2026-11-05T04:18:48Z",39,1367],["2026-11-05T05:50:08Z",40,1371]],"dq":{"row_before":2.9098,"rank_before":19,"row_after":2.8355,"rank_after":27,"x_change_pct":-2.55,"t_before":"2026-11-03T21:29:02Z","t_after":"2026-11-03T21:36:38Z"},"judge_over_clone":0.9787,"fits":{"internal":{"x_per_decade":0.4628,"x_per_decade_ci90":[0.4202,0.4787],"intercept_x_at_1_run":1.3074,"gap_closed_per_decade":0.1903,"gap_closed_per_decade_ci90":[0.1728,0.1968],"fit_range_runs":[28,12256],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":null,"x_to":2.3377,"x_gain":null,"decades":1.0,"gap_closed_per_decade_local":null,"gap_closed_gain":null},{"runs_from":100,"runs_to":1000,"x_from":2.3377,"x_to":2.7275,"x_gain":0.3898,"decades":1.0,"gap_closed_per_decade_local":0.1603,"gap_closed_gain":0.1603},{"runs_from":1000,"runs_to":10000,"x_from":2.7275,"x_to":3.0478,"x_gain":0.3202,"decades":1.0,"gap_closed_per_decade_local":0.1317,"gap_closed_gain":0.1317},{"runs_from":10000,"runs_to":12256,"x_from":3.0478,"x_to":3.1704,"x_gain":0.1227,"decades":0.088,"gap_closed_per_decade_local":null,"gap_closed_gain":0.0504}]},"scored":{"x_per_decade":0.4754,"x_per_decade_ci90":[0.4146,0.5019],"intercept_x_at_1_run":1.0448,"gap_closed_per_decade":0.1997,"gap_closed_per_decade_ci90":[0.1742,0.2109],"fit_range_runs":[10,12256],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":1.4804,"x_to":2.1453,"x_gain":0.6649,"decades":1.0,"gap_closed_per_decade_local":0.2794,"gap_closed_gain":0.2794},{"runs_from":100,"runs_to":1000,"x_from":2.1453,"x_to":2.5287,"x_gain":0.3834,"decades":1.0,"gap_closed_per_decade_local":0.1611,"gap_closed_gain":0.1611},{"runs_from":1000,"runs_to":10000,"x_from":2.5287,"x_to":2.8355,"x_gain":0.3068,"decades":1.0,"gap_closed_per_decade_local":0.1289,"gap_closed_gain":0.1289},{"runs_from":10000,"runs_to":12256,"x_from":2.8355,"x_to":2.9098,"x_gain":0.0743,"decades":0.088,"gap_closed_per_decade_local":null,"gap_closed_gain":0.0312}]}}},{"key":"muse_prefill_layer","short":"pre","label":"Prefill layer","slot":4,"what":"The decoder layer at prefill, up to 16K new tokens over a 32K context.","ceiling":2.27,"final":{"t":"2026-11-05T09:58:09Z","x":1.920697465982476,"rank":104,"entries":899,"leader":"LeshengJin","leader_x":2.1470638260618298,"gap_pct":-10.54,"pct_ceiling":84.6},"first":{"t":"2026-10-22T23:03:25Z","x":1.3469},"best_ever_scored":1.9207,"best":[["2026-10-22T23:17:51Z",1.3469],["2026-10-23T01:04:24Z",1.4692],["2026-10-23T01:51:58Z",1.5263],["2026-10-23T03:30:28Z",1.5313],["2026-10-23T05:34:58Z",1.5744],["2026-10-23T07:39:59Z",1.6317],["2026-10-23T18:13:14Z",1.679],["2026-10-23T21:36:18Z",1.6808],["2026-10-24T00:51:17Z",1.7159],["2026-10-25T02:16:10Z",1.7288],["2026-10-27T01:43:14Z",1.7313],["2026-10-27T17:14:03Z",1.7373],["2026-10-27T19:14:26Z",1.7989],["2026-10-28T18:11:07Z",1.8022],["2026-10-28T22:58:51Z",1.8081],["2026-10-29T05:29:36Z",1.8325],["2026-10-30T11:59:10Z",1.9207]],"leader":[["2026-10-22T22:43:14Z",1.2588],["2026-10-22T22:54:06Z",1.4366],["2026-10-22T22:56:48Z",1.4411],["2026-10-22T23:17:51Z",1.5301],["2026-10-23T00:47:24Z",1.5983],["2026-10-23T01:17:51Z",1.6287],["2026-10-23T03:30:28Z",1.6576],["2026-10-23T05:02:56Z",1.6954],["2026-10-23T18:13:14Z",1.7258],["2026-10-23T19:27:57Z",1.7448],["2026-10-24T17:57:27Z",1.8573],["2026-10-24T19:53:17Z",1.8665],["2026-10-25T07:46:59Z",1.8878],["2026-10-25T18:58:51Z",1.9192],["2026-10-25T21:27:38Z",1.9631],["2026-10-26T19:59:18Z",2.0252],["2026-10-28T12:02:57Z",2.0302],["2026-10-29T00:06:57Z",2.0325],["2026-10-31T14:26:02Z",2.0333],["2026-10-31T16:23:54Z",2.122],["2026-10-31T17:53:38Z",2.1551],["2026-11-01T00:42:55Z",2.24],["2026-11-01T03:33:43Z",2.2482],["2026-11-01T06:40:20Z",2.2777],["2026-11-01T07:36:48Z",2.2815],["2026-11-02T05:43:29Z",2.3277],["2026-11-02T07:44:40Z",2.0727],["2026-11-02T16:53:28Z",2.0997],["2026-11-03T07:39:03Z",2.1252],["2026-11-03T08:47:41Z",2.2172],["2026-11-03T09:45:52Z",2.2522],["2026-11-04T01:18:51Z",2.0754],["2026-11-04T10:46:57Z",2.1471]],"rank":[["2026-10-22T23:17:51Z",3,26],["2026-10-23T00:47:24Z",7,39],["2026-10-23T01:04:24Z",4,40],["2026-10-23T01:22:02Z",5,41],["2026-10-23T01:42:05Z",6,44],["2026-10-23T01:51:58Z",4,44],["2026-10-23T01:53:04Z",5,45],["2026-10-23T03:30:28Z",4,54],["2026-10-23T03:51:58Z",5,55],["2026-10-23T04:19:31Z",6,60],["2026-10-23T05:34:58Z",5,72],["2026-10-23T07:39:59Z",3,79],["2026-10-23T18:35:24Z",4,159],["2026-10-24T00:16:02Z",5,193],["2026-10-24T00:33:35Z",4,194],["2026-10-24T00:51:17Z",3,198],["2026-10-24T05:46:08Z",4,224],["2026-10-24T06:46:00Z",5,231],["2026-10-24T17:57:27Z",14,299],["2026-10-24T19:53:17Z",15,313],["2026-10-24T20:40:02Z",16,317],["2026-10-24T22:00:23Z",15,324],["2026-10-24T22:54:47Z",16,330],["2026-10-25T01:01:50Z",18,335],["2026-10-25T01:51:30Z",19,337],["2026-10-25T01:57:51Z",20,337],["2026-10-25T02:16:10Z",16,338],["2026-10-25T03:34:03Z",17,340],["2026-10-25T07:16:53Z",18,361],["2026-10-25T18:58:51Z",20,411],["2026-10-25T21:27:38Z",21,419],["2026-10-25T22:10:45Z",22,424],["2026-10-26T19:59:18Z",27,476],["2026-10-26T21:16:57Z",28,482],["2026-10-26T23:07:10Z",29,486],["2026-10-26T23:31:54Z",30,487],["2026-10-27T05:11:07Z",31,498],["2026-10-27T05:19:51Z",32,499],["2026-10-27T06:27:17Z",33,501],["2026-10-27T10:19:42Z",34,510],["2026-10-27T15:39:08Z",35,517],["2026-10-27T16:15:32Z",36,518],["2026-10-27T16:33:54Z",37,519],["2026-10-27T16:46:15Z",38,520],["2026-10-27T17:10:26Z",39,520],["2026-10-27T18:42:27Z",40,522],["2026-10-27T18:57:58Z",41,522],["2026-10-27T19:00:20Z",42,523],["2026-10-27T19:14:26Z",22,524],["2026-10-27T19:24:35Z",23,526],["2026-10-27T20:33:26Z",24,526],["2026-10-27T20:41:37Z",25,527],["2026-10-27T21:03:53Z",26,529],["2026-10-27T21:54:22Z",27,530],["2026-10-27T22:01:41Z",28,531],["2026-10-27T22:20:02Z",29,534],["2026-10-28T01:26:49Z",30,540],["2026-10-28T01:41:11Z",31,540],["2026-10-28T04:16:05Z",32,546],["2026-10-28T04:20:26Z",33,547],["2026-10-28T05:13:12Z",34,551],["2026-10-28T07:31:17Z",35,555],["2026-10-28T07:41:53Z",36,555],["2026-10-28T08:33:35Z",37,557],["2026-10-28T09:04:51Z",38,558],["2026-10-28T10:47:47Z",39,560],["2026-10-28T11:29:50Z",40,561],["2026-10-28T12:39:52Z",41,566],["2026-10-28T13:27:52Z",42,570],["2026-10-28T14:03:22Z",43,572],["2026-10-28T15:47:01Z",44,574],["2026-10-28T18:11:07Z",38,578],["2026-10-28T18:39:16Z",39,580],["2026-10-28T19:17:37Z",40,582],["2026-10-28T20:53:07Z",41,584],["2026-10-28T22:58:51Z",36,586],["2026-10-29T01:21:28Z",37,591],["2026-10-29T04:22:51Z",38,598],["2026-10-29T05:29:36Z",24,599],["2026-10-29T16:06:41Z",28,632],["2026-10-29T17:17:24Z",29,634],["2026-10-29T21:11:56Z",30,648],["2026-10-29T21:37:12Z",31,649],["2026-10-30T01:26:17Z",32,656],["2026-10-30T01:39:18Z",33,656],["2026-10-30T05:20:03Z",34,663],["2026-10-30T06:21:16Z",35,665],["2026-10-30T07:34:19Z",36,669],["2026-10-30T10:11:24Z",37,678],["2026-10-30T11:04:41Z",38,681],["2026-10-30T11:59:10Z",19,684],["2026-10-30T14:21:51Z",20,691],["2026-10-30T16:15:10Z",21,699],["2026-10-30T17:56:31Z",22,706],["2026-10-30T22:22:27Z",23,713],["2026-10-30T23:19:09Z",24,716],["2026-10-30T23:44:08Z",25,717],["2026-10-31T00:03:19Z",26,718],["2026-10-31T01:57:49Z",27,721],["2026-10-31T06:13:11Z",28,731],["2026-10-31T09:35:05Z",29,736],["2026-10-31T12:34:41Z",30,742],["2026-10-31T14:26:02Z",33,745],["2026-10-31T17:59:40Z",34,751],["2026-10-31T18:37:16Z",35,754],["2026-11-01T01:02:58Z",36,763],["2026-11-01T01:17:05Z",37,763],["2026-11-01T06:07:13Z",38,773],["2026-11-01T08:23:43Z",39,775],["2026-11-01T09:07:56Z",40,777],["2026-11-01T09:13:58Z",39,777],["2026-11-01T10:31:21Z",40,781],["2026-11-01T13:03:54Z",41,785],["2026-11-01T13:57:04Z",42,787],["2026-11-01T16:17:09Z",43,790],["2026-11-01T17:44:49Z",44,791],["2026-11-01T18:22:16Z",45,792],["2026-11-01T18:28:16Z",44,792],["2026-11-01T18:41:34Z",43,792],["2026-11-02T02:51:29Z",45,800],["2026-11-02T02:57:20Z",44,800],["2026-11-02T05:28:59Z",45,801],["2026-11-02T05:43:29Z",46,801],["2026-11-02T07:44:40Z",45,803],["2026-11-02T08:24:50Z",41,800],["2026-11-02T09:23:07Z",42,801],["2026-11-02T11:25:35Z",43,802],["2026-11-02T11:39:36Z",44,803],["2026-11-02T12:45:11Z",45,804],["2026-11-02T15:18:11Z",46,806],["2026-11-02T17:40:38Z",47,808],["2026-11-02T19:13:26Z",48,809],["2026-11-02T21:51:49Z",50,815],["2026-11-02T22:06:38Z",51,815],["2026-11-02T23:33:21Z",52,819],["2026-11-02T23:45:24Z",53,820],["2026-11-02T23:51:54Z",54,821],["2026-11-03T00:21:37Z",55,823],["2026-11-03T00:49:08Z",56,822],["2026-11-03T01:09:09Z",57,823],["2026-11-03T02:08:14Z",58,823],["2026-11-03T03:51:24Z",59,827],["2026-11-03T05:11:58Z",60,828],["2026-11-03T05:56:02Z",59,828],["2026-11-03T06:07:49Z",60,829],["2026-11-03T06:20:31Z",61,830],["2026-11-03T06:33:39Z",62,832],["2026-11-03T06:40:08Z",63,831],["2026-11-03T07:33:08Z",64,832],["2026-11-03T08:13:28Z",65,834],["2026-11-03T08:27:55Z",66,834],["2026-11-03T10:13:38Z",67,835],["2026-11-03T10:20:59Z",68,836],["2026-11-03T11:36:36Z",69,839],["2026-11-03T15:24:25Z",70,848],["2026-11-03T17:19:26Z",71,849],["2026-11-03T18:46:23Z",72,853],["2026-11-03T20:46:09Z",73,856],["2026-11-03T20:53:46Z",74,856],["2026-11-03T21:44:34Z",73,848],["2026-11-04T00:51:07Z",74,852],["2026-11-04T00:57:35Z",73,851],["2026-11-04T01:06:38Z",71,851],["2026-11-04T01:18:51Z",70,851],["2026-11-04T01:33:22Z",71,852],["2026-11-04T01:47:01Z",70,852],["2026-11-04T02:09:28Z",71,853],["2026-11-04T04:36:19Z",72,856],["2026-11-04T05:10:02Z",73,858],["2026-11-04T07:33:42Z",74,861],["2026-11-04T08:21:37Z",75,864],["2026-11-04T09:40:48Z",76,866],["2026-11-04T10:46:57Z",78,869],["2026-11-04T11:54:05Z",79,872],["2026-11-04T12:34:13Z",80,872],["2026-11-04T13:08:20Z",81,873],["2026-11-04T13:34:35Z",82,874],["2026-11-04T13:47:38Z",83,876],["2026-11-04T13:53:37Z",85,876],["2026-11-04T14:12:54Z",86,876],["2026-11-04T14:44:41Z",87,876],["2026-11-04T15:18:57Z",88,877],["2026-11-04T15:30:53Z",90,877],["2026-11-04T17:58:58Z",91,879],["2026-11-04T18:40:18Z",92,879],["2026-11-04T19:29:02Z",93,879],["2026-11-04T19:36:02Z",94,879],["2026-11-04T20:19:52Z",95,880],["2026-11-04T20:34:12Z",96,882],["2026-11-04T21:14:19Z",97,882],["2026-11-04T21:45:08Z",98,882],["2026-11-05T01:24:06Z",99,886],["2026-11-05T02:36:29Z",100,887],["2026-11-05T03:03:38Z",101,887],["2026-11-05T04:44:54Z",102,889],["2026-11-05T07:35:33Z",103,899],["2026-11-05T07:55:05Z",104,900]],"dq":null,"judge_over_clone":1.0146,"fits":{"internal":{"x_per_decade":0.2367,"x_per_decade_ci90":[0.2189,0.2557],"intercept_x_at_1_run":1.1948,"gap_closed_per_decade":0.1891,"gap_closed_per_decade_ci90":[0.1749,0.2043],"fit_range_runs":[12,4080],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":null,"x_to":1.6741,"x_gain":null,"decades":1.0,"gap_closed_per_decade_local":null,"gap_closed_gain":null},{"runs_from":100,"runs_to":1000,"x_from":1.6741,"x_to":1.8825,"x_gain":0.2084,"decades":1.0,"gap_closed_per_decade_local":0.1665,"gap_closed_gain":0.1665},{"runs_from":1000,"runs_to":4080,"x_from":1.8825,"x_to":2.1419,"x_gain":0.2594,"decades":0.611,"gap_closed_per_decade_local":0.3394,"gap_closed_gain":0.2072}]},"scored":{"x_per_decade":0.1346,"x_per_decade_ci90":[0.1189,0.178],"intercept_x_at_1_run":1.3752,"gap_closed_per_decade":0.106,"gap_closed_per_decade_ci90":[0.0936,0.1402],"fit_range_runs":[10,4080],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":1.4692,"x_to":1.6808,"x_gain":0.2116,"decades":1.0,"gap_closed_per_decade_local":0.1666,"gap_closed_gain":0.1666},{"runs_from":100,"runs_to":1000,"x_from":1.6808,"x_to":1.7313,"x_gain":0.0505,"decades":1.0,"gap_closed_per_decade_local":0.0397,"gap_closed_gain":0.0397},{"runs_from":1000,"runs_to":4080,"x_from":1.7313,"x_to":1.9207,"x_gain":0.1894,"decades":0.611,"gap_closed_per_decade_local":0.2442,"gap_closed_gain":0.1491}]}}},{"key":"muse_bf16_proj","short":"bf16","label":"BF16 projections","slot":5,"what":"The dense BF16 projections around the MoE at small batch: qkv, out, router, pooler.","ceiling":1.99,"final":{"t":"2026-11-05T16:46:28Z","x":1.5375097993769167,"rank":313,"entries":1747,"leader":"gogowhy","leader_x":1.7276592188096174,"gap_pct":-11.01,"pct_ceiling":77.3},"first":{"t":"2026-10-22T22:20:14Z","x":1.1769},"best_ever_scored":1.5712,"best":[["2026-10-22T22:43:14Z",1.1769],["2026-10-23T02:15:07Z",1.1886],["2026-10-23T03:30:28Z",1.2929],["2026-10-23T03:48:58Z",1.3203],["2026-10-23T09:07:36Z",1.3941],["2026-10-23T22:28:51Z",1.4107],["2026-10-24T01:59:34Z",1.4245],["2026-10-24T03:01:39Z",1.4246],["2026-10-24T22:05:34Z",1.4439],["2026-10-27T05:15:24Z",1.4663],["2026-10-27T19:56:57Z",1.5004],["2026-10-27T21:24:33Z",1.5375],["2026-10-27T21:45:10Z",1.5543],["2026-10-30T08:47:35Z",1.5621],["2026-10-30T22:54:20Z",1.5712],["2026-11-03T21:29:02Z",1.5621],["2026-11-03T21:44:34Z",1.5375]],"leader":[["2026-10-22T22:43:14Z",1.2594],["2026-10-22T23:36:58Z",1.3014],["2026-10-22T23:59:27Z",1.3336],["2026-10-23T00:47:24Z",1.3809],["2026-10-23T01:20:31Z",1.3824],["2026-10-23T01:34:12Z",1.3903],["2026-10-23T01:55:57Z",1.4339],["2026-10-23T03:30:28Z",1.4466],["2026-10-24T03:22:33Z",1.4543],["2026-10-24T03:59:08Z",1.4601],["2026-10-24T04:12:00Z",1.501],["2026-10-24T05:46:08Z",1.5084],["2026-10-24T17:57:27Z",1.5916],["2026-10-24T19:57:41Z",1.6099],["2026-10-25T23:11:41Z",1.6238],["2026-10-26T19:59:18Z",1.6367],["2026-10-29T21:11:56Z",1.6498],["2026-10-31T00:03:19Z",1.6523],["2026-11-02T07:31:36Z",1.6596],["2026-11-02T17:33:06Z",1.6636],["2026-11-03T02:01:37Z",1.6665],["2026-11-03T03:18:41Z",1.6796],["2026-11-03T07:27:03Z",1.6797],["2026-11-03T08:40:45Z",1.694],["2026-11-04T11:22:42Z",1.7091],["2026-11-04T11:48:09Z",1.7102],["2026-11-04T11:54:05Z",1.7111],["2026-11-04T17:31:22Z",1.7175],["2026-11-05T03:29:26Z",1.7277]],"rank":[["2026-10-22T22:43:14Z",36,59],["2026-10-22T22:54:06Z",37,60],["2026-10-22T23:17:51Z",41,67],["2026-10-22T23:24:18Z",42,71],["2026-10-22T23:25:48Z",43,72],["2026-10-22T23:29:40Z",44,71],["2026-10-22T23:34:58Z",45,72],["2026-10-22T23:38:58Z",46,73],["2026-10-22T23:45:50Z",47,76],["2026-10-22T23:59:27Z",49,82],["2026-10-23T00:04:21Z",51,84],["2026-10-23T00:47:24Z",53,91],["2026-10-23T01:01:49Z",54,93],["2026-10-23T01:11:24Z",55,94],["2026-10-23T01:13:53Z",57,96],["2026-10-23T01:42:05Z",58,101],["2026-10-23T01:47:19Z",59,105],["2026-10-23T01:51:58Z",60,107],["2026-10-23T02:15:07Z",55,113],["2026-10-23T02:19:41Z",56,115],["2026-10-23T03:30:28Z",8,139],["2026-10-23T03:48:58Z",4,147],["2026-10-23T04:36:01Z",5,166],["2026-10-23T06:46:26Z",6,225],["2026-10-23T09:00:36Z",3,273],["2026-10-23T09:07:36Z",2,276],["2026-10-23T18:13:14Z",3,459],["2026-10-23T19:58:44Z",4,486],["2026-10-23T20:08:51Z",5,491],["2026-10-23T21:36:18Z",6,525],["2026-10-23T22:28:51Z",3,544],["2026-10-23T22:53:33Z",4,548],["2026-10-23T23:43:42Z",5,566],["2026-10-24T00:45:00Z",6,582],["2026-10-24T01:59:34Z",5,598],["2026-10-24T02:18:03Z",6,601],["2026-10-24T03:01:39Z",5,608],["2026-10-24T03:22:33Z",6,612],["2026-10-24T03:59:08Z",7,617],["2026-10-24T04:19:09Z",8,618],["2026-10-24T17:57:27Z",17,771],["2026-10-24T19:23:06Z",18,787],["2026-10-24T20:00:36Z",19,797],["2026-10-24T20:02:34Z",20,797],["2026-10-24T20:40:02Z",22,802],["2026-10-24T21:19:17Z",23,813],["2026-10-24T21:35:35Z",24,817],["2026-10-24T21:55:52Z",26,820],["2026-10-24T22:05:34Z",19,824],["2026-10-24T22:26:05Z",20,832],["2026-10-24T22:54:47Z",21,839],["2026-10-24T23:06:29Z",22,842],["2026-10-24T23:30:27Z",23,846],["2026-10-24T23:38:09Z",24,848],["2026-10-24T23:45:20Z",25,850],["2026-10-25T00:13:14Z",26,855],["2026-10-25T01:01:50Z",27,862],["2026-10-25T01:47:01Z",28,865],["2026-10-25T02:17:36Z",29,868],["2026-10-25T02:30:17Z",31,869],["2026-10-25T03:15:47Z",32,877],["2026-10-25T03:27:17Z",33,878],["2026-10-25T04:55:53Z",34,892],["2026-10-25T05:01:51Z",35,892],["2026-10-25T05:46:39Z",38,898],["2026-10-25T05:56:29Z",39,898],["2026-10-25T06:18:03Z",40,901],["2026-10-25T06:28:34Z",41,901],["2026-10-25T06:33:13Z",42,902],["2026-10-25T06:34:36Z",43,902],["2026-10-25T06:53:18Z",44,903],["2026-10-25T07:16:53Z",46,906],["2026-10-25T08:03:37Z",47,913],["2026-10-25T08:08:24Z",48,913],["2026-10-25T08:43:06Z",49,916],["2026-10-25T18:58:51Z",73,980],["2026-10-25T19:45:37Z",75,988],["2026-10-25T20:04:58Z",76,993],["2026-10-25T20:34:38Z",77,996],["2026-10-25T21:15:19Z",79,1001],["2026-10-25T21:26:16Z",80,1003],["2026-10-25T22:10:45Z",81,1007],["2026-10-25T22:27:40Z",82,1007],["2026-10-25T23:13:49Z",83,1009],["2026-10-25T23:33:39Z",84,1009],["2026-10-26T19:59:18Z",140,1091],["2026-10-26T20:36:42Z",141,1093],["2026-10-26T21:02:39Z",142,1094],["2026-10-26T21:08:36Z",143,1094],["2026-10-26T21:26:49Z",144,1096],["2026-10-26T22:28:44Z",145,1097],["2026-10-26T22:40:11Z",146,1098],["2026-10-26T22:56:28Z",147,1099],["2026-10-26T23:04:05Z",148,1100],["2026-10-26T23:08:40Z",149,1100],["2026-10-26T23:16:29Z",150,1101],["2026-10-26T23:19:41Z",151,1101],["2026-10-26T23:44:04Z",152,1103],["2026-10-27T01:26:50Z",153,1105],["2026-10-27T02:00:01Z",154,1107],["2026-10-27T02:09:10Z",155,1107],["2026-10-27T02:30:31Z",156,1107],["2026-10-27T02:35:43Z",157,1107],["2026-10-27T03:05:02Z",158,1110],["2026-10-27T03:40:16Z",159,1111],["2026-10-27T04:41:36Z",160,1115],["2026-10-27T04:49:24Z",161,1117],["2026-10-27T04:53:54Z",163,1118],["2026-10-27T05:11:07Z",164,1118],["2026-10-27T05:15:24Z",150,1118],["2026-10-27T05:28:34Z",151,1119],["2026-10-27T06:04:28Z",152,1120],["2026-10-27T08:02:32Z",153,1130],["2026-10-27T08:06:15Z",154,1130],["2026-10-27T08:57:05Z",155,1133],["2026-10-27T09:36:55Z",156,1135],["2026-10-27T09:41:27Z",157,1135],["2026-10-27T11:02:15Z",158,1138],["2026-10-27T11:42:21Z",159,1144],["2026-10-27T11:50:29Z",160,1145],["2026-10-27T12:08:12Z",161,1148],["2026-10-27T12:39:13Z",162,1149],["2026-10-27T14:06:10Z",163,1151],["2026-10-27T14:22:22Z",164,1152],["2026-10-27T15:25:00Z",165,1153],["2026-10-27T16:21:00Z",166,1155],["2026-10-27T17:25:39Z",167,1162],["2026-10-27T18:20:01Z",168,1166],["2026-10-27T19:24:35Z",169,1170],["2026-10-27T19:26:42Z",170,1170],["2026-10-27T19:30:43Z",171,1170],["2026-10-27T19:56:57Z",147,1171],["2026-10-27T20:31:32Z",148,1173],["2026-10-27T21:16:04Z",149,1174],["2026-10-27T21:24:33Z",46,1174],["2026-10-27T21:41:40Z",47,1175],["2026-10-27T21:45:10Z",26,1175],["2026-10-27T22:01:41Z",27,1177],["2026-10-27T22:13:55Z",28,1182],["2026-10-27T22:44:23Z",29,1185],["2026-10-28T00:32:53Z",28,1188],["2026-10-28T01:39:16Z",29,1193],["2026-10-28T01:58:37Z",30,1193],["2026-10-28T03:02:39Z",31,1197],["2026-10-28T03:38:58Z",32,1201],["2026-10-28T06:07:17Z",33,1213],["2026-10-28T06:48:19Z",35,1215],["2026-10-28T08:41:12Z",36,1220],["2026-10-28T10:51:26Z",37,1225],["2026-10-28T15:16:39Z",38,1242],["2026-10-28T15:39:45Z",39,1242],["2026-10-28T15:56:42Z",40,1244],["2026-10-28T16:21:08Z",41,1245],["2026-10-28T16:31:37Z",42,1245],["2026-10-28T16:39:52Z",43,1245],["2026-10-28T18:08:57Z",44,1250],["2026-10-28T18:58:34Z",45,1253],["2026-10-28T19:00:28Z",46,1253],["2026-10-28T20:38:03Z",45,1257],["2026-10-28T20:43:55Z",46,1258],["2026-10-28T21:39:01Z",47,1260],["2026-10-28T23:26:41Z",48,1262],["2026-10-28T23:42:36Z",49,1262],["2026-10-29T00:49:19Z",50,1264],["2026-10-29T02:08:31Z",51,1267],["2026-10-29T03:16:00Z",52,1268],["2026-10-29T04:53:01Z",53,1272],["2026-10-29T16:06:41Z",64,1359],["2026-10-29T16:18:22Z",65,1360],["2026-10-29T17:04:32Z",66,1366],["2026-10-29T21:55:22Z",67,1407],["2026-10-29T22:10:56Z",68,1407],["2026-10-29T23:23:55Z",69,1412],["2026-10-30T00:13:04Z",70,1412],["2026-10-30T01:45:32Z",71,1418],["2026-10-30T01:51:42Z",72,1419],["2026-10-30T02:11:39Z",73,1420],["2026-10-30T03:53:44Z",74,1429],["2026-10-30T04:13:13Z",75,1431],["2026-10-30T04:34:17Z",76,1434],["2026-10-30T05:20:03Z",77,1439],["2026-10-30T05:39:19Z",78,1439],["2026-10-30T06:35:38Z",79,1442],["2026-10-30T08:40:35Z",81,1451],["2026-10-30T08:47:35Z",55,1451],["2026-10-30T09:00:34Z",56,1451],["2026-10-30T10:18:07Z",57,1456],["2026-10-30T10:44:56Z",58,1459],["2026-10-30T12:45:30Z",59,1466],["2026-10-30T13:47:19Z",60,1469],["2026-10-30T14:21:51Z",61,1471],["2026-10-30T14:51:20Z",62,1473],["2026-10-30T15:10:58Z",66,1474],["2026-10-30T16:01:47Z",67,1479],["2026-10-30T16:48:53Z",68,1483],["2026-10-30T16:54:50Z",69,1484],["2026-10-30T17:07:31Z",70,1484],["2026-10-30T17:19:51Z",71,1485],["2026-10-30T17:26:22Z",72,1485],["2026-10-30T18:44:28Z",73,1491],["2026-10-30T18:56:59Z",74,1492],["2026-10-30T19:09:35Z",73,1491],["2026-10-30T19:15:30Z",74,1491],["2026-10-30T19:53:29Z",76,1493],["2026-10-30T20:26:53Z",77,1494],["2026-10-30T22:29:27Z",78,1498],["2026-10-30T22:47:55Z",79,1500],["2026-10-30T22:54:20Z",49,1500],["2026-10-30T23:07:04Z",50,1500],["2026-10-30T23:19:09Z",51,1501],["2026-10-31T02:33:06Z",52,1507],["2026-10-31T03:42:12Z",53,1510],["2026-10-31T06:13:11Z",55,1518],["2026-10-31T14:26:02Z",58,1534],["2026-10-31T16:23:54Z",59,1538],["2026-10-31T16:31:35Z",60,1538],["2026-10-31T16:58:34Z",61,1538],["2026-10-31T17:53:38Z",62,1543],["2026-10-31T18:18:15Z",63,1544],["2026-10-31T18:24:12Z",64,1544],["2026-10-31T19:46:11Z",65,1547],["2026-10-31T22:14:56Z",66,1552],["2026-10-31T22:21:02Z",67,1552],["2026-10-31T23:18:10Z",68,1555],["2026-10-31T23:38:12Z",69,1558],["2026-11-01T00:09:48Z",70,1558],["2026-11-01T01:02:58Z",69,1560],["2026-11-01T03:14:37Z",70,1564],["2026-11-01T03:20:37Z",71,1564],["2026-11-01T08:03:37Z",72,1584],["2026-11-01T08:30:52Z",73,1586],["2026-11-01T09:13:58Z",74,1589],["2026-11-01T09:20:55Z",75,1589],["2026-11-01T09:28:12Z",76,1589],["2026-11-01T09:40:59Z",77,1590],["2026-11-01T10:00:05Z",78,1591],["2026-11-01T10:12:18Z",79,1593],["2026-11-01T10:25:16Z",80,1593],["2026-11-01T10:44:23Z",81,1595],["2026-11-01T10:51:22Z",82,1595],["2026-11-01T11:18:34Z",84,1595],["2026-11-01T11:31:32Z",85,1596],["2026-11-01T11:50:36Z",86,1596],["2026-11-01T12:17:42Z",87,1597],["2026-11-01T12:23:40Z",89,1597],["2026-11-01T12:30:41Z",90,1597],["2026-11-01T12:43:47Z",91,1599],["2026-11-01T12:56:48Z",92,1599],["2026-11-01T13:57:04Z",93,1599],["2026-11-01T14:57:18Z",94,1599],["2026-11-01T15:03:18Z",95,1601],["2026-11-01T15:43:27Z",96,1604],["2026-11-01T16:03:52Z",97,1604],["2026-11-01T16:17:09Z",98,1605],["2026-11-01T16:31:18Z",99,1606],["2026-11-01T18:09:32Z",100,1607],["2026-11-01T18:16:17Z",101,1607],["2026-11-01T18:22:16Z",102,1607],["2026-11-01T18:28:16Z",101,1607],["2026-11-01T18:41:34Z",102,1607],["2026-11-01T18:54:58Z",103,1607],["2026-11-01T19:02:24Z",104,1607],["2026-11-01T19:15:09Z",103,1608],["2026-11-01T19:43:03Z",104,1609],["2026-11-01T19:57:37Z",105,1609],["2026-11-01T23:08:04Z",106,1609],["2026-11-01T23:27:14Z",107,1610],["2026-11-02T00:13:40Z",108,1613],["2026-11-02T00:31:48Z",109,1614],["2026-11-02T01:37:56Z",110,1615],["2026-11-02T01:52:08Z",111,1617],["2026-11-02T02:05:12Z",112,1617],["2026-11-02T02:51:29Z",113,1619],["2026-11-02T03:37:55Z",112,1620],["2026-11-02T04:24:02Z",113,1621],["2026-11-02T05:22:05Z",114,1623],["2026-11-02T05:43:29Z",115,1623],["2026-11-02T05:49:08Z",116,1623],["2026-11-02T06:03:20Z",117,1623],["2026-11-02T07:11:28Z",118,1625],["2026-11-02T07:44:40Z",119,1625],["2026-11-02T07:51:42Z",120,1625],["2026-11-02T08:10:41Z",121,1628],["2026-11-02T08:50:53Z",122,1630],["2026-11-02T09:16:54Z",123,1632],["2026-11-02T09:29:05Z",121,1629],["2026-11-02T09:48:05Z",122,1631],["2026-11-02T09:53:46Z",123,1631],["2026-11-02T11:19:35Z",124,1631],["2026-11-02T12:51:21Z",126,1634],["2026-11-02T13:17:49Z",127,1634],["2026-11-02T13:30:49Z",128,1636],["2026-11-02T14:16:33Z",131,1638],["2026-11-02T14:30:31Z",132,1638],["2026-11-02T15:04:13Z",133,1638],["2026-11-02T15:44:20Z",134,1638],["2026-11-02T16:53:28Z",136,1643],["2026-11-02T17:26:42Z",137,1645],["2026-11-02T17:33:06Z",139,1645],["2026-11-02T18:01:17Z",140,1646],["2026-11-02T18:15:11Z",139,1646],["2026-11-02T18:31:11Z",140,1646],["2026-11-02T18:37:34Z",141,1646],["2026-11-02T19:05:12Z",142,1647],["2026-11-02T19:20:34Z",143,1649],["2026-11-02T19:27:45Z",145,1651],["2026-11-02T19:55:50Z",146,1652],["2026-11-02T20:10:10Z",147,1652],["2026-11-02T20:32:39Z",149,1652],["2026-11-02T20:47:21Z",150,1653],["2026-11-02T21:51:49Z",151,1653],["2026-11-02T21:59:23Z",152,1653],["2026-11-02T22:22:11Z",154,1654],["2026-11-02T23:01:22Z",155,1656],["2026-11-02T23:21:04Z",154,1658],["2026-11-02T23:26:43Z",153,1658],["2026-11-02T23:33:21Z",154,1658],["2026-11-02T23:45:24Z",155,1660],["2026-11-03T00:21:37Z",156,1662],["2026-11-03T01:54:10Z",155,1663],["2026-11-03T03:25:25Z",156,1667],["2026-11-03T03:43:54Z",157,1667],["2026-11-03T03:58:39Z",158,1667],["2026-11-03T04:30:56Z",159,1668],["2026-11-03T06:01:46Z",160,1672],["2026-11-03T08:47:41Z",161,1678],["2026-11-03T10:48:59Z",162,1683],["2026-11-03T12:10:44Z",163,1690],["2026-11-03T12:31:28Z",164,1690],["2026-11-03T13:39:26Z",165,1692],["2026-11-03T13:46:27Z",166,1692],["2026-11-03T14:19:26Z",167,1693],["2026-11-03T14:51:44Z",168,1694],["2026-11-03T18:09:33Z",169,1702],["2026-11-03T18:17:14Z",170,1703],["2026-11-03T18:32:08Z",171,1704],["2026-11-03T19:33:06Z",172,1710],["2026-11-03T19:55:08Z",173,1710],["2026-11-03T20:03:40Z",174,1711],["2026-11-03T20:30:50Z",173,1714],["2026-11-03T20:39:13Z",175,1716],["2026-11-03T21:01:14Z",176,1717],["2026-11-03T21:07:33Z",177,1717],["2026-11-03T21:20:35Z",176,1718],["2026-11-03T21:29:02Z",216,1710],["2026-11-03T21:36:38Z",200,1700],["2026-11-03T21:44:34Z",224,1675],["2026-11-03T22:21:03Z",226,1677],["2026-11-03T22:27:09Z",227,1678],["2026-11-03T22:53:09Z",228,1680],["2026-11-03T23:00:48Z",229,1680],["2026-11-03T23:07:21Z",230,1680],["2026-11-03T23:20:48Z",231,1681],["2026-11-03T23:43:58Z",232,1682],["2026-11-04T00:13:00Z",233,1684],["2026-11-04T00:21:45Z",234,1684],["2026-11-04T00:35:56Z",235,1684],["2026-11-04T00:43:38Z",233,1683],["2026-11-04T01:12:40Z",234,1683],["2026-11-04T01:18:51Z",235,1683],["2026-11-04T01:25:22Z",236,1683],["2026-11-04T01:54:25Z",238,1684],["2026-11-04T02:16:42Z",239,1685],["2026-11-04T02:23:24Z",240,1687],["2026-11-04T02:39:07Z",241,1688],["2026-11-04T02:46:29Z",242,1689],["2026-11-04T03:51:38Z",243,1689],["2026-11-04T03:57:42Z",244,1690],["2026-11-04T04:05:01Z",245,1690],["2026-11-04T04:28:20Z",246,1691],["2026-11-04T04:36:19Z",247,1692],["2026-11-04T05:02:46Z",248,1692],["2026-11-04T05:23:24Z",249,1693],["2026-11-04T05:37:14Z",250,1694],["2026-11-04T06:07:00Z",252,1695],["2026-11-04T08:07:20Z",253,1700],["2026-11-04T08:21:37Z",255,1699],["2026-11-04T08:34:53Z",256,1700],["2026-11-04T08:54:47Z",258,1701],["2026-11-04T09:07:11Z",259,1701],["2026-11-04T09:21:03Z",260,1701],["2026-11-04T09:47:39Z",261,1701],["2026-11-04T10:00:51Z",262,1701],["2026-11-04T10:33:51Z",263,1702],["2026-11-04T11:29:19Z",265,1702],["2026-11-04T11:41:07Z",267,1704],["2026-11-04T12:41:17Z",268,1705],["2026-11-04T12:47:20Z",269,1705],["2026-11-04T13:15:29Z",270,1706],["2026-11-04T13:21:30Z",271,1706],["2026-11-04T13:34:35Z",272,1708],["2026-11-04T13:41:30Z",273,1709],["2026-11-04T14:05:37Z",274,1709],["2026-11-04T14:18:59Z",275,1709],["2026-11-04T14:44:41Z",276,1711],["2026-11-04T15:45:29Z",277,1713],["2026-11-04T16:33:37Z",278,1714],["2026-11-04T17:18:28Z",279,1716],["2026-11-04T17:31:22Z",280,1718],["2026-11-04T17:58:58Z",282,1721],["2026-11-04T18:25:22Z",283,1721],["2026-11-04T18:53:38Z",284,1721],["2026-11-04T19:00:27Z",285,1721],["2026-11-04T19:07:24Z",286,1721],["2026-11-04T19:13:48Z",287,1721],["2026-11-04T19:53:08Z",288,1721],["2026-11-04T20:11:22Z",289,1721],["2026-11-04T20:27:14Z",290,1722],["2026-11-04T21:20:24Z",291,1723],["2026-11-04T22:31:12Z",292,1723],["2026-11-04T22:44:07Z",293,1723],["2026-11-04T23:25:38Z",294,1726],["2026-11-04T23:32:22Z",295,1726],["2026-11-04T23:38:21Z",296,1726],["2026-11-05T00:04:29Z",297,1726],["2026-11-05T00:59:59Z",298,1728],["2026-11-05T01:07:07Z",299,1728],["2026-11-05T02:30:22Z",300,1731],["2026-11-05T02:36:29Z",301,1731],["2026-11-05T02:49:20Z",303,1732],["2026-11-05T03:29:26Z",304,1735],["2026-11-05T04:11:39Z",305,1737],["2026-11-05T04:50:45Z",306,1740],["2026-11-05T05:11:54Z",307,1740],["2026-11-05T05:31:34Z",308,1742],["2026-11-05T06:58:12Z",309,1743],["2026-11-05T07:04:25Z",310,1745],["2026-11-05T07:11:32Z",311,1746],["2026-11-05T07:17:25Z",312,1746],["2026-11-05T07:55:05Z",313,1747]],"dq":{"row_before":1.5712,"rank_before":176,"row_after":1.5375,"rank_after":224,"x_change_pct":-2.14,"t_before":"2026-11-03T21:20:35Z","t_after":"2026-11-03T21:44:34Z"},"judge_over_clone":1.0104,"fits":{"internal":{"x_per_decade":0.1559,"x_per_decade_ci90":[0.1396,0.1607],"intercept_x_at_1_run":1.0144,"gap_closed_per_decade":0.1591,"gap_closed_per_decade_ci90":[0.1424,0.164],"fit_range_runs":[35,6026],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":null,"x_to":1.3297,"x_gain":null,"decades":1.0,"gap_closed_per_decade_local":null,"gap_closed_gain":null},{"runs_from":100,"runs_to":1000,"x_from":1.3297,"x_to":1.4835,"x_gain":0.1537,"decades":1.0,"gap_closed_per_decade_local":0.1569,"gap_closed_gain":0.1569},{"runs_from":1000,"runs_to":6026,"x_from":1.4835,"x_to":1.6408,"x_gain":0.1573,"decades":0.78,"gap_closed_per_decade_local":0.2058,"gap_closed_gain":0.1605}]},"scored":{"x_per_decade":0.1605,"x_per_decade_ci90":[0.1237,0.1641],"intercept_x_at_1_run":0.9714,"gap_closed_per_decade":0.1621,"gap_closed_per_decade_ci90":[0.1249,0.1657],"fit_range_runs":[10,6026],"grid_points":30,"bootstrap_reps":200,"by_decade":[{"runs_from":10,"runs_to":100,"x_from":1.1769,"x_to":1.2929,"x_gain":0.116,"decades":1.0,"gap_closed_per_decade_local":0.1172,"gap_closed_gain":0.1172},{"runs_from":100,"runs_to":1000,"x_from":1.2929,"x_to":1.4663,"x_gain":0.1734,"decades":1.0,"gap_closed_per_decade_local":0.1751,"gap_closed_gain":0.1751},{"runs_from":1000,"runs_to":6026,"x_from":1.4663,"x_to":1.5712,"x_gain":0.1049,"decades":0.78,"gap_closed_per_decade_local":0.1359,"gap_closed_gain":0.106}]}}}],"scaling":{"series":[{"key":"muse_decode_layer","points":[[24,1.5373,0.2379,"2026-10-23T01:37:12Z"],[29,1.5756,0.2562,"2026-10-23T01:52:11Z"],[35,1.5912,0.2636,"2026-10-23T02:07:19Z"],[39,1.6118,0.2734,"2026-10-23T02:13:37Z"],[45,1.6454,0.2895,"2026-10-23T02:25:05Z"],[47,1.6864,0.309,"2026-10-23T02:31:01Z"],[61,1.7083,0.3194,"2026-10-23T02:51:44Z"],[77,1.733,0.3312,"2026-10-23T03:34:17Z"],[82,1.7942,0.3604,"2026-10-23T04:06:12Z"],[90,1.7946,0.3606,"2026-10-23T04:16:48Z"],[92,1.8284,0.3767,"2026-10-23T04:20:34Z"],[93,1.8306,0.3778,"2026-10-23T04:24:25Z"],[101,1.8618,0.3926,"2026-10-23T05:14:06Z"],[116,1.8628,0.3931,"2026-10-23T05:36:27Z"],[119,1.9206,0.4207,"2026-10-23T05:42:07Z"],[126,1.9448,0.4322,"2026-10-23T06:11:41Z"],[136,1.9635,0.4411,"2026-10-23T06:40:48Z"],[153,1.9638,0.4412,"2026-10-23T07:42:13Z"],[155,1.9683,0.4434,"2026-10-23T07:54:43Z"],[156,1.9698,0.4441,"2026-10-23T07:59:41Z"],[157,1.9741,0.4462,"2026-10-23T08:00:37Z"],[160,1.9912,0.4543,"2026-10-23T08:15:14Z"],[168,1.9929,0.4551,"2026-10-23T09:17:54Z"],[169,1.9962,0.4567,"2026-10-23T09:22:22Z"],[172,1.9996,0.4583,"2026-10-23T09:31:47Z"],[173,2.0267,0.4713,"2026-10-23T09:38:31Z"],[185,2.0329,0.4742,"2026-10-23T10:10:54Z"],[203,2.0517,0.4832,"2026-10-23T10:38:27Z"],[207,2.0538,0.4842,"2026-10-23T10:43:54Z"],[233,2.0553,0.4849,"2026-10-23T11:28:58Z"],[241,2.06,0.4871,"2026-10-23T11:49:42Z"],[242,2.0795,0.4964,"2026-10-23T11:50:27Z"],[249,2.0845,0.4988,"2026-10-23T12:10:16Z"],[255,2.0929,0.5028,"2026-10-23T12:17:51Z"],[266,2.1175,0.5146,"2026-10-23T13:57:47Z"],[287,2.1227,0.517,"2026-10-23T15:00:25Z"],[293,2.1253,0.5183,"2026-10-23T15:16:28Z"],[295,2.1367,0.5237,"2026-10-23T15:20:39Z"],[301,2.1441,0.5272,"2026-10-23T15:50:54Z"],[309,2.1526,0.5313,"2026-10-23T16:43:50Z"],[316,2.155,0.5324,"2026-10-23T17:43:46Z"],[317,2.158,0.5338,"2026-10-23T17:49:23Z"],[330,2.1583,0.534,"2026-10-23T18:09:09Z"],[355,2.1604,0.535,"2026-10-23T18:45:48Z"],[374,2.1624,0.5359,"2026-10-23T19:05:02Z"],[433,2.1639,0.5367,"2026-10-23T22:17:08Z"],[440,2.1646,0.537,"2026-10-23T22:28:07Z"],[444,2.1651,0.5372,"2026-10-23T22:33:18Z"],[452,2.1664,0.5379,"2026-10-23T22:42:27Z"],[470,2.176,0.5424,"2026-10-23T23:14:19Z"],[471,2.1798,0.5443,"2026-10-23T23:15:46Z"],[473,2.1883,0.5483,"2026-10-23T23:20:25Z"],[475,2.1943,0.5512,"2026-10-23T23:23:08Z"],[484,2.1957,0.5518,"2026-10-23T23:31:56Z"],[525,2.197,0.5524,"2026-10-24T01:23:47Z"],[531,2.212,0.5596,"2026-10-24T01:39:29Z"],[532,2.2121,0.5596,"2026-10-24T01:53:50Z"],[533,2.2192,0.563,"2026-10-24T01:56:24Z"],[598,2.2212,0.564,"2026-10-24T20:41:50Z"],[661,2.241,0.5734,"2026-10-25T00:20:48Z"],[686,2.2487,0.5771,"2026-10-25T01:34:44Z"],[693,2.2568,0.5809,"2026-10-25T01:51:00Z"],[716,2.2769,0.5905,"2026-10-25T02:57:20Z"],[760,2.278,0.5911,"2026-10-25T05:03:33Z"],[768,2.2793,0.5917,"2026-10-25T05:37:54Z"],[778,2.2807,0.5924,"2026-10-25T05:52:20Z"],[800,2.2819,0.5929,"2026-10-25T06:25:37Z"],[820,2.2891,0.5964,"2026-10-25T06:50:18Z"],[839,2.3034,0.6031,"2026-10-25T07:37:49Z"],[998,2.3047,0.6038,"2026-10-26T23:53:01Z"],[1078,2.3391,0.6202,"2026-10-27T00:47:30Z"],[2513,2.3742,0.6369,"2026-10-27T14:51:38Z"],[2663,2.3772,0.6383,"2026-10-27T15:42:06Z"],[2966,2.3834,0.6413,"2026-10-27T17:27:48Z"],[4402,2.3916,0.6452,"2026-10-28T03:09:09Z"],[5850,2.3957,0.6472,"2026-10-28T11:23:02Z"],[6157,2.4245,0.6609,"2026-10-28T13:15:39Z"],[16858,2.4254,0.6614,"2026-10-30T08:25:51Z"],[19199,2.4294,0.6632,"2026-10-30T23:07:26Z"],[19242,2.4306,0.6638,"2026-10-30T23:17:02Z"],[19255,2.4307,0.6639,"2026-10-30T23:20:38Z"],[21363,2.4332,0.6651,"2026-11-01T01:39:17Z"],[22032,2.4368,0.6668,"2026-11-01T04:30:36Z"],[22113,2.4381,0.6674,"2026-11-01T04:59:14Z"],[22365,2.4614,0.6785,"2026-11-01T05:46:17Z"],[23612,2.4766,0.6858,"2026-11-01T11:29:37Z"],[25404,2.4826,0.6886,"2026-11-02T03:23:05Z"],[25509,2.4839,0.6892,"2026-11-02T06:52:42Z"],[25641,2.4876,0.691,"2026-11-02T10:53:11Z"],[25646,2.4898,0.692,"2026-11-02T10:57:34Z"],[26318,2.5081,0.7008,"2026-11-03T18:26:08Z"],[26331,2.5102,0.7018,"2026-11-03T18:41:09Z"],[26344,2.5223,0.7075,"2026-11-03T18:50:57Z"],[26458,2.5261,0.7093,"2026-11-03T23:54:13Z"],[26516,2.545,0.7184,"2026-11-04T01:58:12Z"],[26633,2.5616,0.7263,"2026-11-04T08:47:53Z"],[26749,2.5671,0.7289,"2026-11-04T17:42:26Z"],[26810,2.5729,0.7317,"2026-11-04T21:15:29Z"],[26818,2.5788,0.7345,"2026-11-04T21:43:12Z"],[26823,2.5876,0.7387,"2026-11-04T22:36:16Z"],[26835,2.5895,0.7396,"2026-11-04T23:00:25Z"],[26869,2.5913,0.7404,"2026-11-05T00:15:28Z"],[26959,2.5913,0.7404,"2026-11-05T07:59:59Z"]]},{"key":"muse_moe_w4a16","points":[[22,1.5407,0.2339,"2026-10-23T01:45:21Z"],[24,1.6217,0.2716,"2026-10-23T01:48:09Z"],[27,1.7176,0.3162,"2026-10-23T01:51:27Z"],[32,1.7677,0.3395,"2026-10-23T01:59:58Z"],[36,1.7793,0.3449,"2026-10-23T02:07:18Z"],[38,1.792,0.3508,"2026-10-23T02:10:22Z"],[46,1.8021,0.3555,"2026-10-23T02:16:37Z"],[50,1.841,0.3736,"2026-10-23T02:17:59Z"],[97,1.8543,0.3798,"2026-10-23T03:03:32Z"],[98,1.9115,0.4065,"2026-10-23T03:03:36Z"],[101,1.9223,0.4115,"2026-10-23T03:10:56Z"],[102,1.9463,0.4227,"2026-10-23T03:14:17Z"],[105,1.9533,0.4259,"2026-10-23T03:16:24Z"],[114,2.0337,0.4633,"2026-10-23T03:44:05Z"],[119,2.0904,0.4897,"2026-10-23T03:53:57Z"],[130,2.1565,0.5205,"2026-10-23T04:07:10Z"],[140,2.1642,0.524,"2026-10-23T04:50:29Z"],[142,2.1669,0.5253,"2026-10-23T05:03:07Z"],[145,2.1887,0.5355,"2026-10-23T05:11:49Z"],[147,2.2132,0.5469,"2026-10-23T05:14:51Z"],[153,2.2381,0.5584,"2026-10-23T05:43:27Z"],[158,2.2421,0.5603,"2026-10-23T06:20:05Z"],[159,2.2486,0.5633,"2026-10-23T06:22:23Z"],[160,2.2939,0.5844,"2026-10-23T06:32:05Z"],[177,2.3347,0.6034,"2026-10-23T07:30:19Z"],[196,2.3533,0.6121,"2026-10-23T10:47:10Z"],[199,2.3733,0.6213,"2026-10-23T10:57:21Z"],[216,2.3832,0.626,"2026-10-23T12:08:16Z"],[217,2.3893,0.6288,"2026-10-23T12:23:17Z"],[219,2.4206,0.6434,"2026-10-23T12:44:10Z"],[224,2.4237,0.6448,"2026-10-23T14:00:03Z"],[240,2.4518,0.6579,"2026-10-23T19:08:26Z"],[275,2.4584,0.661,"2026-10-23T23:07:11Z"],[286,2.4701,0.6664,"2026-10-24T00:25:23Z"],[358,2.4865,0.6741,"2026-10-24T20:29:38Z"],[480,2.4875,0.6745,"2026-10-25T03:02:21Z"],[501,2.4964,0.6786,"2026-10-25T04:25:08Z"],[519,2.5108,0.6854,"2026-10-25T06:00:34Z"],[993,2.5172,0.6883,"2026-10-27T05:12:29Z"],[1031,2.523,0.691,"2026-10-27T05:35:59Z"],[1036,2.5257,0.6923,"2026-10-27T05:36:46Z"],[1081,2.5284,0.6935,"2026-10-27T05:58:26Z"],[1905,2.5381,0.698,"2026-10-27T11:51:03Z"],[1911,2.5457,0.7016,"2026-10-27T11:53:30Z"],[1924,2.5533,0.7051,"2026-10-27T12:01:48Z"],[1974,2.5571,0.7069,"2026-10-27T12:26:59Z"],[2036,2.5577,0.7071,"2026-10-27T12:55:09Z"],[2070,2.575,0.7152,"2026-10-27T13:08:25Z"],[2553,2.5958,0.7249,"2026-10-27T16:10:52Z"],[2557,2.6113,0.7321,"2026-10-27T16:12:02Z"],[2570,2.6319,0.7417,"2026-10-27T16:15:42Z"],[5041,2.6527,0.7514,"2026-10-28T11:51:57Z"],[5042,2.6544,0.7522,"2026-10-28T11:53:06Z"],[5320,2.6615,0.7555,"2026-10-28T15:18:30Z"],[5328,2.6675,0.7583,"2026-10-28T15:21:08Z"],[6195,2.6706,0.7597,"2026-10-28T20:05:12Z"],[6223,2.6713,0.76,"2026-10-28T20:18:01Z"],[8849,2.6803,0.7642,"2026-10-31T21:12:50Z"],[9297,2.6803,0.7642,"2026-11-05T07:59:59Z"]]},{"key":"muse_attention_decode","points":[[28,1.718,0.2863,"2026-10-23T01:30:34Z"],[30,1.8574,0.3436,"2026-10-23T01:40:59Z"],[33,1.9629,0.387,"2026-10-23T02:10:02Z"],[41,1.9703,0.3901,"2026-10-23T02:23:06Z"],[51,1.9987,0.4017,"2026-10-23T02:39:29Z"],[53,2.0243,0.4123,"2026-10-23T02:44:01Z"],[65,2.0379,0.4179,"2026-10-23T03:07:29Z"],[66,2.0419,0.4195,"2026-10-23T03:12:20Z"],[67,2.0522,0.4237,"2026-10-23T03:13:37Z"],[70,2.1363,0.4583,"2026-10-23T03:27:19Z"],[76,2.235,0.4989,"2026-10-23T04:02:01Z"],[82,2.2556,0.5074,"2026-10-23T04:35:18Z"],[87,2.2777,0.5165,"2026-10-23T05:17:06Z"],[88,2.3026,0.5267,"2026-10-23T05:17:39Z"],[94,2.3145,0.5316,"2026-10-23T05:55:00Z"],[95,2.3377,0.5411,"2026-10-23T06:11:04Z"],[102,2.361,0.5507,"2026-10-23T10:20:59Z"],[106,2.3634,0.5517,"2026-10-23T10:49:55Z"],[125,2.3759,0.5568,"2026-10-23T12:30:33Z"],[126,2.4286,0.5785,"2026-10-23T12:50:23Z"],[142,2.439,0.5828,"2026-10-23T17:22:10Z"],[160,2.454,0.5889,"2026-10-23T19:53:41Z"],[190,2.458,0.5906,"2026-10-23T22:23:06Z"],[225,2.4711,0.596,"2026-10-24T00:51:02Z"],[226,2.4784,0.599,"2026-10-24T00:53:19Z"],[230,2.4846,0.6016,"2026-10-24T01:15:07Z"],[236,2.5161,0.6145,"2026-10-24T01:25:49Z"],[297,2.5245,0.618,"2026-10-24T05:09:38Z"],[300,2.5499,0.6284,"2026-10-24T05:15:02Z"],[489,2.6111,0.6536,"2026-10-25T00:03:48Z"],[508,2.6143,0.6549,"2026-10-25T00:31:29Z"],[519,2.621,0.6576,"2026-10-25T00:40:45Z"],[529,2.6385,0.6648,"2026-10-25T00:55:12Z"],[586,2.6455,0.6677,"2026-10-25T03:00:54Z"],[601,2.679,0.6815,"2026-10-25T03:06:30Z"],[708,2.7275,0.7014,"2026-10-25T06:50:10Z"],[1086,2.735,0.7045,"2026-10-27T04:59:52Z"],[1116,2.7359,0.7049,"2026-10-27T05:26:37Z"],[1452,2.7576,0.7138,"2026-10-27T09:39:00Z"],[1461,2.7882,0.7264,"2026-10-27T09:46:26Z"],[2155,2.8396,0.7475,"2026-10-27T23:05:58Z"],[2363,2.8507,0.7521,"2026-10-28T02:46:36Z"],[2377,2.8778,0.7632,"2026-10-28T02:57:54Z"],[2386,2.8778,0.7632,"2026-10-28T03:02:35Z"],[2514,2.8809,0.7645,"2026-10-28T04:29:29Z"],[2532,2.9317,0.7854,"2026-10-28T04:37:34Z"],[2541,2.9436,0.7903,"2026-10-28T04:39:48Z"],[3953,2.975,0.8032,"2026-10-28T10:08:27Z"],[5940,3.0147,0.8195,"2026-10-28T19:50:47Z"],[8022,3.0292,0.8255,"2026-10-29T10:10:00Z"],[8804,3.0478,0.8331,"2026-10-29T20:43:44Z"],[10162,3.0498,0.834,"2026-10-30T08:54:42Z"],[10863,3.0669,0.841,"2026-10-31T01:58:25Z"],[11050,3.1106,0.8589,"2026-10-31T03:54:25Z"],[11055,3.1572,0.8781,"2026-10-31T04:06:24Z"],[11391,3.1595,0.8791,"2026-10-31T18:47:48Z"],[11503,3.1704,0.8836,"2026-10-31T19:47:54Z"],[12256,3.1704,0.8836,"2026-11-05T07:59:59Z"]]},{"key":"muse_prefill_layer","points":[[12,1.4871,0.4006,"2026-10-23T01:39:20Z"],[15,1.5064,0.4161,"2026-10-23T01:54:28Z"],[17,1.5114,0.4201,"2026-10-23T02:09:46Z"],[20,1.5132,0.4215,"2026-10-23T02:30:33Z"],[21,1.5407,0.4434,"2026-10-23T02:36:34Z"],[24,1.545,0.4469,"2026-10-23T02:49:02Z"],[25,1.5609,0.4596,"2026-10-23T03:08:15Z"],[33,1.5642,0.4622,"2026-10-23T05:00:34Z"],[37,1.5899,0.4828,"2026-10-23T05:45:35Z"],[39,1.5982,0.4894,"2026-10-23T07:10:58Z"],[46,1.6045,0.4944,"2026-10-23T09:52:10Z"],[49,1.6082,0.4974,"2026-10-23T12:15:28Z"],[51,1.6149,0.5027,"2026-10-23T13:04:36Z"],[55,1.6449,0.5267,"2026-10-23T14:14:32Z"],[58,1.6472,0.5285,"2026-10-23T15:10:48Z"],[59,1.6481,0.5293,"2026-10-23T16:53:10Z"],[61,1.6498,0.5307,"2026-10-23T19:24:28Z"],[68,1.6516,0.5321,"2026-10-23T20:33:12Z"],[73,1.6588,0.5378,"2026-10-23T21:20:52Z"],[90,1.6646,0.5425,"2026-10-23T22:28:04Z"],[97,1.6714,0.5479,"2026-10-23T23:05:36Z"],[98,1.6741,0.55,"2026-10-23T23:11:02Z"],[105,1.6753,0.551,"2026-10-24T01:44:54Z"],[132,1.6753,0.551,"2026-10-24T18:57:34Z"],[137,1.681,0.5555,"2026-10-24T19:20:29Z"],[166,1.6817,0.5561,"2026-10-24T23:07:06Z"],[167,1.6822,0.5565,"2026-10-24T23:13:55Z"],[177,1.6827,0.5569,"2026-10-25T00:20:02Z"],[188,1.6829,0.5571,"2026-10-25T00:30:51Z"],[189,1.6868,0.5602,"2026-10-25T00:37:21Z"],[194,1.6874,0.5607,"2026-10-25T01:29:42Z"],[207,1.6909,0.5635,"2026-10-25T01:56:42Z"],[213,1.6913,0.5638,"2026-10-25T02:34:11Z"],[216,1.6945,0.5663,"2026-10-25T02:45:08Z"],[231,1.6986,0.5696,"2026-10-25T03:04:02Z"],[327,1.701,0.5716,"2026-10-25T07:36:33Z"],[363,1.705,0.5747,"2026-10-25T08:32:02Z"],[505,1.8625,0.7005,"2026-10-27T04:40:13Z"],[857,1.8671,0.7042,"2026-10-27T13:54:09Z"],[874,1.8775,0.7125,"2026-10-27T14:04:43Z"],[985,1.8825,0.7165,"2026-10-27T15:39:57Z"],[1033,1.9752,0.7906,"2026-10-27T16:53:31Z"],[1034,1.9777,0.7926,"2026-10-27T16:56:31Z"],[1036,1.979,0.7936,"2026-10-27T16:59:02Z"],[1906,1.9916,0.8036,"2026-10-28T02:58:32Z"],[1916,1.9949,0.8063,"2026-10-28T03:01:07Z"],[2252,2.0027,0.8125,"2026-10-28T06:45:39Z"],[2254,2.0118,0.8198,"2026-10-28T06:47:20Z"],[2335,2.0537,0.8533,"2026-10-28T08:06:23Z"],[3536,2.0887,0.8813,"2026-10-30T07:57:08Z"],[3953,2.1339,0.9174,"2026-10-30T17:59:25Z"],[3956,2.1419,0.9237,"2026-10-30T18:02:52Z"],[4080,2.1419,0.9237,"2026-11-05T07:59:59Z"]]},{"key":"muse_bf16_proj","points":[[35,1.2104,0.2252,"2026-10-23T01:50:39Z"],[43,1.2219,0.237,"2026-10-23T02:07:42Z"],[46,1.2374,0.2528,"2026-10-23T02:10:44Z"],[60,1.2414,0.2569,"2026-10-23T02:23:06Z"],[61,1.2416,0.2571,"2026-10-23T02:23:12Z"],[67,1.2457,0.2613,"2026-10-23T02:30:10Z"],[72,1.2677,0.2837,"2026-10-23T02:36:17Z"],[83,1.2958,0.3124,"2026-10-23T02:52:27Z"],[94,1.3033,0.3201,"2026-10-23T03:14:32Z"],[97,1.3099,0.3268,"2026-10-23T03:19:29Z"],[99,1.315,0.332,"2026-10-23T03:20:07Z"],[100,1.3297,0.347,"2026-10-23T03:23:27Z"],[135,1.3309,0.3483,"2026-10-23T05:47:10Z"],[136,1.3502,0.3679,"2026-10-23T05:48:29Z"],[149,1.4296,0.4489,"2026-10-23T08:20:31Z"],[152,1.4371,0.4567,"2026-10-23T08:27:07Z"],[154,1.4509,0.4707,"2026-10-23T08:31:27Z"],[167,1.4584,0.4784,"2026-10-23T09:32:54Z"],[291,1.4622,0.4822,"2026-10-24T18:17:20Z"],[324,1.4671,0.4873,"2026-10-24T19:36:39Z"],[422,1.469,0.4892,"2026-10-25T00:33:46Z"],[541,1.4756,0.4959,"2026-10-26T23:41:54Z"],[667,1.4835,0.5039,"2026-10-27T04:58:52Z"],[1103,1.4883,0.5088,"2026-10-27T08:19:29Z"],[2419,1.5129,0.534,"2026-10-27T14:12:12Z"],[2863,1.5141,0.5352,"2026-10-27T18:10:36Z"],[2915,1.5154,0.5365,"2026-10-27T18:38:57Z"],[2920,1.5183,0.5395,"2026-10-27T18:43:06Z"],[3547,1.5379,0.5595,"2026-10-27T21:28:49Z"],[3618,1.5412,0.5628,"2026-10-27T21:54:12Z"],[3626,1.5428,0.5645,"2026-10-27T21:58:35Z"],[3677,1.548,0.5698,"2026-10-27T22:19:01Z"],[3743,1.5516,0.5735,"2026-10-27T22:52:25Z"],[3788,1.5519,0.5738,"2026-10-27T23:08:35Z"],[3888,1.5526,0.5745,"2026-10-27T23:50:52Z"],[3890,1.5549,0.5768,"2026-10-27T23:51:36Z"],[4971,1.5554,0.5773,"2026-10-28T18:03:59Z"],[4995,1.557,0.579,"2026-10-28T18:39:52Z"],[5154,1.6035,0.6264,"2026-10-29T09:06:55Z"],[5529,1.6096,0.6326,"2026-10-30T07:35:59Z"],[5546,1.6291,0.6525,"2026-10-30T07:52:09Z"],[5560,1.6358,0.6594,"2026-10-30T08:00:52Z"],[5982,1.6408,0.6645,"2026-10-31T23:03:46Z"],[6026,1.6408,0.6645,"2026-11-05T07:59:59Z"]]}],"fit_lines":[{"key":"muse_decode_layer","from":[24,0.3416745846969914],"to":[26959,0.8039299141909951],"slope":0.146},{"key":"muse_moe_w4a16","from":[22,0.3809247448376577],"to":[9297,0.8899980570356641],"slope":0.1867},{"key":"muse_attention_decode","from":[28,0.4105650155063777],"to":[12256,0.9241545388903912],"slope":0.1903},{"key":"muse_prefill_layer","from":[12,0.35452141806257703],"to":[4080,0.8263332760656492],"slope":0.1891},{"key":"muse_bf16_proj","from":[35,0.2576971799133413],"to":[6026,0.6098045868200921],"slope":0.1591}],"corr":{"muse_decode_layer":0.9376951533684155,"muse_moe_w4a16":0.9088673101155033,"muse_attention_decode":0.9747060880211842,"muse_prefill_layer":0.9716902268616215,"muse_bf16_proj":0.9472276589266995},"corr_min":0.9088673101155033,"corr_max":0.9747060880211842,"pooled":0.1734,"pooled_ci":[0.168,0.1767],"pooled_scored":0.1719,"records":{"b_runs":0.543,"b_runs_ci90":[0.477,0.62],"b_cum":-0.517,"b_cum_ci90":[-0.604,-0.458],"problem_hours":1609,"records":224},"per1k":[{"day":"2026-10-22","gain":3055.686,"records":13,"runs":48},{"day":"2026-10-23","gain":153.799,"records":80,"runs":1269},{"day":"2026-10-24","gain":10.382,"records":14,"runs":763},{"day":"2026-10-25","gain":11.778,"records":15,"runs":1054},{"day":"2026-10-26","gain":21.034,"records":13,"runs":353},{"day":"2026-10-27","gain":2.665,"records":28,"runs":11765},{"day":"2026-10-28","gain":0.802,"records":15,"runs":15109},{"day":"2026-10-29","gain":0.541,"records":3,"runs":7539},{"day":"2026-10-30","gain":1.704,"records":13,"runs":10347},{"day":"2026-10-31","gain":1.846,"records":8,"runs":3428},{"day":"2026-11-01","gain":0.491,"records":4,"runs":5056},{"day":"2026-11-02","gain":3.405,"records":3,"runs":730},{"day":"2026-11-03","gain":29.129,"records":6,"runs":128},{"day":"2026-11-04","gain":34.022,"records":9,"runs":306},{"day":"2026-11-05","gain":0.0,"records":0,"runs":99}],"caption":"Dots: each new best in the clone protocol, where the swarm made its decisions, plus one point per day. Dashed: a least-squares fit of gap closed on log10 of runs, one per problem. Switch to speedup to see the raw clone x."},"runs_daily":{"days":["2026-10-22","2026-10-23","2026-10-24","2026-10-25","2026-10-26","2026-10-27","2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-04","2026-11-05"],"series":[{"label":"H200 boxes (clone of the judge)","values":[0,751,752,1023,360,11633,15063,7503,10309,3416,5039,724,112,284,92]},{"label":"Modal clone","values":[14,382,0,30,1,43,0,0,0,0,0,0,0,0,0]},{"label":"Serverless judge protocol","values":[0,0,0,0,0,0,0,0,0,0,0,28,430,96,0]},{"label":"Organizers' judge","values":[34,143,43,20,0,93,46,36,38,12,17,6,16,22,7]}]},"climb":{"start":"2026-10-22T22:00:00Z","end":"2026-11-05T07:59:59Z","counters":[["2026-10-22T22:00:00Z",12,4.0,0,12],["2026-10-22T23:00:00Z",48,4.0,0,14],["2026-10-23T00:00:00Z",70,6.0,4279,39],["2026-10-23T01:00:00Z",145,24.0,4279,49],["2026-10-23T02:00:00Z",331,113.3,4279,85],["2026-10-23T03:00:00Z",419,209.9,4279,93],["2026-10-23T04:00:00Z",474,306.5,4279,141],["2026-10-23T05:00:00Z",547,399.1,4279,154],["2026-10-23T06:00:00Z",596,490.3,4279,145],["2026-10-23T07:00:00Z",627,573.0,4279,45],["2026-10-23T08:00:00Z",654,600.9,4279,12],["2026-10-23T09:00:00Z",693,635.6,4279,18],["2026-10-23T10:00:00Z",747,697.5,4279,20],["2026-10-23T11:00:00Z",810,750.2,4279,16],["2026-10-23T12:00:00Z",848,802.1,4279,10],["2026-10-23T13:00:00Z",861,836.1,4279,12],["2026-10-23T14:00:00Z",892,872.8,4279,13],["2026-10-23T15:00:00Z",917,914.1,4279,11],["2026-10-23T16:00:00Z",932,959.4,4279,11],["2026-10-23T17:00:00Z",950,1004.0,4279,9],["2026-10-23T18:00:00Z",1010,1066.6,4279,15],["2026-10-23T19:00:00Z",1063,1121.9,4279,19],["2026-10-23T20:00:00Z",1103,1190.6,4279,34],["2026-10-23T21:00:00Z",1169,1293.8,4279,90],["2026-10-23T22:00:00Z",1243,1387.1,4279,53],["2026-10-23T23:00:00Z",1324,1480.4,4279,53],["2026-10-24T00:00:00Z",1368,1527.0,5412,26],["2026-10-24T01:00:00Z",1430,1666.2,5412,72],["2026-10-24T02:00:00Z",1477,1864.1,5412,72],["2026-10-24T03:00:00Z",1538,2087.3,5412,84],["2026-10-24T04:00:00Z",1565,2286.5,5412,91],["2026-10-24T05:00:00Z",1596,2378.4,5412,24],["2026-10-24T06:00:00Z",1648,2432.4,5412,41],["2026-10-24T07:00:00Z",1661,2442.4,5412,44],["2026-10-24T08:00:00Z",1661,2443.1,5412,11],["2026-10-24T09:00:00Z",1661,2443.1,5412,5],["2026-10-24T10:00:00Z",1661,2443.1,5412,7],["2026-10-24T11:00:00Z",1661,2443.7,5412,8],["2026-10-24T12:00:00Z",1661,2443.7,5412,3],["2026-10-24T13:00:00Z",1661,2444.4,5412,11],["2026-10-24T14:00:00Z",1661,2444.4,5412,6],["2026-10-24T15:00:00Z",1661,2444.4,5412,2],["2026-10-24T16:00:00Z",1661,2445.1,5412,6],["2026-10-24T17:00:00Z",1662,2451.1,5412,21],["2026-10-24T18:00:00Z",1715,2555.0,5412,90],["2026-10-24T19:00:00Z",1806,2662.3,5412,79],["2026-10-24T20:00:00Z",1880,2893.4,5412,110],["2026-10-24T21:00:00Z",1951,3014.7,5412,95],["2026-10-24T22:00:00Z",2062,3066.0,5412,56],["2026-10-24T23:00:00Z",2119,3153.3,5412,72],["2026-10-25T00:00:00Z",2263,3697.6,6910,182],["2026-10-25T01:00:00Z",2342,4220.6,6910,185],["2026-10-25T02:00:00Z",2451,4647.6,6910,183],["2026-10-25T03:00:00Z",2617,4780.9,6910,170],["2026-10-25T04:00:00Z",2650,4831.5,6910,76],["2026-10-25T05:00:00Z",2749,4925.5,6910,76],["2026-10-25T06:00:00Z",2879,5031.4,6910,98],["2026-10-25T07:00:00Z",3055,5108.0,6910,82],["2026-10-25T08:00:00Z",3192,5163.3,6910,45],["2026-10-25T09:00:00Z",3192,5168.6,6910,17],["2026-10-25T10:00:00Z",3192,5174.0,6910,0],["2026-10-25T11:00:00Z",3192,5174.0,6910,0],["2026-10-25T12:00:00Z",3192,5174.0,6910,0],["2026-10-25T13:00:00Z",3192,5174.0,6910,0],["2026-10-25T14:00:00Z",3192,5174.0,6910,0],["2026-10-25T15:00:00Z",3192,5174.0,6910,0],["2026-10-25T16:00:00Z",3192,5174.0,6910,0],["2026-10-25T17:00:00Z",3192,5174.6,6910,0],["2026-10-25T18:00:00Z",3192,5178.6,6910,0],["2026-10-25T19:00:00Z",3192,5180.0,6910,0],["2026-10-25T20:00:00Z",3192,5181.3,6910,25],["2026-10-25T21:00:00Z",3192,5181.3,6910,0],["2026-10-25T22:00:00Z",3192,5181.3,6910,0],["2026-10-25T23:00:00Z",3192,5181.3,6910,0],["2026-10-26T00:00:00Z",3192,5181.3,7119,0],["2026-10-26T01:00:00Z",3192,5181.3,7119,0],["2026-10-26T02:00:00Z",3192,5181.3,7119,0],["2026-10-26T03:00:00Z",3192,5181.3,7119,0],["2026-10-26T04:00:00Z",3192,5181.3,7119,0],["2026-10-26T05:00:00Z",3192,5181.3,7119,0],["2026-10-26T06:00:00Z",3192,5181.3,7119,0],["2026-10-26T07:00:00Z",3192,5181.3,7119,0],["2026-10-26T08:00:00Z",3192,5181.3,7119,0],["2026-10-26T09:00:00Z",3192,5181.3,7119,0],["2026-10-26T10:00:00Z",3192,5181.3,7119,0],["2026-10-26T11:00:00Z",3192,5181.3,7119,0],["2026-10-26T12:00:00Z",3192,5181.3,7119,0],["2026-10-26T13:00:00Z",3192,5181.3,7119,0],["2026-10-26T14:00:00Z",3192,5181.3,7119,0],["2026-10-26T15:00:00Z",3192,5181.3,7119,0],["2026-10-26T16:00:00Z",3192,5181.3,7119,0],["2026-10-26T17:00:00Z",3192,5181.3,7119,0],["2026-10-26T18:00:00Z",3192,5181.3,7119,0],["2026-10-26T19:00:00Z",3192,5181.3,7119,0],["2026-10-26T20:00:00Z",3197,5186.0,7119,0],["2026-10-26T21:00:00Z",3243,5234.6,7119,17],["2026-10-26T22:00:00Z",3308,5297.2,7119,49],["2026-10-26T23:00:00Z",3553,5362.5,7119,75],["2026-10-27T00:00:00Z",3878,5434.5,9341,80],["2026-10-27T01:00:00Z",4118,5499.7,9341,119],["2026-10-27T02:00:00Z",4228,5529.7,9341,94],["2026-10-27T03:00:00Z",4267,5549.7,9341,76],["2026-10-27T04:00:00Z",4539,5609.7,9341,131],["2026-10-27T05:00:00Z",4987,5682.3,9341,30],["2026-10-27T06:00:00Z",5529,5767.6,9341,12],["2026-10-27T07:00:00Z",6048,5848.2,9341,9],["2026-10-27T08:00:00Z",6597,5904.8,9341,15],["2026-10-27T09:00:00Z",7084,5972.1,9341,15],["2026-10-27T10:00:00Z",7698,6058.1,9341,17],["2026-10-27T11:00:00Z",8243,6143.3,9341,12],["2026-10-27T12:00:00Z",8897,6229.3,9341,15],["2026-10-27T13:00:00Z",9459,6312.6,9341,9],["2026-10-27T14:00:00Z",10152,6398.5,9341,15],["2026-10-27T15:00:00Z",10707,6483.8,9341,17],["2026-10-27T16:00:00Z",11223,6569.7,9341,35],["2026-10-27T17:00:00Z",11675,6655.0,9341,40],["2026-10-27T18:00:00Z",12136,6744.3,9341,38],["2026-10-27T19:00:00Z",12684,6856.9,9341,36],["2026-10-27T20:00:00Z",13442,6956.8,9341,11],["2026-10-27T21:00:00Z",14141,7050.1,9341,35],["2026-10-27T22:00:00Z",14776,7158.7,9341,34],["2026-10-27T23:00:00Z",15322,7272.6,9341,104],["2026-10-28T00:00:00Z",15825,7411.8,11210,42],["2026-10-28T01:00:00Z",16439,7546.4,11210,20],["2026-10-28T02:00:00Z",17022,7681.7,11210,41],["2026-10-28T03:00:00Z",17421,7820.2,11210,44],["2026-10-28T04:00:00Z",18289,7958.8,11210,133],["2026-10-28T05:00:00Z",19129,8094.7,11210,85],["2026-10-28T06:00:00Z",19656,8234.0,11210,60],["2026-10-28T07:00:00Z",20231,8376.5,11210,74],["2026-10-28T08:00:00Z",20749,8502.5,11210,32],["2026-10-28T09:00:00Z",21381,8619.7,11210,29],["2026-10-28T10:00:00Z",21912,8737.0,11210,16],["2026-10-28T11:00:00Z",22377,8848.9,11210,20],["2026-10-28T12:00:00Z",22926,8958.8,11210,27],["2026-10-28T13:00:00Z",23489,9082.1,11210,59],["2026-10-28T14:00:00Z",24167,9216.7,11210,35],["2026-10-28T15:00:00Z",24911,9354.6,11210,39],["2026-10-28T16:00:00Z",25564,9638.4,11210,193],["2026-10-28T17:00:00Z",26409,9960.8,11210,157],["2026-10-28T18:00:00Z",27342,10161.4,11210,94],["2026-10-28T19:00:00Z",27906,10431.2,11210,320],["2026-10-28T20:00:00Z",28831,10665.0,11210,228],["2026-10-28T21:00:00Z",29650,10919.6,11210,201],["2026-10-28T22:00:00Z",30175,11212.0,11210,216],["2026-10-28T23:00:00Z",30431,11335.3,11210,259],["2026-10-29T00:00:00Z",30503,11539.1,12411,481],["2026-10-29T01:00:00Z",30742,11795.6,12411,349],["2026-10-29T02:00:00Z",31049,12230.7,12411,348],["2026-10-29T03:00:00Z",31282,12510.5,12411,228],["2026-10-29T04:00:00Z",31462,12890.9,12411,393],["2026-10-29T05:00:00Z",31889,13068.8,12411,340],["2026-10-29T06:00:00Z",32333,13297.3,12411,173],["2026-10-29T07:00:00Z",32895,13513.2,12411,176],["2026-10-29T08:00:00Z",33399,13725.1,12411,187],["2026-10-29T09:00:00Z",33762,13853.0,12411,106],["2026-10-29T10:00:00Z",34098,13992.2,12411,123],["2026-10-29T11:00:00Z",34331,14086.8,12411,102],["2026-10-29T12:00:00Z",34539,14176.1,12411,81],["2026-10-29T13:00:00Z",34609,14255.4,12411,105],["2026-10-29T14:00:00Z",34620,14437.9,12411,182],["2026-10-29T15:00:00Z",34760,14635.1,12411,166],["2026-10-29T16:00:00Z",34981,14766.4,12411,133],["2026-10-29T17:00:00Z",35177,14884.3,12411,111],["2026-10-29T18:00:00Z",35555,15064.8,12411,148],["2026-10-29T19:00:00Z",35857,15271.4,12411,157],["2026-10-29T20:00:00Z",36374,15551.9,12411,136],["2026-10-29T21:00:00Z",36887,15771.7,12411,112],["2026-10-29T22:00:00Z",37437,15956.3,12411,82],["2026-10-29T23:00:00Z",37970,16142.8,12411,73],["2026-10-30T00:00:00Z",38358,16258.7,13096,53],["2026-10-30T01:00:00Z",39115,16357.3,13096,45],["2026-10-30T02:00:00Z",39812,16461.9,13096,42],["2026-10-30T03:00:00Z",40447,16559.2,13096,37],["2026-10-30T04:00:00Z",41126,16664.5,13096,38],["2026-10-30T05:00:00Z",41712,16769.7,13096,43],["2026-10-30T06:00:00Z",42300,16877.7,13096,30],["2026-10-30T07:00:00Z",43189,16987.6,13096,27],["2026-10-30T08:00:00Z",43937,17096.2,13096,23],["2026-10-30T09:00:00Z",44616,17200.1,13096,23],["2026-10-30T10:00:00Z",45067,17309.4,13096,24],["2026-10-30T11:00:00Z",45677,17410.7,13096,28],["2026-10-30T12:00:00Z",46163,17505.9,13096,32],["2026-10-30T13:00:00Z",46568,17612.5,13096,30],["2026-10-30T14:00:00Z",46813,17714.5,13096,27],["2026-10-30T15:00:00Z",47028,17814.4,13096,18],["2026-10-30T16:00:00Z",47401,17896.3,13096,25],["2026-10-30T17:00:00Z",47518,17944.3,13096,10],["2026-10-30T18:00:00Z",47583,17976.3,13096,4],["2026-10-30T19:00:00Z",47629,17992.9,13096,4],["2026-10-30T20:00:00Z",47636,17999.6,13096,3],["2026-10-30T21:00:00Z",47765,18024.9,13096,4],["2026-10-30T22:00:00Z",47854,18114.9,13096,54],["2026-10-30T23:00:00Z",48317,18230.8,13096,54],["2026-10-31T00:00:00Z",48573,18297.4,13668,40],["2026-10-31T01:00:00Z",48800,18420.7,13668,54],["2026-10-31T02:00:00Z",49173,18557.9,13668,52],["2026-10-31T03:00:00Z",49386,18629.2,13668,43],["2026-10-31T04:00:00Z",49453,18681.8,13668,27],["2026-10-31T05:00:00Z",49504,18713.8,13668,16],["2026-10-31T06:00:00Z",49504,18723.8,13668,13],["2026-10-31T07:00:00Z",49504,18734.5,13668,13],["2026-10-31T08:00:00Z",49504,18743.8,13668,13],["2026-10-31T09:00:00Z",49504,18751.8,13668,13],["2026-10-31T10:00:00Z",49504,18758.4,13668,13],["2026-10-31T11:00:00Z",49504,18767.8,13668,13],["2026-10-31T12:00:00Z",49504,18779.1,13668,13],["2026-10-31T13:00:00Z",49504,18794.4,13668,13],["2026-10-31T14:00:00Z",49504,18800.4,13668,13],["2026-10-31T15:00:00Z",49504,18808.4,13668,13],["2026-10-31T16:00:00Z",49528,18847.1,13668,46],["2026-10-31T17:00:00Z",49838,18979.6,13668,58],["2026-10-31T18:00:00Z",50242,19156.8,13668,60],["2026-10-31T19:00:00Z",50787,19378.7,13668,59],["2026-10-31T20:00:00Z",51246,19502.0,13668,59],["2026-10-31T21:00:00Z",51419,19557.9,13668,18],["2026-10-31T22:00:00Z",51499,19593.2,13668,22],["2026-10-31T23:00:00Z",51745,19661.9,13668,23],["2026-11-01T00:00:00Z",52004,19735.8,14401,23],["2026-11-01T01:00:00Z",52231,19847.7,14401,28],["2026-11-01T02:00:00Z",52477,19960.3,14401,28],["2026-11-01T03:00:00Z",52794,20064.9,14401,26],["2026-11-01T04:00:00Z",53026,20157.5,14401,27],["2026-11-01T05:00:00Z",53425,20282.1,14401,27],["2026-11-01T06:00:00Z",53716,20390.7,14401,27],["2026-11-01T07:00:00Z",53982,20494.0,14401,29],["2026-11-01T08:00:00Z",54277,20603.2,14401,29],["2026-11-01T09:00:00Z",54486,20706.5,14401,28],["2026-11-01T10:00:00Z",54748,20823.1,14401,27],["2026-11-01T11:00:00Z",55063,20947.0,14401,27],["2026-11-01T12:00:00Z",55320,21060.3,14401,27],["2026-11-01T13:00:00Z",55525,21162.9,14401,26],["2026-11-01T14:00:00Z",55713,21258.8,14401,25],["2026-11-01T15:00:00Z",55835,21366.1,14401,32],["2026-11-01T16:00:00Z",56004,21465.3,14401,44],["2026-11-01T17:00:00Z",56136,21572.6,14401,42],["2026-11-01T18:00:00Z",56286,21657.9,14401,36],["2026-11-01T19:00:00Z",56396,21762.5,14401,44],["2026-11-01T20:00:00Z",56506,21895.7,14401,29],["2026-11-01T21:00:00Z",56607,22021.0,14401,30],["2026-11-01T22:00:00Z",56720,22148.9,14401,28],["2026-11-01T23:00:00Z",56801,22276.1,14401,27],["2026-11-02T00:00:00Z",56879,22390.7,14743,37],["2026-11-02T01:00:00Z",56942,22509.3,14743,32],["2026-11-02T02:00:00Z",56997,22637.9,14743,23],["2026-11-02T03:00:00Z",57037,22749.2,14743,26],["2026-11-02T04:00:00Z",57075,22845.8,14743,23],["2026-11-02T05:00:00Z",57094,22935.1,14743,19],["2026-11-02T06:00:00Z",57134,23026.3,14743,20],["2026-11-02T07:00:00Z",57181,23124.3,14743,21],["2026-11-02T08:00:00Z",57221,23209.5,14743,21],["2026-11-02T09:00:00Z",57236,23330.1,14743,22],["2026-11-02T10:00:00Z",57261,23450.1,14743,21],["2026-11-02T11:00:00Z",57292,23548.7,14743,19],["2026-11-02T12:00:00Z",57321,23643.9,14743,15],["2026-11-02T13:00:00Z",57332,23741.9,14743,17],["2026-11-02T14:00:00Z",57340,23817.8,14743,18],["2026-11-02T15:00:00Z",57358,23900.4,14743,15],["2026-11-02T16:00:00Z",57372,23981.0,14743,15],["2026-11-02T17:00:00Z",57394,24086.3,14743,35],["2026-11-02T18:00:00Z",57431,24190.2,14743,35],["2026-11-02T19:00:00Z",57465,24289.5,14743,37],["2026-11-02T20:00:00Z",57491,24378.8,14743,26],["2026-11-02T21:00:00Z",57511,24458.1,14743,23],["2026-11-02T22:00:00Z",57535,24538.7,14743,20],["2026-11-02T23:00:00Z",57559,24614.6,14743,18],["2026-11-03T00:00:00Z",57590,24735.2,14899,22],["2026-11-03T01:00:00Z",57637,24845.8,14899,16],["2026-11-03T02:00:00Z",57666,24972.4,14899,16],["2026-11-03T03:00:00Z",57702,25093.6,14899,21],["2026-11-03T04:00:00Z",57731,25216.2,14899,18],["2026-11-03T05:00:00Z",57748,25315.5,14899,9],["2026-11-03T06:00:00Z",57773,25392.8,14899,9],["2026-11-03T07:00:00Z",57788,25462.1,14899,8],["2026-11-03T08:00:00Z",57812,25549.3,14899,7],["2026-11-03T09:00:00Z",57829,25657.9,14899,8],["2026-11-03T10:00:00Z",57844,25759.2,14899,9],["2026-11-03T11:00:00Z",57860,25861.8,14899,9],["2026-11-03T12:00:00Z",57888,25988.4,14899,9],["2026-11-03T13:00:00Z",57895,26077.0,14899,10],["2026-11-03T14:00:00Z",57913,26161.0,14899,11],["2026-11-03T15:00:00Z",57929,26238.2,14899,8],["2026-11-03T16:00:00Z",57939,26297.5,14899,13],["2026-11-03T17:00:00Z",57958,26388.1,14899,17],["2026-11-03T18:00:00Z",58001,26477.4,14899,14],["2026-11-03T19:00:00Z",58026,26568.0,14899,12],["2026-11-03T20:00:00Z",58050,26657.3,14899,12],["2026-11-03T21:00:00Z",58076,26703.9,14899,13],["2026-11-03T22:00:00Z",58097,26747.2,14899,9],["2026-11-03T23:00:00Z",58117,26790.5,14899,20],["2026-11-04T00:00:00Z",58143,26829.2,14935,22],["2026-11-04T01:00:00Z",58171,26866.5,14935,19],["2026-11-04T02:00:00Z",58183,26898.5,14935,19],["2026-11-04T03:00:00Z",58193,26934.4,14935,22],["2026-11-04T04:00:00Z",58212,26959.8,14935,19],["2026-11-04T05:00:00Z",58239,27025.7,14935,19],["2026-11-04T06:00:00Z",58261,27077.7,14935,17],["2026-11-04T07:00:00Z",58281,27120.3,14935,19],["2026-11-04T08:00:00Z",58291,27159.6,14935,16],["2026-11-04T09:00:00Z",58294,27185.6,14935,4],["2026-11-04T10:00:00Z",58316,27213.6,14935,5],["2026-11-04T11:00:00Z",58370,27244.9,14935,4],["2026-11-04T12:00:00Z",58378,27272.9,14935,6],["2026-11-04T13:00:00Z",58386,27292.9,14935,4],["2026-11-04T14:00:00Z",58388,27324.9,14935,6],["2026-11-04T15:00:00Z",58388,27355.5,14935,7],["2026-11-04T16:00:00Z",58391,27433.5,14935,8],["2026-11-04T17:00:00Z",58408,27484.1,14935,7],["2026-11-04T18:00:00Z",58423,27542.1,14935,7],["2026-11-04T19:00:00Z",58455,27622.7,14935,12],["2026-11-04T20:00:00Z",58466,27683.3,14935,10],["2026-11-04T21:00:00Z",58476,27744.6,14935,11],["2026-11-04T22:00:00Z",58492,27821.9,14935,9],["2026-11-04T23:00:00Z",58519,27891.8,14935,9],["2026-11-05T00:00:00Z",58543,27923.1,14936,8],["2026-11-05T01:00:00Z",58566,27987.8,14936,13],["2026-11-05T02:00:00Z",58588,28043.7,14936,12],["2026-11-05T03:00:00Z",58604,28077.7,14936,7],["2026-11-05T04:00:00Z",58616,28114.3,14936,6],["2026-11-05T05:00:00Z",58617,28131.7,14936,2],["2026-11-05T06:00:00Z",58617,28131.7,14936,1],["2026-11-05T07:00:00Z",58618,28131.7,14936,3]],"marks":[{"t":"2026-10-24T00:20:23Z","label":"#1 on the decode board"},{"t":"2026-11-04T00:43:38Z","label":"disqualification"}],"counter_labels":["GPU runs so far","cost of the work so far","ideas registered so far","agent sessions alive this hour"]},"models":{"rows":[{"model":"GPT-6 Astra","harness":"Codex CLI","lanes":85,"lane_unit":"derived GPU sessions","hours":170.4,"gpu_runs":598,"wins":24,"rate":14.08450704225352,"rlo":9.71127420504965,"rhi":19.807024726838964,"p":0.14118,"lo":0.0901,"hi":0.2144,"per_k_usd":null,"per_100_runs":4.01,"odds":{"odds_ratio":0.99,"ci90":[0.65,1.49],"n_valid_runs":538,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":1,"note":"84 of 85 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Claude Opus 5.5","harness":"Claude Code","lanes":633,"lane_unit":"derived GPU sessions","hours":1690.5,"gpu_runs":20565,"wins":107,"rate":6.329488317065957,"rlo":5.35767258104114,"rhi":7.4323654618912185,"p":0.08215,"lo":0.0659,"hi":0.10189999999999999,"per_k_usd":null,"per_100_runs":0.52,"odds":{"odds_ratio":1.0,"ci90":[1.0,1.0],"n_valid_runs":19925,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":2,"note":"596 of 633 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Gemini 4 Argon","harness":"pi","lanes":1103,"lane_unit":"mixed","hours":300.0,"gpu_runs":594,"wins":19,"rate":6.333333333333333,"rlo":4.146917576829817,"rhi":9.29250922030975,"p":0.013600000000000001,"lo":0.0089,"hi":0.0206,"per_k_usd":3.75,"per_100_runs":3.2,"odds":{"odds_ratio":1.52,"ci90":[0.96,2.28],"n_valid_runs":585,"note":"controls ln(runs so far on the problem) + problem"},"usd":4531.0,"billing":"API","rank":3,"note":"14 of 1103 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"GPT-6.1 Sol","harness":"pi","lanes":1162,"lane_unit":"ledger","hours":237.0,"gpu_runs":740,"wins":9,"rate":3.7974683544303796,"rlo":1.9800765286605277,"rhi":6.625092138984318,"p":0.00688,"lo":0.0039000000000000003,"hi":0.012199999999999999,"per_k_usd":14.88,"per_100_runs":1.22,"odds":{"odds_ratio":0.82,"ci90":[0.46,1.25],"n_valid_runs":698,"note":"controls ln(runs so far on the problem) + problem"},"usd":605.0,"billing":"API","rank":4,"note":"active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Gemini 4 Argon","harness":"Antigravity (agy)","lanes":67,"lane_unit":"derived GPU sessions","hours":55.0,"gpu_runs":259,"wins":3,"rate":5.454545454545454,"rlo":1.4730838452456676,"rhi":14.08179469120214,"p":0.04478,"lo":0.018000000000000002,"hi":0.1069,"per_k_usd":null,"per_100_runs":1.16,"odds":{"odds_ratio":0.94,"ci90":[0.35,1.85],"n_valid_runs":244,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":5,"note":"67 of 67 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses)"},{"model":"Kimi K3.1 Pro","harness":"Kimi Code CLI","lanes":350,"lane_unit":"mixed","hours":301.6,"gpu_runs":2874,"wins":4,"rate":1.326259946949602,"rlo":0.45112678929681843,"rhi":3.03261955435398,"p":0.00857,"lo":0.0034000000000000002,"hi":0.0213,"per_k_usd":null,"per_100_runs":0.14,"odds":{"odds_ratio":0.35,"ci90":[0.19,0.64],"n_valid_runs":2837,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":6,"note":"45 of 350 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Kimi K3","harness":"pi","lanes":15913,"lane_unit":"ledger","hours":699.5,"gpu_runs":2378,"wins":5,"rate":0.7147962830593281,"rlo":0.2809986047541031,"rhi":1.502068278039031,"p":0.00025,"lo":0.0001,"hi":0.0006,"per_k_usd":2.47,"per_100_runs":0.21,"odds":{"odds_ratio":1.93,"ci90":[0.85,3.37],"n_valid_runs":2319,"note":"controls ln(runs so far on the problem) + problem"},"usd":2027.0,"billing":"API","rank":7,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"Claude Sonnet 5.5","harness":"pi","lanes":32929,"lane_unit":"ledger","hours":471.7,"gpu_runs":5855,"wins":2,"rate":0.423998304006784,"rlo":0.07315787417252727,"rhi":1.3323563662201099,"p":6e-05,"lo":0.0,"hi":0.0002,"per_k_usd":0.99,"per_100_runs":0.03,"odds":{"odds_ratio":0.45,"ci90":[0.23,0.89],"n_valid_runs":5607,"note":"controls ln(runs so far on the problem) + problem"},"usd":2015.0,"billing":"API","rank":8,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"GPT-6.1 Sol","harness":"Codex CLI","lanes":896,"lane_unit":"mixed","hours":738.3,"gpu_runs":1006,"wins":2,"rate":0.27089259108763375,"rlo":0.04674057869047964,"rhi":0.8512427169795825,"p":0.00223,"lo":0.0007000000000000001,"hi":0.0067,"per_k_usd":null,"per_100_runs":0.2,"odds":{"odds_ratio":0.48,"ci90":[0.25,0.86],"n_valid_runs":993,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":9,"note":"32 of 896 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses)"},{"model":"GPT-6.1 Terra","harness":"Codex CLI","lanes":6110,"lane_unit":"mixed","hours":1847.6,"gpu_runs":1196,"wins":1,"rate":0.05412426932236415,"rlo":0.0021377452724046994,"rhi":0.25592077967962956,"p":0.00016,"lo":0.0,"hi":0.0007000000000000001,"per_k_usd":null,"per_100_runs":0.08,"odds":{"odds_ratio":0.64,"ci90":[0.39,1.13],"n_valid_runs":1183,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":10,"note":"72 of 6110 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses)"},{"model":"Claude Haiku 5.5","harness":"pi","lanes":5147,"lane_unit":"ledger","hours":140.4,"gpu_runs":1226,"wins":0,"rate":0.0,"rlo":0.0,"rhi":2.114485264885942,"p":0.0,"lo":0.0,"hi":0.0005,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.41,"ci90":[0.37,0.47],"n_valid_runs":1221,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":63.0,"billing":"API","rank":11,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"Claude Opus 5.5","harness":"pi","lanes":2352,"lane_unit":"mixed","hours":447.1,"gpu_runs":1484,"wins":0,"rate":0.0,"rlo":0.0,"rhi":0.6639985041153796,"p":0.0,"lo":0.0,"hi":0.0011,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.51,"ci90":[0.46,0.58],"n_valid_runs":1450,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":3617.0,"billing":"API","rank":12,"note":"49 of 2352 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"GPT-6.1 Terra","harness":"pi","lanes":3502,"lane_unit":"ledger","hours":56.5,"gpu_runs":274,"wins":0,"rate":0.0,"rlo":0.0,"rhi":5.254402321946659,"p":0.0,"lo":0.0,"hi":0.0008,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.74,"ci90":[0.69,0.79],"n_valid_runs":274,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":1250.0,"billing":"API","rank":13,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"Claude Sonnet 5.5","harness":"Claude Managed Agents","lanes":1490,"lane_unit":"ledger","hours":207.6,"gpu_runs":996,"wins":0,"rate":0.0,"rlo":0.0,"rhi":1.4300276068881803,"p":0.0,"lo":0.0,"hi":0.0018,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.6,"ci90":[0.56,0.67],"n_valid_runs":511,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":1664.0,"billing":"API","rank":14,"note":"active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"GPT-6.1 Terra","harness":"OpenAI Agents","lanes":1538,"lane_unit":"ledger","hours":100.4,"gpu_runs":460,"wins":0,"rate":0.0,"rlo":0.0,"rhi":2.956909673207034,"p":0.0,"lo":0.0,"hi":0.0018,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.62,"ci90":[0.58,0.68],"n_valid_runs":459,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":219.0,"billing":"API","rank":15,"note":"active only from 10-28, when the problems were heavily mined (records rare for every family)"}],"writers":[{"who":"Claude Opus 5.5 via Claude Code","records":123,"share":0.583,"gain":430.77},{"who":"GPT-6 Astra via Codex CLI","records":25,"share":0.11800000000000001,"gain":46.85},{"who":"Gemini 4 Argon via pi","records":19,"share":0.09,"gain":11.83},{"who":"Unknown writer","records":12,"share":0.056,"gain":16.18},{"who":"GPT-6.1 Sol via pi","records":9,"share":0.043,"gain":6.74},{"who":"Gemini 4 Argon via Antigravity (agy)","records":8,"share":0.038,"gain":3.98},{"who":"Kimi K3.1 Pro via Kimi Code CLI","records":7,"share":0.033,"gain":33.66},{"who":"GPT-6.1 Sol via Codex CLI","records":5,"share":0.024,"gain":43.6},{"who":"Claude Sonnet 5.5 via pi","records":1,"share":0.005,"gain":0.48},{"who":"Claude Sonnet 5.5 via pi (API writer tier)","records":1,"share":0.005,"gain":0.61},{"who":"Claude Opus 5.5 via pi","records":1,"share":0.005,"gain":0.28}],"mix":{"days":["2026-10-22","2026-10-23","2026-10-24","2026-10-25","2026-10-26","2026-10-27","2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-04","2026-11-05"],"series":[{"label":"Claude Opus 5.5","values":[52.0,170.0,60.0,11.0,47.0,277.0,249.0,316.0,419.0,172.0,284.0,79.0,34.0,65.0,0.0]},{"label":"Claude Sonnet 5.5","values":[0,0,0,0,0,20.0,580.0,664.0,53.0,126.0,118.0,29.0,0.0,0,0.0]},{"label":"GPT-6.1 Terra","values":[0,0,0,0,4.0,52.0,349.0,53.0,15.0,2.0,0,0,0,0,0]},{"label":"GPT-6.1 Sol","values":[0,1.0,280.0,45.0,41.0,85.0,0.0,0,0,0,0,0,0,0,0]},{"label":"Kimi K3","values":[0,0,0,0,0,0,117.0,175.0,46.0,59.0,22.0,3.0,0.0,0,0]},{"label":"Gemini 4 Argon","values":[0,174.0,125.0,41.0,7.0,0,9.0,0,0,0,0,0,0,0,0]},{"label":"Other and unattributed","values":[117.0,24.0,37.0,13.0,118.0,245.0,255.0,78.0,88.0,71.0,142.0,105.0,24.0,2.0,0.0]}]},"best_family":{"name":"pi frontier (main-briefed) | Gemini 4 Argon | writer+runner","odds":4.34,"ci":[2.37,7.17],"n":121,"records":13},"baseline_family":"Claude Code seats | Claude Opus 5.5 | writer+runner","writer_measured":4,"writer_believed":205,"caption":"24-hour records per 100 lane-hours (summed session time), with a 90% Poisson interval, sorted by the lower end of the interval. Claude Code, Codex and Antigravity sessions are rebuilt from GPU activity, which favours them against pi lanes. Associations across 14 days of mixed roles and phases, not a controlled test.","brief":{"auc":0.913,"n":4277,"auc_1029":0.877,"raise":["keys","attn sweeper","sweeper","warps","case are","blocked lane","clusters","unblock"],"lower":["the row","row file","decode order","dec","over the","mean"]},"ab":{"lanes":[11,13],"beat":[0,0],"verdict":"INCONCLUSIVE (< 12 lanes per arm)"},"policy":[{"t":"2026-10-22T17:56-07:00","change":"Codex islands launched: 5 GPT-6 Astra agents, one per problem","quote":null,"src":"git 05e714d3a92 (bench/tools/codex_island.py)","tag":"copied from markdown/config/git"},{"t":"2026-10-23T00:25-07:00","change":"Codex islands OFF (no Codex compute left)","quote":"- **Codex islands are OFF** (owner, 10-23 07:25 UTC: \"i dont have anymore codex compute\"; account limit 07:18 UTC). The","src":"AGENTS.md Current reality (owner 10-23 07:25 UTC)","tag":"copied from markdown/config/git"},{"t":"2026-10-23T02:00-07:00","change":"Coding-agent compute on SUBSCRIPTIONS only; pi + Gemini 4 Argon allowed as the API exception","quote":"- Coding-agent compute runs on SUBSCRIPTIONS only, never API-billed usage (owner, 10-23 06:53 UTC): Codex on the","src":"AGENTS.md Hard rules (owner 10-23 06:53 UTC)","tag":"copied from markdown/config/git"},{"t":"2026-10-23T14:55-07:00","change":"Pi islands on Gemini 4 Argon replace Codex research; 15 islands live","quote":"- **Pi islands replace the Codex islands** (owner, 10-23 21:55 UTC: \"replace all codex research with Pi and Argon harness\";","src":"AGENTS.md (owner 10-23 21:55 UTC)","tag":"copied from markdown/config/git"},{"t":"2026-10-23T16:20-07:00","change":"Gemini 3.7 Flash fallback banned (only Argon or DeepSeek)","quote":"auto switcher\", \"no open router\", \"compaction yes\"; 23:20: \"never use gemini flash ... only ARGON or DEEPSEEK\").","src":"AGENTS.md (owner 10-23 23:20 UTC)","tag":"copied from markdown/config/git"},{"t":"2026-10-23T22:50-07:00","change":"GPT-6.1 Sol direct from OpenAI for cheap pi workflows; no Argon","quote":"- **pi runs use GPT-6.1 Sol direct from OpenAI** (owner 10-24 ~05:50 UTC: \"use gpt 6.1 sol for many cheap workflows ... no argon, we are spending way too much\"; \"not deepseek, just 6.1 sol from OpenAI directly\"; \"keep track of the cost and usage\"): `pi_agent.py launch` (auto) runs `openai/gpt-6.1-sol` (thinking medium); key in ~/.config/okc/llm.env (llm.env wins over the shell env); Argon only whe","src":"AGENTS.md (owner 10-24 ~05:50 UTC)","tag":"copied from markdown/config/git"},{"t":"2026-10-24T12:40-07:00","change":"Fleet mix: Argon first (first lane per case + hypothesis generator), Sol second","quote":"- **Fleet model mix: Argon first** (owner 10-24 12:40 PDT: \"use Gemini Argon more than Sol, Argon is pretty good at novel ideas\"): in the manager fleet the first lane of every case and the hypothesis generator run on Gemini 4 Argon ($6 / 75 min); the second lane per case, the adversary and forks run on Sol. ~10 Argon lanes live (Gemini quota binds above 16).","src":"AGENTS.md (owner 10-24 12:40 PDT)","tag":"copied from markdown/config/git"},{"t":"2026-10-25T00:30-07:00","change":"Codex model knob: GPT-6.1 Sol (Astra weekly-limited to 10-29)","quote":null,"src":"git 7390b29788e (bench/tools/codex_env.py)","tag":"copied from markdown/config/git"},{"t":"2026-10-26T17:08-07:00","change":"Codex founders account ($2,500 credits): GPT-6 Astra default, 12 Astra islands","quote":null,"src":"git 22ef5ccb370 (bench/tools/codex_env.py)","tag":"copied from markdown/config/git"},{"t":"2026-10-27T15:20-07:00","change":"DeepSeek V4.1 Flash default brain for pi launches and manager workers","quote":"- **Coordinator + DeepSeek PI workers (owner 10-27).** `pi_agent.py launch` and manager workers default to DeepSeek V4.1 Flash at high reasoning. The `scientist_coordinator` brief turns verified success patterns into bounded advisory tasks; `scientist_deepseek_worker` executes them and returns artifacts and receipts. Explicit model overrides remain. API launches still require `OKC_ALLOW_API=1`; do","src":"AGENTS.md (owner 10-27)","tag":"copied from markdown/config/git"},{"t":"2026-10-27T19:20-07:00","change":"Provenance gate: code is written only by Opus 5.5; DeepSeek/Haiku/Terra/Argon/Kimi become runners","quote":"- 10-27 7:20 PM PT owner: DeepSeek is the default brain; code is written only by Opus 5.5; enforced by the provenance gate (bench/pi/writer_registry.jsonl).","src":"AGENTS.md; bench/tools/provenance.py docstring","tag":"copied from markdown/config/git"},{"t":"2026-10-28T00:39-07:00","change":"Codex default GPT-6.1 Terra high; pi Haiku 5.5 cloud fleet scaled","quote":null,"src":"git b66c752a595 (bench/tools/codex_env.py)","tag":"copied from markdown/config/git"},{"t":"2026-10-28T09:25-07:00","change":"pi agent mix: Haiku 60%, Terra 20%, Argon 10% runners + Sonnet 5.5 writers (writer tier moved from Opus to Sonnet)","quote":"# THE PI-AGENT MODEL MIX (owner 10-28 ~09:25 PT, lane-scheduler): of every live pi agent, Haiku 5.5 60%, GPT-6.1 Terra","src":"bench/tools/brain.py comment (owner 10-28 ~09:25 PT); bench/tools/pi_scale.py docstring 'Sonnet 5.5 ... (owner 10-28; was Opus 5.5)'","tag":"copied from markdown/config/git"},{"t":"2026-10-28T11:15-07:00","change":"Argon OFF (economics); later 10-29 AGY stopped too","quote":"owner 10-28 11:15 PT: \"let's not use Argon for now ... economically it makes no sense\". Delete this file to allow new Argon runs again.","src":"bench/manager/ARGON_OFF","tag":"copied from markdown/config/git"},{"t":"2026-10-28T20:25-07:00","change":"OpenAI Agents (Terra raw API) OFF after an $800 bill; move to Kimi","quote":"owner 10-28 ~8:25 PM PT: \"$800 bill on 6.1 Terra in raw API ... stop the Terra lanes once they are done and move to Kimi\". Present = oai_worker starts no new OpenAI Agents session; delete to re-enable.","src":"bench/manager/OAI_OFF","tag":"copied from markdown/config/git"},{"t":"2026-10-28T22:40-07:00","change":"GPT-6.1 Terra OFF for pi runs ($513 in 3 h, 2 GPU runs, 0 hits)","quote":"owner 10-28 ~10:40 PM PT: \"let them finish and then STOP 6.1 Terra for now ... write down the learning\".","src":"bench/manager/TERRA_OFF","tag":"copied from markdown/config/git"},{"t":"2026-10-30T12:55-07:00","change":"Kimi K3 OFF: Sonnet 5.5 and Opus instead","quote":"owner 10-30 ~12:55 PT: \"use the Sonnet 5.5 in there actually instead of kimi\" and \"use opus for some, no need to use kimi anymore\".","src":"bench/manager/KIMI_OFF","tag":"copied from markdown/config/git"},{"t":"2026-10-30T16:30-07:00","change":"AWS suspends Bedrock on both accounts; every Bedrock lane paused","quote":"AWS suspended Amazon Bedrock on accounts 432342421250 and 263998146836 (~2026-10-30 23:30Z, 'Error 002: Access to Bedrock models is not allowed for this account'). Every Bedrock lane is paused and nothing re-probes. Re-enable: a probe OK (python3 bench/tools/swarm.py probe) AND the owner deletes this file.","src":"bench/manager/BEDROCK_OFF","tag":"copied from markdown/config/git"},{"t":"2026-10-30T17:35-07:00","change":"Kimi K3 back on River credits ($50k) for pi runner lanes (KIMI_OFF superseded)","quote":"SUPERSEDED for pi runner lanes (main 5d16f8bd, 11-01 12:05 PM PST): owner 10-30 ~5:35 PM PT, \"why is the river ai with 50k$ credits and kimi k3 is not being utilised???\". RIVER.json + 5 bench/manager/arm_policy.json allowances (kimi3r-high, until 2026-11-05T07:59Z, $2,000/day pool) run Kimi K3 on pi lanes; 448 runs on 11-01 were sanctioned. This file still disables the swarm brain kimi and the pau","src":"bench/manager/KIMI_OFF; bench/manager/RIVER.json","tag":"copied from markdown/config/git"},{"t":"2026-10-30T18:15-07:00","change":"OpenRouter grant ($4,999.99): Sonnet 5.5 runners via OpenRouter, Opus 5.5 writers on Anthropic API","quote":null,"src":"bench/manager/BRAIN.json why field (owner 10-30 ~6:15 PM PT)","tag":"copied"},{"t":"2026-10-31T14:15-07:00","change":"Decode strike lanes on pi + Claude Code seats","quote":null,"src":"bench/tools/strike_lanes.py docstring (owner 10-31 ~2:15 PM PT)","tag":"copied"},{"t":"2026-11-01T16:30-07:00","change":"Writer pi lanes moved to OpenRouter after the Anthropic API pool overspent ($624 of $450/day)","quote":null,"src":"bench/manager/BRAIN.json why field (11-01 23:30Z)","tag":"copied"},{"t":"2026-11-03T00:00-08:00","change":"All agents cloud-first (Claude Code cloud sessions)","quote":"- **How every agent works (owner 11-03: \"make sure they are all only cloud agents in there, nothing local really\"): `HOW_AGENTS_WORK.md`.** One rule, same bytes as stormy-mono; it wins over older lines here. New agents start in the cloud. Lanes that still run on the Mac (`claude_agent.py`) are the old form and move as their seats get GitHub connected.","src":"AGENTS.md Hard rules (owner 11-03)","tag":"copied from markdown/config/git"}]},"classes_sets":[{"id":"rows","label":"Scored-row moves, all 14 days","rows":[{"name":"Structural rewrite","id":"STRUCTURAL-REWRITE","n":16,"share":0.306,"median":0.049850000000000005,"desc":"A new kernel design replaces an existing stage (new dataflow, new persistent kernel, wgmma path).","examples":["moe162: grouped W2 K256 + small TMA routing broker (Codex)","Codex fragment v3 (<96) + own W2 wide kernel (>=96)","Codex D1D2: static V6 + split-K pooler/qkv (SM-fill fix)"]},{"name":"Compose","id":"COMPOSE","n":21,"share":0.2,"median":0.01551,"desc":"Stack confirmed or near-miss wins (\"genes\") on the head; includes keeps.py small-win bundles.","examples":["dec8_v4_e_r31_lum: the night bundle of ~19 h of island + audit work (row frozen 19 h because the day's 10 scored runs were spent); #1 decode","v17pl_752_wd: night bundle of Pi-island and Claude-fix work","f20_r2_sm8 (pi Argon frontier lane): night bundle"]},{"name":"Port","id":"PORT","n":10,"share":0.138,"median":0.02932,"desc":"Move a kernel or trick that won elsewhere (sibling shape, sibling problem, or a rival Apache/MIT repo).","examples":["v9: import the MoE problem's agent kernel at T<=192, ours above; cat ffn; CUDA attention","v15: route_pre in CUDA + Codex MoE fragment build","ZhimaoL fmha7 hybrid-exp2 attention port (+5.18% paired) + DQ3 code layout (+0.27%)"]},{"name":"Tweak","id":"TWEAK","n":19,"share":0.13,"median":0.01159,"desc":"A small (<= ~50-line) schedule edit of the incumbent: trigger, stages, warps, prefetch, barrier.","examples":["[sealed until Nov 19]","E5-E7: uniform item loop + KSTG table","p4_fixed: TMA maps per call (state-copy safe)"]},{"name":"From scratch","id":"FROM-SCRATCH","n":2,"share":0.09,"median":0.13661,"desc":"A first version written from nothing (the day-1 root kernels and their first rewrites).","examples":["v3: 17-warp persistent kernel, reducer warp, 4 slots, cp.async","v6: own MoE kernel at T<=4 and T>=96 + Triton PDL + fused ffn"]},{"name":"Route","id":"ROUTE","n":2,"share":0.024,"median":0.033929999999999995,"desc":"Send a shape / T range to an existing faster kernel or config (per-case hybrid dispatch, held-out routes).","examples":["v8 hybrid (own kernels + SGLang per case)","pre-338: 3-class T32 128/64/32 MoE split"]},{"name":"Pin","id":"PIN","n":3,"share":0.02,"median":0.021709999999999997,"desc":"Freeze a config found by exhaustive offline search (cuBLASLt algo/tile/split-K), no run-time heuristic.","examples":["big_m d1: frozen cuBLASLt configs from an offline search (NathenFernandes, MIT)","moe204: static wide flags","big_m d2u: frozen cuBLASLt configs"]}]},{"id":"chain","label":"Chain landings, from Oct 30","rows":[{"name":"Compose","id":"COMPOSE","n":26,"share":0.445,"median":0.00192,"desc":"Stack confirmed or near-miss wins (\"genes\") on the head; includes keeps.py small-win bundles.","examples":["98e6a26d: paired CONFIRMED +1.34% vs cand12, never composed with cand13","RSWG2_HQKS_V1F_LOGCODES","paired b7b96a9f9651e5c6 (okc/muse_attention_decode/claude_astack3/as3_c132_lz.py)"]},{"name":"Tweak","id":"TWEAK","n":35,"share":0.434,"median":0.00184,"desc":"A small (<= ~50-line) schedule edit of the incumbent: trigger, stages, warps, prefetch, barrier.","examples":["64c6be4b CONFIRMED +1.39% vs db3572f0 (box-20261030T231708-821915775372449db0d163a5b6d9f419.py)","[sealed until Nov 19]","b7c7aa45 CONFIRMED +1.03% vs 4a7e6565 (box-20261030T132438-709e4eece7ef41d88d0d297b146faea8.py)"]},{"name":"Port","id":"PORT","n":8,"share":0.10400000000000001,"median":0.00156,"desc":"Move a kernel or trick that won elsewhere (sibling shape, sibling problem, or a rival Apache/MIT repo).","examples":["KMOE_STREAMK_T32 on cand66 with PIN_UNP4 = {} (PIN_UNP4 knocks KMOE out: jc17 with PIN_UNP4 on reads -0.01%); keell B4FIN WGS stream-K MoE at 16<T<=32 (pi-b26-1 kmoe_v1); jexact +0.99% [+0.72,+1.26] vs cand66 @40 (d32s -10.8, d32f -11.2 us); packed --min 200 (942 KB)","F8X c6-attn: attention row f8 tile genes in decode verify-full wg_tile_f8 (fexp Schraudolph e4m3 P codes [ZhimaoL/okc_patch Apache-2.0 via opus-bigslide] + c4-q3 rswg tensor-pipe row sums + claude-attn-sink TU headroom), q2h kept, kill switch DEC_F8X; jexact +0.18% CI [0.05,0.665] vs cand30 (rcpt 20","STRIKE_PORTS dq2h_pab 0ad79dba = cand25 + dq2h (2-term FP8 query, c6 FP8 P.V verify full) + pab (b32 full CTA pairing)"]},{"name":"Route","id":"ROUTE","n":3,"share":0.008,"median":0.0005600000000000001,"desc":"Send a shape / T range to an existing faster kernel or config (per-case hybrid dispatch, held-out routes).","examples":["QO32F d32s o on rsc MW=32 ST6 FL4114 with the FIXPRE fixer (fixpre_land h32_st6 2e3d0fca; FIXPRE d16s block present with FIXPRE=0): jexact_scoped judge router salts 13/17/19/23 +0.077/+0.061/+0.086/+0.091%, all CI low > 0, A/A -0.005/-0.017/+0.014/+0.009; d32s -0.58..-1.23 us","[sealed until Nov 19]","QO32F_FULL: d32f o (T=32 full) also on the cand84 rsc MW=32 ST6 FL4114 route (fixpre_land hf_st6 23876c21; head d32f o was q12d proj_rs12): jexact_scoped judge router 8 salts vs cand84 +0.022/+0.069/+0.047/+0.003/+0.024/+0.065/+0.070/+0.052%, pool +0.043% [+0.013,+0.074]; d32f -0.58..-1.70 us"]},{"name":"Pin","id":"PIN","n":2,"share":0.006,"median":0.00063,"desc":"Freeze a config found by exhaustive offline search (cuBLASLt algo/tile/split-K), no run-time heuristic.","examples":["INSURE_DQ2_V32LT: (1) two-term FP8 query (hi=e4m3(q), lo=e4m3(q-hi), 2 extra k32 QK wgmmas/tile) on verify-sliding b16/b32 (r7s4) = routing insurance (hard draw 9312402/13240 v32s D +0.0002 -> +0.03; v16s/v32s min D 15 seeds 0.027/0.028); (2) with that margin, T=256 qkv as ONE bf16 cuBLASLt GEMM (TN","D34_PIN4 (d34-combine 11-03): bf16 cand14 LT_PINS cfg [66,20,1,0,0,4,35,0,2] on decode's T=256 NN4 unpooler and o (PIN256 override, 222 B, kill switch). jexact_scoped 9 runs vs cand105 judge router, pooled +0.033% [+0.004,+0.062], A/A mean +0.006%; v32s -0.4, v32f -0.4 us; strict w29304/w9312402/w91"]},{"name":"Structural rewrite","id":"STRUCTURAL-REWRITE","n":1,"share":0.003,"median":0.00061,"desc":"A new kernel design replaces an existing stage (new dataflow, new persistent kernel, wgmma path).","examples":["R1F0_JOINT on cand74 (pi-b23-2 joint_v3 / c6-moe-t8 j3c67 genes; coordinator-dec port)"]}]}],"classes":[{"name":"Structural rewrite","id":"STRUCTURAL-REWRITE","n":16,"share":0.306,"median":0.049850000000000005,"desc":"A new kernel design replaces an existing stage (new dataflow, new persistent kernel, wgmma path).","examples":["moe162: grouped W2 K256 + small TMA routing broker (Codex)","Codex fragment v3 (<96) + own W2 wide kernel (>=96)","Codex D1D2: static V6 + split-K pooler/qkv (SM-fill fix)"]},{"name":"Compose","id":"COMPOSE","n":21,"share":0.2,"median":0.01551,"desc":"Stack confirmed or near-miss wins (\"genes\") on the head; includes keeps.py small-win bundles.","examples":["dec8_v4_e_r31_lum: the night bundle of ~19 h of island + audit work (row frozen 19 h because the day's 10 scored runs were spent); #1 decode","v17pl_752_wd: night bundle of Pi-island and Claude-fix work","f20_r2_sm8 (pi Argon frontier lane): night bundle"]},{"name":"Port","id":"PORT","n":10,"share":0.138,"median":0.02932,"desc":"Move a kernel or trick that won elsewhere (sibling shape, sibling problem, or a rival Apache/MIT repo).","examples":["v9: import the MoE problem's agent kernel at T<=192, ours above; cat ffn; CUDA attention","v15: route_pre in CUDA + Codex MoE fragment build","ZhimaoL fmha7 hybrid-exp2 attention port (+5.18% paired) + DQ3 code layout (+0.27%)"]},{"name":"Tweak","id":"TWEAK","n":19,"share":0.13,"median":0.01159,"desc":"A small (<= ~50-line) schedule edit of the incumbent: trigger, stages, warps, prefetch, barrier.","examples":["[sealed until Nov 19]","E5-E7: uniform item loop + KSTG table","p4_fixed: TMA maps per call (state-copy safe)"]},{"name":"From scratch","id":"FROM-SCRATCH","n":2,"share":0.09,"median":0.13661,"desc":"A first version written from nothing (the day-1 root kernels and their first rewrites).","examples":["v3: 17-warp persistent kernel, reducer warp, 4 slots, cp.async","v6: own MoE kernel at T<=4 and T>=96 + Triton PDL + fused ffn"]},{"name":"Route","id":"ROUTE","n":2,"share":0.024,"median":0.033929999999999995,"desc":"Send a shape / T range to an existing faster kernel or config (per-case hybrid dispatch, held-out routes).","examples":["v8 hybrid (own kernels + SGLang per case)","pre-338: 3-class T32 128/64/32 MoE split"]},{"name":"Pin","id":"PIN","n":3,"share":0.02,"median":0.021709999999999997,"desc":"Freeze a config found by exhaustive offline search (cuBLASLt algo/tile/split-K), no run-time heuristic.","examples":["big_m d1: frozen cuBLASLt configs from an offline search (NathenFernandes, MIT)","moe204: static wide flags","big_m d2u: frozen cuBLASLt configs"]}],"waterfall":{"problem":"muse_decode_layer","title":"The decode layer, 32 row moves from 1.168x to 2.451x","steps":[{"t":"2026-10-22T22:56:48Z","x_before":null,"x_after":1.1682,"cls":"from scratch","label":"v1: SGLang path + fused glue + Triton GEMM table"},{"t":"2026-10-23T00:08:58Z","x_before":1.1682,"x_after":1.2598,"cls":"from scratch","label":"v6: own MoE kernel at T<=4 and T>=96 + Triton PDL + fused ffn"},{"t":"2026-10-23T01:08:55Z","x_before":1.2598,"x_after":1.4319,"cls":"port","label":"v9: import the MoE problem's agent kernel at T<=192, ours above; cat ffn; CUDA attention"},{"t":"2026-10-23T02:03:37Z","x_before":1.4319,"x_after":1.4952,"cls":"structural rewrite","label":"v10: wgmma experts at T>192 + attention VM5/VM1"},{"t":"2026-10-23T02:19:41Z","x_before":1.4952,"x_after":1.4953,"cls":"resend","label":"same bytes as a check, re-scored (owner: submit results)"},{"t":"2026-10-23T03:30:28Z","x_before":1.4953,"x_after":1.6325,"cls":"port","label":"v15: route_pre in CUDA + Codex MoE fragment build"},{"t":"2026-10-23T03:50:28Z","x_before":1.6325,"x_after":1.6906,"cls":"structural rewrite","label":"Codex island decode build dec142 (details not in the note)"},{"t":"2026-10-23T04:50:54Z","x_before":1.6906,"x_after":1.7547,"cls":"port","label":"v21: dec142 + MoE problem kernels moe162 (wide, T>=128) + moe173 decoder/LPT"},{"t":"2026-10-24T00:54:59Z","x_before":1.7547,"x_after":2.0821,"cls":"compose","label":"dec8_v4_e_r31_lum: the night bundle of ~19 h of island + audit work (row frozen 19 h because the day's 10 scored runs were spent); #1 decode"},{"t":"2026-10-24T01:06:00Z","x_before":2.0821,"x_after":2.1008,"cls":"tweak","label":"dec8_v3_r31_lum variant check promoted (pi Argon frontier lane)"},{"t":"2026-10-25T00:04:44Z","x_before":2.1008,"x_after":2.1146,"cls":"tweak","label":"dec8_v11 promotion"},{"t":"2026-10-25T06:06:47Z","x_before":2.1146,"x_after":2.162,"cls":"structural rewrite","label":"decode D + own norm kernel nr_D_v1"},{"t":"2026-10-27T02:44:41Z","x_before":2.162,"x_after":2.204,"cls":"compose","label":"owner checkpoint send of the accumulated head (10-26 6:40 PM PT)"},{"t":"2026-10-27T17:29:03Z","x_before":2.204,"x_after":2.2134,"cls":"compose","label":"reset_send of kept small wins (keeps.py)"},{"t":"2026-10-28T22:36:53Z","x_before":2.2134,"x_after":2.2496,"cls":"unattributed","label":"promoter: best judge check +1.37% over the row"},{"t":"2026-10-29T22:10:56Z","x_before":2.2496,"x_after":2.264,"cls":"port","label":"attention problem's dlw T=1 sliding kernel ported into decode; paired +0.40%"},{"t":"2026-10-31T21:18:51Z","x_before":2.264,"x_after":2.2866,"cls":"compose","label":"cand25 = strike harvest stackR + pwq1off (T=1 qkv trigger off)"},{"t":"2026-10-31T23:38:12Z","x_before":2.2866,"x_after":2.3317,"cls":"compose","label":"cand28 = cand22..27 stack (jc1 trigger drop +1.23%, harvests) + PCSE_WG_OFF"},{"t":"2026-11-01T12:37:48Z","x_before":2.3317,"x_after":2.3408,"cls":"compose","label":"cand56: stacked scoped wins + held-out routes"},{"t":"2026-11-02T05:22:05Z","x_before":2.3408,"x_after":2.3499,"cls":"compose","label":"cand77: includes KMOE stream-K port (keell, +0.99%) and R1F0_JOINT (Fork-and-Flush)"},{"t":"2026-11-02T15:50:48Z","x_before":2.3499,"x_after":2.3726,"cls":"compose","label":"cand81 ROWSAFE: held-out OOB crash fix + stacked genes"},{"t":"2026-11-03T21:29:02Z","x_before":2.3726,"x_after":2.3655,"cls":"row lost","label":"encoded-source disqualification (46 runs, 11-03): the row fell back to an older legal run"},{"t":"2026-11-03T21:44:34Z","x_before":2.3655,"x_after":2.3647,"cls":"row lost","label":"encoded-source disqualification (46 runs, 11-03): the row fell back to an older legal run"},{"t":"2026-11-04T00:43:38Z","x_before":2.3647,"x_after":2.2496,"cls":"row lost","label":"encoded-source disqualification (46 runs, 11-03): the row fell back to an older legal run"},{"t":"2026-11-04T01:33:22Z","x_before":2.2496,"x_after":2.3687,"cls":"recovery","label":"first readable-source re-send after the DQ (d34 clean line)"},{"t":"2026-11-04T03:22:28Z","x_before":2.3687,"x_after":2.3702,"cls":"recovery","label":"readable-source re-send"},{"t":"2026-11-04T05:51:49Z","x_before":2.3702,"x_after":2.3765,"cls":"unattributed","label":"clean_s7 (d34_clean profiles); content not in the run note"},{"t":"2026-11-04T06:45:23Z","x_before":2.3765,"x_after":2.3967,"cls":"unattributed","label":"clean_s8 (d34_clean profiles); content not in the run note"},{"t":"2026-11-04T13:28:24Z","x_before":2.3967,"x_after":2.4019,"cls":"tweak","label":"b2z arm (d34_fresh)"},{"t":"2026-11-04T20:02:57Z","x_before":2.4019,"x_after":2.4034,"cls":"compose","label":"clean_s28 = r22 + c_pks + x2 + cm16s + qpc6g + pu6m (paired +0.46..+0.56% vs cand119)"},{"t":"2026-11-05T00:11:49Z","x_before":2.4034,"x_after":2.4234,"cls":"tweak","label":"clean_s31: T=8 cat ring 7, proj_q o ring 3, OKC_MFQ 39"},{"t":"2026-11-05T03:16:27Z","x_before":2.4234,"x_after":2.4514,"cls":"compose","label":"cand125 clean_s38: f1c, yl8, yd8 (one-CTA final pass, 16 B/lane), keep_s2, h2w12; paired +0.19%"}],"by_class":{"PORT":25.94,"COMPOSE":25.35,"STRUCTURAL-REWRITE":10.04,"FROM-SCRATCH":7.55,"RECOVERY":5.22,"UNATTRIBUTED":2.73,"TWEAK":2.59,"RESEND":0.01,"ROW-LOST":-5.32}},"milestones":[{"t":"2026-10-22T22:12:00Z","kind":"tool","label":"okc_run.py: one sender for judge checks and scored runs","detail":"The send tool and board.py were the first tools committed.","date_only":false},{"t":"2026-10-22T23:17:00Z","kind":"rank","label":"prefill board best rank #3","detail":"first read at rank 3; held until 2026-10-24T04:54:11Z","date_only":false},{"t":"2026-10-23T00:36:00Z","kind":"tool","label":"sci.py ledger: every idea pre-registered with a prediction","detail":"First idea row (pre-001) logged with gain and p; 14,667 ideas and 13,638 trials by 11-05.","date_only":false},{"t":"2026-10-23T00:47:00Z","kind":"win","label":"moe row +19.5% faster (FROM-SCRATCH)","detail":"1.2712 -> 1.5188: v3: 17-warp persistent kernel, reducer warp, 4 slots, cp.async","date_only":false},{"t":"2026-10-23T03:55:00Z","kind":"incident","label":"All 5 MoE judge FAILs traced to state copies","detail":"The judge runs kernel_fn on tensor-cloned copies; kernels that baked raw addresses at prepare read stale buffers (rel L2 1.414). Gate: statecopy.py.","date_only":false},{"t":"2026-10-23T03:59:00Z","kind":"win","label":"moe row +17.2% faster (STRUCTURAL-REWRITE)","detail":"1.6724 -> 1.9607: moe162: grouped W2 K256 + small TMA routing broker (Codex)","date_only":false},{"t":"2026-10-23T04:02:00Z","kind":"rank","label":"attention board best rank #2","detail":"first read at rank 2; held until 2026-10-23T10:39:58Z","date_only":false},{"t":"2026-10-23T04:50:00Z","kind":"incident","label":"Decode spent 10 of 10 scored runs; row froze ~20 h","detail":"The day's scored runs were gone by 9:50 PM PDT 10-22; the decode row stood at 1.7547 until 10-24 00:54Z while checks rose; rank 2 -> 8+. Gate: per-day pacing in okc_run.py.","date_only":false},{"t":"2026-10-23T05:11:00Z","kind":"incident","label":"Branch merge reset the shared checkout; ledger writes lost","detail":"A worktree-branch merge with git's hard reset overwrote events.jsonl and island files; ~10 ledger writes lost. Rule: only git add <paths>, commit, push_sync.","date_only":false},{"t":"2026-10-23T07:25:00Z","kind":"owner","label":"Codex islands off: \"no more codex compute\"","detail":"Codex account limit at 07:18Z; the island supervisor was stopped; Claude owner agents carry every front.","date_only":false},{"t":"2026-10-23T09:00:00Z","kind":"owner","label":"pi harness with Gemini 4 Argon allowed","detail":"Owner asked for sub-agents on the pi harness with Argon, with a $ cap per run.","date_only":false},{"t":"2026-10-23T09:07:00Z","kind":"rank","label":"bf16 board best rank #2","detail":"first read at rank 2; held until 2026-10-23T10:39:58Z","date_only":false},{"t":"2026-10-23T21:55:00Z","kind":"owner","label":"Pi islands replace Codex islands; \"launch many many agents\"","detail":"15 Pi/Argon islands live from 22:00Z (dec 5, MoE 3, attn 3, pre 2, bf16 2).","date_only":false},{"t":"2026-10-23T22:45:00Z","kind":"owner","label":"\"$2k on Gemini Argon\": brain auto-switcher and context compaction","detail":"Owner cut spend: DeepSeek for clear briefs, Argon for open ones, compaction at ~150k tokens.","date_only":false},{"t":"2026-10-24T00:16:00Z","kind":"owner","label":"One sender, one send per problem at a time","detail":"After two main sessions both acted on one hand-off (N3641), okc_run.py holds a per-problem lock.","date_only":false},{"t":"2026-10-24T00:20:00Z","kind":"rank","label":"#1 on the decode board (2.0821x, then 2.1008x)","detail":"Decode #1 from 00:20Z to at least 07:06Z; MoE #1 from 01:59Z. Overall #1 is a proxy: our overall rank matched our decode rank in every snapshot that recorded both (none in this window).","date_only":false},{"t":"2026-10-24T00:54:00Z","kind":"win","label":"decode row +18.7% faster (COMPOSE)","detail":"1.7547 -> 2.0821: dec8_v4_e_r31_lum: the night bundle of ~19 h of island + audit work (row frozen 19 h because the day's 10 scored runs were spent); #1 decode","date_only":false},{"t":"2026-10-24T01:59:00Z","kind":"rank","label":"moe board best rank #1","detail":"first read at rank 1; held until 2026-10-24T07:06:00Z","date_only":false},{"t":"2026-10-24T04:25:00Z","kind":"incident","label":"Google monthly Argon cap hit; every Argon lane died","detail":"Spend was tracked per run and per day, never against the provider's monthly cap.","date_only":false},{"t":"2026-10-24T17:57:00Z","kind":"rank","label":"First read at #3 after an 11-hour gap in board reads","detail":"accountforgoofystuff and colinlikescode passed 2.1008x; overall #3 at 19:15Z.","date_only":false},{"t":"2026-10-24T19:40:00Z","kind":"owner","label":"Fleet model mix: Argon first","detail":"\"Argon is pretty good at novel ideas\": first lane per case on Argon, second on Sol.","date_only":false},{"t":"2026-10-25T07:44:00Z","kind":"tool","label":"growth.jsonl new-best log stops","detail":"380 rows from 2026-10-22T22:11:25Z; successor is the chain ledger","date_only":false},{"t":"2026-10-26T12:00:00Z","kind":"owner","label":"Phase change: below 1% now decides places","detail":"+0.25% on decode = 2 places; every candidate that passes strict checks is raced, small wins stacked (PROTOCOL rule 13).","date_only":true},{"t":"2026-10-26T23:23:00Z","kind":"tool","label":"race.py + keeps.py: race and stack small wins","detail":"P(better) per candidate; gpu_fill reruns until CONFIRMED or LOSING; keeps.py bundles kept small wins.","date_only":false},{"t":"2026-10-27T09:55:00Z","kind":"tool","label":"paired_confirm.py: same-box paired A/B decides","detail":"First CONFIRMED paired deltas carried into scored sends (MoE +1.31% 10-27).","date_only":false},{"t":"2026-10-27T16:20:00Z","kind":"tool","label":"judge.py: median of 5 overhead reads after glitch reads","detail":"7 of 2,997 receipts read 20-75 us overhead and faked 4.42x decode / 3.36x MoE; now median of 5, > 10 us = INFRA.","date_only":false},{"t":"2026-10-27T19:13:00Z","kind":"win","label":"Rival-code ports move 3 rows in one afternoon","detail":"vjeux prefill port 1.737 -> 1.799, attention per-case hybrid 2.597 -> 2.620, bf16 vjeux + frozen cuBLASLt 1.466 -> 1.5375.","date_only":false},{"t":"2026-10-27T22:17:00Z","kind":"tool","label":"SLEEP cycle: on-policy audits of every lane","detail":"sleep.py rebuilds a pruned core every 3 h and logs on-policy violations into MISTAKES.md.","date_only":false},{"t":"2026-10-28T02:20:00Z","kind":"owner","label":"DeepSeek default brain; code written only by Opus 5.5","detail":"Provenance gate on writers (later the writer tier ran Sonnet 5.5: 2,046 of 3,862 files).","date_only":false},{"t":"2026-10-28T23:15:00Z","kind":"incident","label":"Three git worktrees (67 GB) filled the Mac disk","detail":"The confirm loop could not write results: 380 plans aborted NO_RESULT, 43 \"No space left\" errors; nothing shipped for ~16 h.","date_only":false},{"t":"2026-10-29T12:00:00Z","kind":"owner","label":"Post-mortem: the prizes are on the serving board","detail":"Read on day 8 of 15: kernel boards pay nothing; 1.4% of claimed lane gain reached the board; 54% of paired verdicts aborted.","date_only":true},{"t":"2026-10-30T05:30:00Z","kind":"owner","label":"Approximation allowed inside the error bound","detail":"\"push the limits\": approximate exp2 / PWL etc. allowed when the check passes with margin and every key is attended.","date_only":false},{"t":"2026-10-30T22:37:00Z","kind":"tool","label":"chain.py: single-writer candidate chain per problem","detail":"All scored rows imported as cand0..N; from here every landing is a paired diff on the head.","date_only":false},{"t":"2026-10-30T23:30:00Z","kind":"incident","label":"Bedrock suspended: decode on 1 of 33 lanes for ~19 h","detail":"All 12 decode lanes sat on one provider; rank fell #19 -> #44.","date_only":false},{"t":"2026-10-31T20:18:00Z","kind":"tool","label":"jexact: judge-exact cold-L2 replay","detail":"First jexact landing: dropping a T=1 PDL trigger read +1.23% (cand22) that warm-L2 clones hid.","date_only":false},{"t":"2026-10-31T22:30:00Z","kind":"owner","label":"Systems thinking: the decode push drawn as a system","detail":"Gap to #1 closing 0.3-0.46%/day; +5.4..+6.1% projected at the close; marginal gain per landing falling +1.08 -> +0.12%.","date_only":false},{"t":"2026-10-31T23:38:00Z","kind":"win","label":"Decode 2.264 -> 2.3317 after a 48 h flat","detail":"jexact + strike harvest stacked 14 genes (cand19-28).","date_only":false},{"t":"2026-11-01T04:57:00Z","kind":"tool","label":"jexact_scoped + route_identity: +-0.02% verdicts","detail":"Bit-identical untouched cases zero their variance; CI +-0.35% -> +-0.02%.","date_only":false},{"t":"2026-11-01T06:13:00Z","kind":"tool","label":"Held-out landings (heldout_land.py)","detail":"Routes for held-out shapes land without touching scored routes (route_identity 12/12).","date_only":false},{"t":"2026-11-01T19:43:00Z","kind":"tool","label":"Fork-and-Flush on for structural lanes","detail":"K=4 forks from the leader's files every 4 runs, 3 flushed; ON for d32s and v32s_norms only.","date_only":false},{"t":"2026-11-01T19:45:00Z","kind":"owner","label":"SPEC v3: incumbent-first search","detail":"\"hit rate is pretty low\": candidates start as <= 50-line diffs of the head; >= 70% of lane-hours on ROUTE/PORT/COMPOSE/PIN/TWEAK.","date_only":false},{"t":"2026-11-01T23:40:00Z","kind":"owner","label":"Two sources of truth: internal decides, external checks","detail":"The board row is a lucky max (+1.7% over its bytes' median); a revert needs two low reads.","date_only":false},{"t":"2026-11-02T16:00:00Z","kind":"owner","label":"Organizers: readable source only (not read until 11-03)","detail":"Rule of 11-02 08:00 PT; no tool watched llms.txt \"Latest updates\"; lanes kept packing for 37 h.","date_only":false},{"t":"2026-11-02T17:20:00Z","kind":"incident","label":"Writer seats at session / weekly limits; pipeline stalled","detail":"claude6 session limit + claude3/4/5 weekly limits: 37 writer tasks waited up to 6.5 h; pi runners idled (16/24 live).","date_only":false},{"t":"2026-11-03T12:00:00Z","kind":"owner","label":"Cloud first: all agents in the cloud","detail":"\"make sure they are all only cloud agents\"; tether.py engine; Mac lanes are the old form.","date_only":true},{"t":"2026-11-03T21:29:00Z","kind":"incident","label":"46 runs disqualified for compressed source","detail":"Board rows fell: decode 2.3726 -> 2.2496 (rank 215 at 11-04 00:43Z), MoE 2.5343 -> 2.4292, attention 2.9098 -> 2.8355, bf16 1.5712 -> 1.5375.","date_only":false},{"t":"2026-11-04T02:11:00Z","kind":"incident","label":"BSD pkill -n matched every process with -n","detail":"Three pkill events in one night killed gate runs and ~2 h of paired results; rule: kill one exact PID.","date_only":false},{"t":"2026-11-04T12:00:00Z","kind":"incident","label":"qkv gate race found only by a 14,640-call stress","detail":"Strict x3, statecopy and a 1,440-call stress all passed; gate now 36,000 calls at v1f and d1f.","date_only":true},{"t":"2026-11-05T03:11:00Z","kind":"win","label":"Final decode row 2.4514x (readable source), #22","detail":"cand125 clean_s38; the next head read 2.4161x: files 0.09% apart read 1.5% apart on the judge.","date_only":false},{"t":"2026-11-05T07:59:00Z","kind":"rank","label":"Kernel track closes: decode #25 of 1,300 entries","detail":"Decode 2.4514 vs leader pjmullin3 2.6144 (+6.65%); MoE #122, attention #42, prefill #104, bf16 #313 in the last reads.","date_only":false}],"mistakes":[{"name":"sleep / poll wait loops in agent turns","n":40,"cost":"unknown","gate":"note to c3-evolve-moe; the rule is in bench/manager/CORE.md: mailbox.py catchup --agent <agent> every step","after":"98%"},{"name":"skipped start-of-turn reads (CORE.md / mailbox / first moves)","n":29,"cost":"unknown","gate":"note to c3-dec-moe-verify; the rule is in bench/manager/CORE.md: LOOP: `python3 bench/tools/mailbox.py catchup --agent {agent}` every step","after":"76%"},{"name":"off-brief work","n":18,"cost":"~1 h of box 14 time. 4 MoE misses. wq-16 still not done.","gate":"note to c5-stack-moe; the rule is in bench/manager/CORE.md: Take your task from the brief (wq-16 prefill)","after":"94%"},{"name":"raw writes to shared state / git in background lanes","n":17,"cost":"unknown","gate":"note to unblock-manager; the rule is in bench/manager/CORE.md: Run no git command except git add <explicit paths>, git commit and push_sync.py","after":"94%"},{"name":"git / squash / shared-checkout overwrite","n":16,"cost":"~10 ledger writes, 1 island round's files, 5 agents' time","gate":"mechanism (main-c): a FAILED non-fast-forward merge runs `read-tree --reset -u HEAD` (hard reset of every tracked file) and then `stash apply`; anything written in between is los\u2026","after":"94%"},{"name":"decision on a single read / noise / unrecorded number","n":13,"cost":"1 check","gate":"judge.py checks EVERY replay (N621, N630)","after":"77%"},{"name":"provider quota / spend surprise","n":13,"cost":"~16 min x 30 agents; owner added credits at 04:52 (Q4)","gate":"islands.py: on the CLI's ERROR line it pauses all launches 10 min, probes with the next launch, and sends ONE Telegram ask per outage (ask_human.py)","after":"92%"},{"name":"pre-registration / ledger-id hygiene","n":12,"cost":"1 paired run (box 6) with no prediction","gate":"note to c3-evolve-moe; the rule is in bench/manager/CORE.md: Pre-register each change with a numeric prediction (sci.py add) before you measure","after":"75%"},{"name":"stale or wrong pairing base","n":9,"cost":"unknown","gate":"note to c3-evolve-moe; the rule is in bench/manager/CORE.md: Pair against the current base (PAIRED_BASE / campaign base), never an old one","after":"89%"},{"name":"disk full","n":8,"cost":"claude6 commits held ~13 h; 5 boxes' sync stopped","gate":"refilled from the index (checkout-index -f on 0-byte files with non-empty index blobs), re-marked excluded absents, restored kept ones; sparse.py now prunes first and restores on\u2026","after":"88%"}],"spend":{"days":["2026-10-22","2026-10-23","2026-10-24","2026-10-25","2026-10-26","2026-10-27","2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-04","2026-11-05"],"series":[{"label":"GPU rental","values":[5.8,1711.11,447.61,767.54,271.82,2866.94,4366.73,2753.69,2972.09,985.24,2342.43,2164.45,2512.28,1608.28,388.56]},{"label":"API-billed model calls","values":[0,505.01,2120.15,2220.28,0,0,1739.38,4456.38,159.06,1170.92,1586.19,1337.06,754.06,40.26,0]}]},"unit":[{"v":"$27,892","l":"cost of the work: GPU plus API models, measured waste removed"},{"v":"$0.48","l":"per GPU run"},{"v":"$133","l":"per 24-hour record, averaged over 14 days"},{"v":"62","l":"pi-harness lanes running at once, at the peak"}],"periods":[{"id":"P1","from":"2026-10-22","to":"2026-10-24","label":"opening: Codex islands, Gemini Argon / GPT-6.1 Sol pi lanes, first Modal boxes","usd":4789.68,"gpu_runs":2082,"records":107,"per_record":44.76336448598131,"per_run":2.301},{"id":"P2","from":"2026-10-25","to":"2026-10-27","label":"slow-down night 10-25, paid API paused, $25k Modal grant from 10-26","usd":6126.58,"gpu_runs":13173,"records":49,"per_record":125.03224489795919,"per_run":0.465},{"id":"P3","from":"2026-10-28","to":"2026-10-31","label":"API swarm (Anthropic, Kimi, Terra), 16-25 boxes, Daytona, chain from 10-30","usd":18603.489999999998,"gpu_runs":36423,"records":31,"per_record":600.1125806451612,"per_run":0.511},{"id":"P4","from":"2026-11-01","to":"2026-11-04","label":"decode-only focus, OpenRouter Opus writers, serverless fn, close","usd":12345.01,"gpu_runs":6774,"records":22,"per_record":561.1368181818182,"per_run":1.822}],"records_summary":{"runs_considered":57994,"records":209,"gain_lt_1pct":130,"gain_lt_05pct":85,"with_sha":94,"reruns":76,"new_bytes":18,"late_with_sha":72,"late_reruns":62,"by_kind":{"check":19,"scored":7,"modal":36,"box":147},"by_period":[{"id":"P1","from":"2026-10-22","to":"2026-10-24","records":107,"new_bytes":0,"reruns":0,"no_sha":107},{"id":"P2","from":"2026-10-25","to":"2026-10-27","records":49,"new_bytes":12,"reruns":29,"no_sha":8},{"id":"P3","from":"2026-10-28","to":"2026-10-31","records":31,"new_bytes":2,"reruns":29,"no_sha":0},{"id":"P4","from":"2026-11-01","to":"2026-11-04","records":22,"new_bytes":4,"reruns":18,"no_sha":0}],"rule":"valid all-shape run > +0.2% over the best valid run of the previous 24 h on its problem; kinds box/modal/check/scored/episode pooled (bench/tools/success_model.py)"},"decode_top":{"first":"2026-10-24T00:20:23Z","last":"2026-10-24T07:06:00Z","observed_h":6.8,"next_read":"2026-10-24T17:57:27Z","next_rank":3,"upper_h":17.6,"reads_at_1":38},"econ":{"usd_total":41864.77,"usd_gpu":25776.02,"usd_llm":16088.75,"usd_sub_equiv":29204.06,"per_run":0.716,"per_record":200.3099043062201,"per_landing":329.644,"busy_share":0.266,"util":17.8,"peak_lanes":463,"peak_claude":104,"commits_raw":36127,"commits":8021,"rented_gpu_hours":3329.2,"outages":[{"what":"Codex account usage limit: islands, moonshot and research agents died at start","from":"2026-10-23T04:36Z","hours":0.27,"scale":"~30 agents","src":"bench/MISTAKES.md (10-23 04:36 row)","status":"measured"},{"what":"Mac disk full: three 22 GB worktrees (67 GB) on a 78 GB .git; confirm loop could not write, 380 plans aborted NO_RESULT, nothing shipped","from":"2026-10-28T23:15Z","hours":16,"src":"AGENTS.md ('Mac disk (10-29)' section)","status":"measured (duration as written)"},{"what":"hourly review (HOURLY.md) frozen by one union-merged JSONL line","from":"2026-10-28T19:35Z","hours":26,"src":"bench/MISTAKES.md (10-29 2:15 PM PT row)","status":"measured"},{"what":"decode writer pipeline stalled: claude6 session limit + claude3/4/5 weekly limits; 37 writer tasks waited up to 6.5 h, pi runners idle (16/24 live)","from":"2026-11-02T10:50Z","hours":6.5,"src":"bench/MISTAKES.md (11-02 17:20 UTC row)","status":"measured"},{"what":"landing loop stopped on 4 problems, nothing alarmed: 13 candidates waited 3.6-25.4 h unjudged","from":"2026-10-31","hours":25.4,"src":"bench/MISTAKES.md (2026-10-31 row)","status":"measured (max wait)"},{"what":"Mac git packs grew to 117 GB, disk 98% full","from":"2026-11-02T15:04Z","hours":0.8,"src":"bench/MISTAKES.md (11-02 19:50 UTC row)","status":"measured"},{"what":"Claude 'main' seat: zero messages on Pacific days 10-25..10-28 and 11-03..11-04 with rate-limit errors logged (weekly limit)","from":"2026-10-25","hours":null,"src":"bench/manager/USAGE.json claude.main.days","status":"measured (days), duration in hours not derivable"}],"usd_work":27891.67,"usd_gpu_work":18615.92,"usd_llm_work":9275.75,"per_run_work":0.4758209082534375,"per_record_work":133.45296650717702},"decode_pct":{"dq_minutes":50,"series":[["2026-10-22T22:56:48Z",81.818,27,33],["2026-10-22T23:21:48Z",71.795,28,39],["2026-10-22T23:28:21Z",70.732,29,41],["2026-10-22T23:31:50Z",73.81,31,42],["2026-10-22T23:34:58Z",75.0,33,44],["2026-10-22T23:42:58Z",75.556,34,45],["2026-10-22T23:59:27Z",73.469,36,49],["2026-10-23T00:08:58Z",55.102,27,49],["2026-10-23T00:47:24Z",51.786,29,56],["2026-10-23T00:54:52Z",53.448,31,58],["2026-10-23T01:07:43Z",52.542,31,59],["2026-10-23T01:17:51Z",16.667,10,60],["2026-10-23T01:25:26Z",16.393,10,61],["2026-10-23T01:54:26Z",15.873,10,63],["2026-10-23T02:06:27Z",9.375,6,64],["2026-10-23T02:10:55Z",9.231,6,65],["2026-10-23T02:20:58Z",9.091,6,66],["2026-10-23T03:31:30Z",6.024,5,83],["2026-10-23T03:48:58Z",5.682,5,88],["2026-10-23T03:53:58Z",2.247,2,89],["2026-10-23T04:01:07Z",2.174,2,92],["2026-10-23T04:08:58Z",2.128,2,94],["2026-10-23T04:15:34Z",3.125,3,96],["2026-10-23T04:19:31Z",3.061,3,98],["2026-10-23T04:33:58Z",3.0,3,100],["2026-10-23T04:37:11Z",2.941,3,102],["2026-10-23T04:43:58Z",2.857,3,105],["2026-10-23T04:50:54Z",1.887,2,106],["2026-10-23T05:26:38Z",1.852,2,108],["2026-10-23T05:31:07Z",1.818,2,110],["2026-10-23T05:37:58Z",1.786,2,112],["2026-10-23T05:53:58Z",1.754,2,114],["2026-10-23T06:07:01Z",1.709,2,117],["2026-10-23T06:27:55Z",1.667,2,120],["2026-10-23T06:38:24Z",1.639,2,122],["2026-10-23T06:47:58Z",1.613,2,124],["2026-10-23T07:00:01Z",1.575,2,127],["2026-10-23T07:05:38Z",1.55,2,129],["2026-10-23T07:11:58Z",1.515,2,132],["2026-10-23T07:15:38Z",1.481,2,135],["2026-10-23T07:39:59Z",1.418,2,141],["2026-10-23T09:00:36Z",1.29,2,155],["2026-10-23T09:25:47Z",1.274,2,157],["2026-10-23T09:29:54Z",1.258,2,159],["2026-10-23T09:38:20Z",1.242,2,161],["2026-10-23T09:43:59Z",1.22,2,164],["2026-10-23T09:56:44Z",1.19,2,168],["2026-10-23T10:00:55Z",1.176,2,170],["2026-10-23T10:05:23Z",1.163,2,172],["2026-10-23T10:08:59Z",1.143,2,175],["2026-10-23T10:35:02Z",1.087,2,184],["2026-10-23T18:33:54Z",2.92,8,274],["2026-10-23T18:44:50Z",2.909,8,275],["2026-10-23T18:59:22Z",3.237,9,278],["2026-10-23T19:05:59Z",3.203,9,281],["2026-10-23T19:23:03Z",3.169,9,284],["2026-10-23T19:47:00Z",3.147,9,286],["2026-10-23T19:52:08Z",3.484,10,287],["2026-10-23T20:15:08Z",3.46,10,289],["2026-10-23T20:40:32Z",3.425,10,292],["2026-10-23T20:47:13Z",3.356,10,298],["2026-10-23T20:49:54Z",3.333,10,300],["2026-10-23T21:37:48Z",3.226,10,310],["2026-10-23T21:48:02Z",3.205,10,312],["2026-10-23T22:08:00Z",3.145,10,318],["2026-10-23T22:12:41Z",3.448,11,319],["2026-10-23T22:21:22Z",3.738,12,321],["2026-10-23T22:25:49Z",3.715,12,323],["2026-10-23T22:38:40Z",3.681,12,326],["2026-10-23T22:43:40Z",3.659,12,328],["2026-10-23T22:52:03Z",3.636,12,330],["2026-10-23T23:08:51Z",3.614,12,332],["2026-10-23T23:43:42Z",3.835,13,339],["2026-10-24T00:16:02Z",3.736,13,348],["2026-10-24T00:35:14Z",0.287,1,349],["2026-10-24T00:54:59Z",0.285,1,351],["2026-10-24T01:09:17Z",0.284,1,352],["2026-10-24T01:59:34Z",0.274,1,365],["2026-10-24T02:19:32Z",0.272,1,368],["2026-10-24T02:28:00Z",0.27,1,371],["2026-10-24T02:32:00Z",0.268,1,373],["2026-10-24T02:54:11Z",0.267,1,375],["2026-10-24T03:01:39Z",0.265,1,377],["2026-10-24T03:31:53Z",0.262,1,382],["2026-10-24T03:59:08Z",0.258,1,387],["2026-10-24T04:14:00Z",0.257,1,389],["2026-10-24T04:48:35Z",0.256,1,391],["2026-10-24T05:46:08Z",0.249,1,401],["2026-10-24T05:55:16Z",0.248,1,403],["2026-10-24T05:58:10Z",0.247,1,405],["2026-10-24T06:32:19Z",0.245,1,408],["2026-10-24T06:54:00Z",0.244,1,410],["2026-10-24T07:06:00Z",0.243,1,412],["2026-10-24T18:14:09Z",0.574,3,523],["2026-10-24T18:22:12Z",0.571,3,525],["2026-10-24T18:34:50Z",0.569,3,527],["2026-10-24T18:38:59Z",0.567,3,529],["2026-10-24T18:46:44Z",0.564,3,532],["2026-10-24T19:24:37Z",0.558,3,538],["2026-10-24T19:45:41Z",0.556,3,540],["2026-10-24T19:53:17Z",0.554,3,542],["2026-10-24T20:04:35Z",0.551,3,544],["2026-10-24T20:42:13Z",0.733,4,546],["2026-10-24T21:19:17Z",0.723,4,553],["2026-10-24T21:24:51Z",0.721,4,555],["2026-10-24T21:30:51Z",0.718,4,557],["2026-10-24T21:37:38Z",0.714,4,560],["2026-10-24T22:03:39Z",0.712,4,562],["2026-10-24T22:26:05Z",0.708,4,565],["2026-10-24T22:57:39Z",0.703,4,569],["2026-10-24T23:09:11Z",0.701,4,571],["2026-10-24T23:30:27Z",0.698,4,573],["2026-10-24T23:36:48Z",0.696,4,575],["2026-10-24T23:43:50Z",0.693,4,577],["2026-10-25T00:04:44Z",0.69,4,580],["2026-10-25T00:18:39Z",1.031,6,582],["2026-10-25T00:26:46Z",1.027,6,584],["2026-10-25T01:04:09Z",1.024,6,586],["2026-10-25T02:09:00Z",1.02,6,588],["2026-10-25T02:17:36Z",1.02,6,588],["2026-10-25T02:33:21Z",1.184,7,591],["2026-10-25T02:54:15Z",1.18,7,593],["2026-10-25T03:25:32Z",1.173,7,597],["2026-10-25T03:49:36Z",1.169,7,599],["2026-10-25T04:55:53Z",1.157,7,605],["2026-10-25T05:46:39Z",1.314,8,609],["2026-10-25T06:08:56Z",0.82,5,610],["2026-10-25T06:19:34Z",0.814,5,614],["2026-10-25T06:34:36Z",0.812,5,616],["2026-10-25T07:16:53Z",0.971,6,618],["2026-10-25T07:43:37Z",0.968,6,620],["2026-10-25T07:49:53Z",0.965,6,622],["2026-10-25T07:55:54Z",0.962,6,624],["2026-10-25T08:03:37Z",0.962,6,624],["2026-10-25T08:51:19Z",0.958,6,626],["2026-10-25T19:09:00Z",1.49,10,671],["2026-10-25T19:47:04Z",1.486,10,673],["2026-10-25T20:27:31Z",1.479,10,676],["2026-10-25T21:15:19Z",1.466,10,682],["2026-10-25T21:24:46Z",1.466,10,682],["2026-10-25T22:15:40Z",1.456,10,687],["2026-10-25T22:29:38Z",1.594,11,690],["2026-10-25T22:43:39Z",1.59,11,692],["2026-10-25T23:15:39Z",1.585,11,694],["2026-10-25T23:43:48Z",1.58,11,696],["2026-10-26T20:03:17Z",2.008,15,747],["2026-10-26T20:36:42Z",2.003,15,749],["2026-10-26T21:11:08Z",1.997,15,751],["2026-10-26T21:59:14Z",1.992,15,753],["2026-10-26T23:18:10Z",2.122,16,754],["2026-10-27T01:26:50Z",2.114,16,757],["2026-10-27T02:44:41Z",1.449,11,759],["2026-10-27T03:49:42Z",1.444,11,762],["2026-10-27T05:04:53Z",1.44,11,764],["2026-10-27T05:37:42Z",1.567,12,766],["2026-10-27T06:37:14Z",1.562,12,768],["2026-10-27T07:16:01Z",1.558,12,770],["2026-10-27T07:37:39Z",1.552,12,773],["2026-10-27T09:09:10Z",1.548,12,775],["2026-10-27T09:41:27Z",1.542,12,778],["2026-10-27T10:17:40Z",1.538,12,780],["2026-10-27T10:36:00Z",1.535,12,782],["2026-10-27T11:09:55Z",1.66,13,783],["2026-10-27T11:42:21Z",1.654,13,786],["2026-10-27T12:27:05Z",1.65,13,788],["2026-10-27T12:45:16Z",1.774,14,789],["2026-10-27T13:00:48Z",1.77,14,791],["2026-10-27T13:36:01Z",1.77,14,791],["2026-10-27T14:18:23Z",1.765,14,793],["2026-10-27T14:50:40Z",1.761,14,795],["2026-10-27T16:15:32Z",1.757,14,797],["2026-10-27T16:50:37Z",1.752,14,799],["2026-10-27T17:29:03Z",1.625,13,800],["2026-10-27T17:48:36Z",1.621,13,802],["2026-10-27T18:05:39Z",1.743,14,803],["2026-10-27T18:55:38Z",1.739,14,805],["2026-10-27T19:10:13Z",1.735,14,807],["2026-10-27T19:30:43Z",1.731,14,809],["2026-10-27T19:49:02Z",1.728,14,810],["2026-10-27T20:09:07Z",1.847,15,812],["2026-10-27T20:37:32Z",1.843,15,814],["2026-10-27T20:51:55Z",1.838,15,816],["2026-10-27T21:22:09Z",1.834,15,818],["2026-10-27T22:26:07Z",1.829,15,820],["2026-10-27T23:01:55Z",1.946,16,822],["2026-10-28T01:54:38Z",2.066,17,823],["2026-10-28T03:06:55Z",2.309,19,823],["2026-10-28T03:20:45Z",2.303,19,825],["2026-10-28T04:10:01Z",2.297,19,827],["2026-10-28T04:28:22Z",2.415,20,828],["2026-10-28T04:54:27Z",2.533,21,829],["2026-10-28T05:16:51Z",2.647,22,831],["2026-10-28T05:38:34Z",2.641,22,833],["2026-10-28T06:04:06Z",2.635,22,835],["2026-10-28T06:39:29Z",2.751,23,836],["2026-10-28T07:01:51Z",2.741,23,839],["2026-10-28T07:31:17Z",2.735,23,841],["2026-10-28T08:21:05Z",2.728,23,843],["2026-10-28T08:38:26Z",2.722,23,845],["2026-10-28T08:58:14Z",2.715,23,847],["2026-10-28T10:02:16Z",2.709,23,849],["2026-10-28T10:08:43Z",2.703,23,851],["2026-10-28T10:44:05Z",2.693,23,854],["2026-10-28T12:08:36Z",2.687,23,856],["2026-10-28T12:33:20Z",2.678,23,859],["2026-10-28T13:13:00Z",2.671,23,861],["2026-10-28T13:25:55Z",2.665,23,863],["2026-10-28T14:16:37Z",2.659,23,865],["2026-10-28T15:14:28Z",2.653,23,867],["2026-10-28T15:39:45Z",2.647,23,869],["2026-10-28T16:36:22Z",2.647,23,869],["2026-10-28T18:00:49Z",2.641,23,871],["2026-10-28T19:02:32Z",2.749,24,873],["2026-10-28T19:52:03Z",2.746,24,874],["2026-10-28T20:33:51Z",2.74,24,876],["2026-10-28T21:36:17Z",2.733,24,878],["2026-10-28T22:36:53Z",1.706,15,879],["2026-10-28T22:50:59Z",1.703,15,881],["2026-10-29T00:27:39Z",1.814,16,882],["2026-10-29T01:21:28Z",1.81,16,884],["2026-10-29T02:46:43Z",1.806,16,886],["2026-10-29T03:10:05Z",1.8,16,889],["2026-10-29T05:29:36Z",1.906,17,892],["2026-10-29T05:54:18Z",2.016,18,893],["2026-10-29T06:33:18Z",2.007,18,897],["2026-10-29T16:06:41Z",2.206,21,952],["2026-10-29T16:24:40Z",2.201,21,954],["2026-10-29T16:44:27Z",2.194,21,957],["2026-10-29T17:17:24Z",2.296,22,958],["2026-10-29T17:31:19Z",2.289,22,961],["2026-10-29T17:45:46Z",2.285,22,963],["2026-10-29T17:59:46Z",2.275,22,967],["2026-10-29T18:20:18Z",2.268,22,970],["2026-10-29T18:39:33Z",2.261,22,973],["2026-10-29T18:52:44Z",2.357,23,976],["2026-10-29T19:04:20Z",2.349,23,979],["2026-10-29T19:30:28Z",2.345,23,981],["2026-10-29T19:49:49Z",2.333,23,986],["2026-10-29T20:37:19Z",2.429,24,988],["2026-10-29T21:11:56Z",2.424,24,990],["2026-10-29T21:37:12Z",2.419,24,992],["2026-10-29T22:01:31Z",2.412,24,995],["2026-10-29T22:30:08Z",1.904,19,998],["2026-10-29T22:48:48Z",1.9,19,1000],["2026-10-30T01:00:40Z",1.896,19,1002],["2026-10-30T02:11:39Z",2.094,21,1003],["2026-10-30T03:27:41Z",2.09,21,1005],["2026-10-30T04:13:13Z",2.183,22,1008],["2026-10-30T04:47:47Z",2.279,23,1009],["2026-10-30T05:05:42Z",2.275,23,1011],["2026-10-30T05:51:36Z",2.27,23,1013],["2026-10-30T06:42:44Z",2.266,23,1015],["2026-10-30T07:34:19Z",2.262,23,1017],["2026-10-30T08:19:10Z",2.257,23,1019],["2026-10-30T08:40:35Z",2.351,24,1021],["2026-10-30T09:13:16Z",2.441,25,1024],["2026-10-30T09:32:31Z",2.537,26,1025],["2026-10-30T09:57:51Z",2.532,26,1027],["2026-10-30T10:18:07Z",2.527,26,1029],["2026-10-30T11:32:49Z",2.619,27,1031],["2026-10-30T12:12:07Z",2.614,27,1033],["2026-10-30T12:31:23Z",2.606,27,1036],["2026-10-30T13:26:21Z",2.601,27,1038],["2026-10-30T15:10:58Z",2.594,27,1041],["2026-10-30T15:43:52Z",2.589,27,1043],["2026-10-30T16:15:10Z",2.667,28,1050],["2026-10-30T16:42:24Z",2.659,28,1053],["2026-10-30T17:01:09Z",2.746,29,1056],["2026-10-30T17:26:22Z",2.838,30,1057],["2026-10-30T18:02:17Z",2.833,30,1059],["2026-10-30T18:32:21Z",2.925,31,1060],["2026-10-30T18:50:43Z",3.016,32,1061],["2026-10-30T19:21:20Z",3.107,33,1062],["2026-10-30T20:34:03Z",3.198,34,1063],["2026-10-30T21:51:14Z",3.289,35,1064],["2026-10-30T22:15:25Z",3.283,35,1066],["2026-10-30T22:47:55Z",3.277,35,1068],["2026-10-30T23:37:05Z",3.364,36,1070],["2026-10-31T00:21:33Z",3.355,36,1073],["2026-10-31T01:13:58Z",3.346,36,1076],["2026-10-31T02:12:21Z",3.34,36,1078],["2026-10-31T02:25:36Z",3.429,37,1079],["2026-10-31T02:45:25Z",3.519,38,1080],["2026-10-31T03:29:13Z",3.512,38,1082],["2026-10-31T06:13:11Z",3.585,39,1088],["2026-10-31T12:34:41Z",3.836,42,1095],["2026-10-31T16:23:54Z",3.909,43,1100],["2026-10-31T16:58:34Z",3.996,44,1101],["2026-10-31T17:53:38Z",3.978,44,1106],["2026-10-31T18:11:39Z",3.971,44,1108],["2026-10-31T18:50:56Z",4.058,45,1109],["2026-10-31T20:19:34Z",4.14,46,1111],["2026-10-31T20:54:37Z",4.313,48,1113],["2026-10-31T21:37:16Z",3.674,41,1116],["2026-10-31T22:02:56Z",3.667,41,1118],["2026-10-31T22:39:11Z",3.661,41,1120],["2026-10-31T23:38:12Z",1.779,20,1124],["2026-11-01T00:09:48Z",1.775,20,1127],["2026-11-01T00:30:02Z",1.95,22,1128],["2026-11-01T00:42:55Z",1.947,22,1130],["2026-11-01T01:02:58Z",1.943,22,1132],["2026-11-01T01:23:19Z",1.94,22,1134],["2026-11-01T01:50:13Z",2.116,24,1134],["2026-11-01T02:21:26Z",2.203,25,1135],["2026-11-01T03:40:38Z",2.199,25,1137],["2026-11-01T03:59:39Z",2.283,26,1139],["2026-11-01T04:58:51Z",2.277,26,1142],["2026-11-01T05:10:49Z",2.271,26,1145],["2026-11-01T05:49:10Z",2.271,26,1145],["2026-11-01T06:20:24Z",2.356,27,1146],["2026-11-01T06:40:20Z",2.441,28,1147],["2026-11-01T07:36:48Z",2.437,28,1149],["2026-11-01T08:36:48Z",2.431,28,1152],["2026-11-01T09:40:59Z",2.513,29,1154],["2026-11-01T10:18:11Z",2.509,29,1156],["2026-11-01T10:44:23Z",2.593,30,1157],["2026-11-01T11:31:32Z",2.586,30,1160],["2026-11-01T12:10:49Z",2.759,32,1160],["2026-11-01T12:50:48Z",2.672,31,1160],["2026-11-01T13:23:57Z",2.668,31,1162],["2026-11-01T14:03:04Z",2.754,32,1162],["2026-11-01T15:03:18Z",2.749,32,1164],["2026-11-01T15:37:24Z",2.744,32,1166],["2026-11-01T18:16:17Z",2.74,32,1168],["2026-11-01T19:43:03Z",2.735,32,1170],["2026-11-01T23:47:34Z",2.733,32,1171],["2026-11-02T00:51:49Z",2.728,32,1173],["2026-11-02T01:32:03Z",2.899,34,1173],["2026-11-02T02:51:29Z",2.894,34,1175],["2026-11-02T04:49:57Z",2.891,34,1176],["2026-11-02T05:28:59Z",2.464,29,1177],["2026-11-02T06:32:18Z",2.46,29,1179],["2026-11-02T07:57:39Z",2.456,29,1181],["2026-11-02T08:24:50Z",2.288,27,1180],["2026-11-02T09:29:05Z",2.373,28,1180],["2026-11-02T10:52:34Z",2.371,28,1181],["2026-11-02T11:39:36Z",2.453,29,1182],["2026-11-02T12:45:11Z",2.534,30,1184],["2026-11-02T14:08:07Z",2.53,30,1186],["2026-11-02T15:04:13Z",2.698,32,1186],["2026-11-02T15:50:48Z",1.853,22,1187],["2026-11-02T17:00:29Z",1.849,22,1190],["2026-11-02T17:26:42Z",1.846,22,1192],["2026-11-02T18:08:02Z",1.844,22,1193],["2026-11-02T18:50:55Z",1.841,22,1195],["2026-11-02T19:20:34Z",1.756,21,1196],["2026-11-02T19:34:39Z",1.756,21,1196],["2026-11-02T21:51:49Z",2.002,24,1199],["2026-11-02T22:37:55Z",1.997,24,1202],["2026-11-02T23:01:22Z",1.993,24,1204],["2026-11-02T23:39:21Z",2.073,25,1206],["2026-11-02T23:51:54Z",2.066,25,1210],["2026-11-03T00:36:39Z",2.063,25,1212],["2026-11-03T02:46:54Z",2.059,25,1214],["2026-11-03T03:18:41Z",1.974,24,1216],["2026-11-03T04:05:11Z",2.054,25,1217],["2026-11-03T05:28:01Z",2.135,26,1218],["2026-11-03T06:14:03Z",2.133,26,1219],["2026-11-03T06:40:08Z",2.211,27,1221],["2026-11-03T06:59:38Z",2.291,28,1222],["2026-11-03T07:21:06Z",2.369,29,1224],["2026-11-03T08:07:43Z",2.367,29,1225],["2026-11-03T09:39:40Z",2.363,29,1227],["2026-11-03T10:56:22Z",2.441,30,1229],["2026-11-03T11:51:07Z",2.514,31,1233],["2026-11-03T12:37:23Z",2.514,31,1233],["2026-11-03T13:33:26Z",2.593,32,1234],["2026-11-03T13:53:25Z",2.672,33,1235],["2026-11-03T14:57:53Z",2.668,33,1237],["2026-11-03T15:16:49Z",2.744,34,1239],["2026-11-03T17:04:20Z",2.744,34,1239],["2026-11-03T18:17:14Z",2.738,34,1242],["2026-11-03T18:54:07Z",2.733,34,1244],["2026-11-03T19:17:24Z",2.731,34,1245],["2026-11-03T19:47:26Z",2.727,34,1247],["2026-11-03T20:39:13Z",2.722,34,1249],["2026-11-03T20:53:46Z",2.718,34,1251],["2026-11-03T21:29:02Z",3.042,38,1249],["2026-11-03T21:44:34Z",2.836,35,1234],["2026-11-03T22:21:03Z",2.91,36,1237],["2026-11-03T23:00:48Z",2.989,37,1238],["2026-11-03T23:14:23Z",3.065,38,1240],["2026-11-03T23:35:15Z",3.06,38,1242],["2026-11-04T00:13:00Z",3.055,38,1244],["2026-11-04T00:57:35Z",17.244,214,1241],["2026-11-04T01:12:40Z",17.164,213,1241],["2026-11-04T01:25:22Z",17.23,214,1242],["2026-11-04T01:39:23Z",2.818,35,1242],["2026-11-04T02:00:58Z",2.814,35,1244],["2026-11-04T02:23:24Z",2.892,36,1245],["2026-11-04T02:59:46Z",2.967,37,1247],["2026-11-04T03:51:38Z",2.887,36,1247],["2026-11-04T04:21:26Z",2.965,37,1248],["2026-11-04T04:36:19Z",3.042,38,1249],["2026-11-04T05:23:24Z",3.12,39,1250],["2026-11-04T05:45:40Z",3.197,40,1251],["2026-11-04T06:21:25Z",3.115,39,1252],["2026-11-04T06:51:50Z",2.632,33,1254],["2026-11-04T07:54:04Z",2.627,33,1256],["2026-11-04T08:41:34Z",2.623,33,1258],["2026-11-04T09:13:56Z",2.617,33,1261],["2026-11-04T10:07:18Z",2.692,34,1263],["2026-11-04T11:22:42Z",2.686,34,1266],["2026-11-04T13:21:30Z",2.762,35,1267],["2026-11-04T13:47:38Z",2.681,34,1268],["2026-11-04T14:18:59Z",2.681,34,1268],["2026-11-04T14:50:53Z",2.758,35,1269],["2026-11-04T15:45:29Z",2.752,35,1272],["2026-11-04T17:03:33Z",2.752,35,1272],["2026-11-04T17:31:22Z",2.826,36,1274],["2026-11-04T20:02:57Z",2.826,36,1274],["2026-11-04T20:19:52Z",2.819,36,1277],["2026-11-04T21:14:19Z",2.893,37,1279],["2026-11-04T21:32:28Z",2.893,37,1279],["2026-11-04T22:03:59Z",2.893,37,1279],["2026-11-04T22:17:14Z",2.893,37,1279],["2026-11-04T22:31:12Z",2.893,37,1279],["2026-11-04T23:18:24Z",2.888,37,1281],["2026-11-04T23:32:22Z",2.96,38,1284],["2026-11-04T23:51:25Z",3.035,39,1285],["2026-11-05T00:11:49Z",2.568,33,1285],["2026-11-05T01:07:07Z",2.644,34,1286],["2026-11-05T01:44:13Z",2.64,34,1288],["2026-11-05T02:49:20Z",2.636,34,1290],["2026-11-05T03:29:26Z",1.705,22,1290],["2026-11-05T03:51:36Z",1.701,22,1293],["2026-11-05T04:25:44Z",1.7,22,1294],["2026-11-05T06:03:07Z",1.698,22,1296],["2026-11-05T07:17:25Z",1.773,23,1297],["2026-11-05T07:41:34Z",1.849,24,1298],["2026-11-05T07:55:05Z",1.923,25,1300]],"reads":860,"share_top5":0.95,"median":2.347,"best":["2026-10-24T07:02:01Z",0.243,1,411],"first":["2026-10-22T22:56:48Z",81.818,27,33],"last":["2026-11-05T07:55:05Z",1.923,25,1300],"src":"bench/board_history.jsonl (our_rank / entries per decode board read, to the close)"},"rank_caption":"In 95% of 860 board reads we were in the top 5% of the decode field; the median read was the top 2.3%. Board reads show #1 from 00:20 to 07:06 UTC on Oct 24. The one spike is the Nov 4 disqualification for compressed source (#215 for 50 minutes, until a clean rebuild was scored). #25 of 1,300 at the close, the top 1.9%. Our overall place equalled our decode place in 402 of 402 snapshots that recorded both.","experiment":{"box_min_per_run":3.0049335723128663,"fn_calls":3404,"fn_gpu_s":305.6574618096357,"fn_build_s":305.7135722679201,"fn_usd":0.6110370152761456,"per_run_work":0.4758209082534375,"idea_to_trial_min":6.516,"paired_runs":1612,"paired_confirmed":130,"work_ratio":0.666232490946445},"money_curve":[["2026-10-22T23:00:00Z",4.0,1.2836,27],["2026-10-23T02:00:00Z",113.3,1.489,10],["2026-10-23T05:00:00Z",399.1,1.7543,2],["2026-10-23T08:00:00Z",600.9,1.7762,2],["2026-10-23T11:00:00Z",750.2,1.7956,2],["2026-10-23T14:00:00Z",872.8,1.7956,2],["2026-10-23T17:00:00Z",1004.0,1.7956,2],["2026-10-23T20:00:00Z",1190.6,1.8059,10],["2026-10-23T23:00:00Z",1480.4,1.8106,12],["2026-10-24T02:00:00Z",1864.1,1.9506,1],["2026-10-24T05:00:00Z",2378.4,1.9603,1],["2026-10-24T08:00:00Z",2443.1,1.9603,1],["2026-10-24T11:00:00Z",2443.7,1.9603,1],["2026-10-24T14:00:00Z",2444.4,1.9603,1],["2026-10-24T17:00:00Z",2451.1,1.9603,1],["2026-10-24T20:00:00Z",2893.4,1.9603,3],["2026-10-24T23:00:00Z",3153.3,1.9676,4],["2026-10-25T02:00:00Z",4647.6,1.988,6],["2026-10-25T05:00:00Z",4925.5,1.988,7],["2026-10-25T08:00:00Z",5163.3,2.0056,6],["2026-10-25T11:00:00Z",5174.0,2.0056,6],["2026-10-25T14:00:00Z",5174.0,2.0056,6],["2026-10-25T17:00:00Z",5174.6,2.0056,6],["2026-10-25T20:00:00Z",5181.3,2.0056,10],["2026-10-25T23:00:00Z",5181.3,2.0056,11],["2026-10-26T02:00:00Z",5181.3,2.0056,11],["2026-10-26T05:00:00Z",5181.3,2.0056,11],["2026-10-26T08:00:00Z",5181.3,2.0056,11],["2026-10-26T11:00:00Z",5181.3,2.0056,11],["2026-10-26T14:00:00Z",5181.3,2.0056,11],["2026-10-26T17:00:00Z",5181.3,2.0056,11],["2026-10-26T20:00:00Z",5186.0,2.0056,15],["2026-10-26T23:00:00Z",5362.5,2.0056,15],["2026-10-27T02:00:00Z",5529.7,2.0207,16],["2026-10-27T05:00:00Z",5682.3,2.0283,11],["2026-10-27T08:00:00Z",5904.8,2.0283,12],["2026-10-27T11:00:00Z",6143.3,2.0283,13],["2026-10-27T14:00:00Z",6398.5,2.0357,14],["2026-10-27T17:00:00Z",6655.0,2.0428,14],["2026-10-27T20:00:00Z",6956.8,2.0701,15],["2026-10-27T23:00:00Z",7272.6,2.0946,16],["2026-10-28T02:00:00Z",7681.7,2.0946,17],["2026-10-28T05:00:00Z",8094.7,2.0946,21],["2026-10-28T08:00:00Z",8502.5,2.0946,23],["2026-10-28T11:00:00Z",8848.9,2.1025,23],["2026-10-28T14:00:00Z",9216.7,2.1025,23],["2026-10-28T17:00:00Z",9960.8,2.1025,23],["2026-10-28T20:00:00Z",10665.0,2.1043,24],["2026-10-28T23:00:00Z",11335.3,2.1415,15],["2026-10-29T02:00:00Z",12230.7,2.1415,16],["2026-10-29T05:00:00Z",13068.8,2.1473,16],["2026-10-29T08:00:00Z",13725.1,2.1473,18],["2026-10-29T11:00:00Z",14086.8,2.1473,18],["2026-10-29T14:00:00Z",14437.9,2.1473,18],["2026-10-29T17:00:00Z",14884.3,2.1473,22],["2026-10-29T20:00:00Z",15551.9,2.1473,23],["2026-10-29T23:00:00Z",16142.8,2.1533,19],["2026-10-30T02:00:00Z",16461.9,2.1537,20],["2026-10-30T05:00:00Z",16769.7,2.1537,23],["2026-10-30T08:00:00Z",17096.2,2.1559,23],["2026-10-30T11:00:00Z",17410.7,2.1762,26],["2026-10-30T14:00:00Z",17714.5,2.1762,27],["2026-10-30T17:00:00Z",17944.3,2.1762,29],["2026-10-30T20:00:00Z",17999.6,2.1762,34],["2026-10-30T23:00:00Z",18230.8,2.1901,36],["2026-10-31T02:00:00Z",18557.9,2.1901,36],["2026-10-31T05:00:00Z",18713.8,2.1901,38],["2026-10-31T08:00:00Z",18743.8,2.1901,39],["2026-10-31T11:00:00Z",18767.8,2.1901,39],["2026-10-31T14:00:00Z",18800.4,2.1901,42],["2026-10-31T17:00:00Z",18979.6,2.1901,44],["2026-10-31T20:00:00Z",19502.0,2.1901,45],["2026-10-31T23:00:00Z",19661.9,2.203,41],["2026-11-01T02:00:00Z",19960.3,2.203,24],["2026-11-01T05:00:00Z",20282.1,2.203,26],["2026-11-01T08:00:00Z",20603.2,2.203,28],["2026-11-01T11:00:00Z",20947.0,2.203,30],["2026-11-01T14:00:00Z",21258.8,2.2047,32],["2026-11-01T17:00:00Z",21572.6,2.2047,32],["2026-11-01T20:00:00Z",21895.7,2.2047,32],["2026-11-01T23:00:00Z",22276.1,2.2047,32],["2026-11-02T02:00:00Z",22637.9,2.2047,34],["2026-11-02T05:00:00Z",22935.1,2.2064,34],["2026-11-02T08:00:00Z",23209.5,2.2064,29],["2026-11-02T11:00:00Z",23548.7,2.2064,28],["2026-11-02T14:00:00Z",23817.8,2.2064,30],["2026-11-02T17:00:00Z",24086.3,2.2107,22],["2026-11-02T20:00:00Z",24378.8,2.2107,21],["2026-11-02T23:00:00Z",24614.6,2.2107,24],["2026-11-03T02:00:00Z",24972.4,2.2107,25],["2026-11-03T05:00:00Z",25315.5,2.2107,25],["2026-11-03T08:00:00Z",25549.3,2.2107,28],["2026-11-03T11:00:00Z",25861.8,2.2107,30],["2026-11-03T14:00:00Z",26161.0,2.2107,33],["2026-11-03T17:00:00Z",26388.1,2.2107,34],["2026-11-03T20:00:00Z",26657.3,2.2107,34],["2026-11-03T23:00:00Z",26790.5,2.2107,37],["2026-11-04T02:00:00Z",26898.5,2.2107,35],["2026-11-04T05:00:00Z",27025.7,2.2114,38],["2026-11-04T08:00:00Z",27159.6,2.2152,33],["2026-11-04T11:00:00Z",27244.9,2.2152,34],["2026-11-04T14:00:00Z",27324.9,2.2161,34],["2026-11-04T17:00:00Z",27484.1,2.2161,35],["2026-11-04T20:00:00Z",27683.3,2.2164,36],["2026-11-04T23:00:00Z",27891.8,2.2164,37],["2026-11-05T02:00:00Z",28043.7,2.2201,34],["2026-11-05T05:00:00Z",28131.7,2.2252,22],["2026-11-05T07:00:00Z",28131.7,2.2252,22]],"planner":[{"key":"muse_decode_layer","label":"Decode layer","slot":1,"ceiling":3.02,"a":1.2677,"b":0.3061,"runs":26959,"x_final":2.4513545799748853},{"key":"muse_moe_w4a16","label":"MoE W4A16","slot":2,"ceiling":3.07,"a":1.2498,"b":0.4013,"runs":9297,"x_final":2.4292289478838875},{"key":"muse_attention_decode","label":"Attention decode","slot":3,"ceiling":3.38,"a":1.3074,"b":0.4628,"runs":12256,"x_final":2.835522644274862},{"key":"muse_prefill_layer","label":"Prefill layer","slot":4,"ceiling":2.27,"a":1.1948,"b":0.2367,"runs":4080,"x_final":1.920697465982476},{"key":"muse_bf16_proj","label":"BF16 projections","slot":5,"ceiling":1.99,"a":1.0144,"b":0.1559,"runs":6026,"x_final":1.5375097993769167}],"title":"Same model, 2,014 teams: what makes an agent swarm win","findings":[{"title":"Progress tracked the logarithm of compute.","text":"Each tenfold increase in GPU runs on a problem went with about 17.3 percentage points more of the gap between SGLang parity and the model ceiling, pooled over five problems (15 to 19 points per problem). The correlation between log experiments and gap closed is r = 0.91 to 0.97 per problem. It is an association over time, not a controlled test."},{"title":"The check was scarcer than the ideas.","text":"The swarm registered 14,936 ideas and timed 7,621 on a GPU. 236 passed a full check with a same-box control. The median idea predicted 0.25% and measured 0.01%."},{"title":"Most late records were re-measurements.","text":"The swarm set 209 24-hour records. 130 gained less than 1%, inside run noise. From Oct 27, 62 of the 72 records with a code hash re-ran bytes that had already run."},{"title":"One writer model wrote most of the new code.","text":"By inference from the first lane that ran each file, Claude Opus 5.5 in Claude Code wrote about 123 of 211 record files. Sonnet, Haiku, Kimi K3 and Terra lanes, 66,629 lane runs between them, wrote 2."},{"title":"Late wins were small edits of the incumbent.","text":"Our notes record that from Oct 31 to Nov 1 all 44 decode hits edited the incumbent file (believed). Through the close, full decode trials from Oct 31 hit 2.3% (32 of 1,398)."},{"title":"Rival code did better than papers, weakly.","text":"Ideas from rival or peer code passed a controlled check 10 times in 76 tries. Ideas from papers passed 0 times in 22. One-sided Fisher p = 0.068: suggestive, not significant."},{"title":"We stayed in the top few percent as the field grew 40 times.","text":"We reached #1 on the decode board on Oct 24, about a day in, and held it for at least 6.8 hours. Over the two weeks the field grew from 33 teams to 1,300. In 95% of 860 board reads we were in its top 5%. We finished #25 of 2,014 overall, the top 1.2%."}],"classes_caption":"Share of the summed log gain by kind of change. Kinds are hand-tagged from run notes (believed). Re-sends, recoveries after the disqualification and unattributed sends are left out. Toggle to see the late chain.","bench":{"name":"FloorBench-H200","version":"grader 0.2.0; public sample 0.3.0","canary":"7d3c1f0e-5b2a-4c8e-9a61-2f4e8b0d93c7","pitch":"RL environments for GPU kernel optimisation on an NVIDIA H200: real serving kernels, scored against the physical floor, graded in isolated processes with anti-cheat gates.","families":[{"id":"dense_proj_bf16","tier":"kernel","n_tasks":30,"what":"BF16 dense projections (qkv, o, pooler, unpooler) and an FP32 router GEMM of a MoE decoder layer at decode and verify batch sizes 1-256.","needs":"H200 for timing, CPU for every gate"},{"id":"moe_nvfp4_ep8","tier":"kernel","n_tasks":6,"what":"The routed SwiGLU experts of a top-8-of-256 MoE layer with NVFP4 weights on the busiest GPU of an 8-way expert-parallel deployment.","needs":"H200 for timing, CPU for every gate"},{"id":"attn_decode_fp8kv","tier":"kernel","n_tasks":12,"what":"GQA decode and 8-token verify attention over a paged FP8 E4M3 KV cache (80K context, sliding and full layers), incl. QK-norm, RoPE and the cache write.","needs":"H200 for timing, CPU for every gate"},{"id":"decoder_layer_tp8","tier":"kernel","n_tasks":12,"what":"One whole TP8/EP8 decoder layer (attention, norms, router, NVFP4 MoE, residuals) at decode and verify shapes.","needs":"H200 for timing, CPU for every gate"},{"id":"prefill_layer","tier":"kernel","n_tasks":8,"what":"The same decoder layer at prefill shapes: 753 new tokens on an 80K conversation and 16K-token first-turn chunks.","needs":"H200 for timing, CPU for every gate"},{"id":"isa_instruction_budget","tier":"chip","n_tasks":2,"what":"Cut the warp instructions a kernel executes toward the minimal instruction budget of its math, without a slowdown.","needs":"nvcc>=12.4, cuobjdump, ncu, H200"},{"id":"ptx_roundtrip_surgery","tier":"compiler","n_tasks":7,"what":"Edit the PTX of a naive sm_90a kernel; graded on the SASS ptxas builds and timed against the STARTER.","needs":"ptxas, cuobjdump, H200"},{"id":"sass_static_verifier_gates","tier":"compiler","n_tasks":2,"what":"Repair kernels whose correctness depends on an ordering only a SASS static verifier can prove (PDL wait order, async-proxy fence).","needs":"nvcc>=12.4, cuobjdump, H200"},{"id":"sm90a_pipe_microbench","tier":"chip","n_tasks":6,"what":"Keep one Hopper execution pipe busy while computing a checked function; scored by Nsight Compute pipe counters.","needs":"nvcc>=12.4, cuobjdump, ncu, H200"}],"gates":[{"name":"Static policy (AST)","catches":"introspection, dynamic import, sys.modules / builtins access, monkeypatching (incl. through aliases), process and network access, secret reads, timer hooks, tensor subclasses"},{"name":"Process isolation + integrity snapshot","catches":"native code that patches torch or harness functions (C API), a slowed in-process baseline"},{"name":"First call (rel L2 + precision probe rows)","catches":"wrong math, silent precision drops (FP8 weights), skipped norms, missing cache writes, truncated attention context, dropped router bias"},{"name":"Inputs unchanged","catches":"a kernel that writes its inputs"},{"name":"Fresh inputs at the same address","catches":"output caches keyed on data_ptr / id / shape"},{"name":"Tensor-cloned state copy","catches":"baked addresses: TMA descriptors, pointer tables, graphs captured in prepare()"},{"name":"Poisoned original state","catches":"closures and hidden references to the original weights"},{"name":"CUDA-graph capture + timed-output check","catches":"a timed graph that does less work than the eager call (capture-dependent behaviour)"},{"name":"Outputs settle after sync","catches":"deferred work on side streams or threads, outside the timed window"},{"name":"Wall-clock cross-check","catches":"faked or patched timers"},{"name":"Physics bound","catches":"any case faster than half its speed-of-light floor"},{"name":"Race gate (dataset labels)","catches":"live warpgroup races in decode / attention CUDA kernels"},{"name":"PTX / device-source static policy","catches":"%globaltimer / %clock reads, extern calls (printf), device malloc, system-scope access"},{"name":"Action schema + launch limits","catches":"self-reported speed or counters, oversize launches, changed fixed launches"},{"name":"SASS static verifier (pdl_order, fence_order)","catches":"[sealed until Nov 19]"}],"receipt_gates":["static policy","first call","fresh inputs","state copy","poisoned original","CUDA graph","inputs unchanged"],"receipt_episode":"argon-bf16-v2-20261102","redteam":[{"id":"rt01_patch_checker","cheat":"the checker lives in the same process; replace it, return garbage. (Old grader: regex missed sys.modules.)","caught_by":"static policy rejected: ['banned_import']","family":"dense_proj_bf16"},{"id":"rt02_obfuscated_process","cheat":"string-built import dodges a regex scan and reaches os.system (could read the grader, patch files, phone home).","caught_by":"static policy rejected: ['banned_call']","family":"dense_proj_bf16"},{"id":"rt03_timer_patch","cheat":"correct math, but every CUDA event reads 1 ns: infinite speedup. (Old regex only caught `torch.cuda.Event =`.)","caught_by":"static policy rejected: ['monkeypatch']","family":"dense_proj_bf16"},{"id":"rt04_slow_baseline","cheat":"the baseline calls F.linear in the same process; make F.linear 10x slower, use matmul ourselves.","caught_by":"static policy rejected: ['monkeypatch']","family":"dense_proj_bf16"},{"id":"rt05_lazy_subclass","cheat":"return a Tensor subclass. The timed call does no math; the result is computed only when the checker reads it.","caught_by":"static policy rejected: ['tensor_subclass']","family":"dense_proj_bf16"},{"id":"rt06_alias_patch","cheat":"patch the timer through an imported alias, so no `torch.cuda.Event` text appears on the assignment line.","caught_by":"static policy rejected: ['monkeypatch']","family":"dense_proj_bf16"},{"id":"rt07_native_patch","cheat":"the AST cannot see C++. Compiled code patches torch.nn.functional.linear through the Python C API, which slows the in-process baseline. The integrity snapshot (worker.py) must catch the changed attribute.","caught_by":"load failed: cand worker: RuntimeError: integrity: harness or torch attribute changed: t5:","family":"dense_proj_bf16"},{"id":"rt08_graph_skip","cheat":"honest on every test shape (all gates pass) and on eager calls. On the SCORED shapes, which the prompt lists, a graph capture records only a copy of the last eager result, so the timed graph does almost no work. The old grader checked one eager call of each scored case before timing, never the timed graph itself.","caught_by":"H200 timed-output check (receipt 20261102T183017Z-box-sh-56241)","family":"dense_proj_bf16"},{"id":"rt09_secret_read","cheat":"read secrets / the grader's environment.","caught_by":"static policy rejected: ['env_secret']","family":"dense_proj_bf16"},{"id":"rt10_thread_launch","cheat":"a background thread keeps working after kernel_fn returns (work outside the timed window).","caught_by":"static policy rejected: ['banned_import']","family":"dense_proj_bf16"},{"id":"rt6_empty","cheat":"empty kernel: outputs stay NaN","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"ptx_roundtrip_surgery"},{"id":"rt6_partial","cheat":"skipping work: half the rows unwritten","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"ptx_roundtrip_surgery"},{"id":"rt6_cache","cheat":"returning cached output: wrong on the fresh second input set","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"ptx_roundtrip_surgery"},{"id":"rt6_globaltimer","cheat":"reads the global timer","caught_by":"device-source static policy (timer reads, extern calls)","family":"ptx_roundtrip_surgery"},{"id":"rt6_extern","cheat":"external symbol (vprintf)","caught_by":"device-source static policy (timer reads, extern calls)","family":"ptx_roundtrip_surgery"},{"id":"rt6_lie_key","cheat":"the action reports its own speed","caught_by":"action schema: unknown action keys (self-reported numbers) are INVALID","family":"ptx_roundtrip_surgery"},{"id":"rt6_launch_oversize","cheat":"block of 2048 threads","caught_by":"launch limits / fixed launch","family":"ptx_roundtrip_surgery"},{"id":"rt7_pdl_nowait","cheat":"PDL kernel with no wait (race by construction)","caught_by":"SASS pdl_order gate (reads before griddepcontrol.wait)","family":"sass_static_verifier_gates"},{"id":"rt7_pdl_skipwait","cheat":"wait that can be skipped","caught_by":"SASS pdl_order gate (reads before griddepcontrol.wait)","family":"sass_static_verifier_gates"},{"id":"rt7_pdl_zero","cheat":"deleted kernel output, gate alone would pass","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"sass_static_verifier_gates"},{"id":"rt7_pdl_launch","cheat":"changes the fixed launch","caught_by":"launch limits / fixed launch","family":"sass_static_verifier_gates"},{"id":"rt7_fence_after","cheat":"fence after the wgmma","caught_by":"SASS fence_order gate (fence.proxy.async before every wgmma)","family":"sass_static_verifier_gates"},{"id":"rt7_fence_dead","cheat":"fence in a dead branch","caught_by":"SASS fence_order gate (fence.proxy.async before every wgmma)","family":"sass_static_verifier_gates"},{"id":"rt7_fence_nowgmma","cheat":"no wgmma at all","caught_by":"SASS fence_order gate (fence.proxy.async before every wgmma)","family":"sass_static_verifier_gates"},{"id":"rt8_xu_garbage","cheat":"pipe busy but wrong output","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"sm90a_pipe_microbench"},{"id":"rt8_xu_clock","cheat":"reads clock64","caught_by":"device-source static policy (timer reads, extern calls)","family":"sm90a_pipe_microbench"},{"id":"rt8_lie_key","cheat":"the action reports its own pipe %","caught_by":"action schema: unknown action keys (self-reported numbers) are INVALID","family":"sm90a_pipe_microbench"},{"id":"rt8_tma_ldg","cheat":"LDG/STG copy, no UBLKCP","caught_by":"SASS opcode rules (required / forbidden instructions)","family":"sm90a_pipe_microbench"},{"id":"rt8_wgmma_noop","cheat":"no HGMMA","caught_by":"SASS opcode rules (required / forbidden instructions)","family":"sm90a_pipe_microbench"},{"id":"rt9_attn_halfkeys","cheat":"skipping work","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"isa_instruction_budget"},{"id":"rt9_attn_cache","cheat":"cached output","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"isa_instruction_budget"},{"id":"rt9_deq_halfrows","cheat":"skipping rows","caught_by":"checked outputs: the timed graph replays on a fresh second input set (H200)","family":"isa_instruction_budget"},{"id":"rt7_fence_dropped","cheat":"proxy fence dropped for speed (a real race on the H200)","caught_by":"SASS fence_order gate (fence.proxy.async before every wgmma)","family":"sass_static_verifier_gates"}],"episodes":[{"id":"argon-bf16-v2-20261102","model":"Gemini 4 Argon High","harness":"Antigravity 1.4.12","family":"dense_proj_bf16","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":51.14314098497682},"T2":{"x":1.4463634627802266,"pct":67.24688226589855},"T3":{"x":1.8927269255604537,"pct":88.0},"T4":{"x":1.9357434465959187,"pct":90.0}},"steps":[{"step":1,"x":0.9204010713015859,"lo":0.9168907826490328,"hi":0.9238674159665393,"pct":42.99131716964966,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.49,"ours_us":37.44,"base_us":18.36,"floor_pct":26.4},{"load":"decode 1","shape":"m=1 o","x":0.827,"ours_us":11.17,"base_us":9.23,"floor_pct":49.2},{"load":"decode 1","shape":"m=1 pooler","x":0.85,"ours_us":32.56,"base_us":27.66,"floor_pct":49.1},{"load":"decode 1","shape":"m=1 unpooler","x":0.728,"ours_us":11.9,"base_us":8.66,"floor_pct":31.5},{"load":"decode 1","shape":"m=1 router","x":0.954,"ours_us":12.64,"base_us":12.06,"floor_pct":29.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.488,"ours_us":37.7,"base_us":18.41,"floor_pct":26.4},{"load":"decode 16","shape":"m=16 o","x":0.818,"ours_us":11.58,"base_us":9.47,"floor_pct":48.0},{"load":"decode 16","shape":"m=16 pooler","x":0.856,"ours_us":32.65,"base_us":27.96,"floor_pct":49.2},{"load":"decode 16","shape":"m=16 unpooler","x":0.77,"ours_us":32.88,"base_us":25.33,"floor_pct":48.8},{"load":"decode 16","shape":"m=16 router","x":2.232,"ours_us":12.82,"base_us":28.62,"floor_pct":29.7},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.467,"ours_us":40.46,"base_us":18.91,"floor_pct":24.7},{"load":"decode 32","shape":"m=32 o","x":0.739,"ours_us":12.79,"base_us":9.45,"floor_pct":43.9},{"load":"decode 32","shape":"m=32 pooler","x":0.827,"ours_us":34.65,"base_us":28.64,"floor_pct":46.6},{"load":"decode 32","shape":"m=32 unpooler","x":0.725,"ours_us":35.21,"base_us":25.52,"floor_pct":45.8},{"load":"decode 32","shape":"m=32 router","x":1.568,"ours_us":13.88,"base_us":21.76,"floor_pct":27.8},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.494,"ours_us":37.28,"base_us":18.4,"floor_pct":26.6},{"load":"verify 1","shape":"m=8 o","x":0.822,"ours_us":11.24,"base_us":9.24,"floor_pct":49.2},{"load":"verify 1","shape":"m=8 pooler","x":0.856,"ours_us":32.69,"base_us":27.98,"floor_pct":49.0},{"load":"verify 1","shape":"m=8 unpooler","x":0.747,"ours_us":11.86,"base_us":8.86,"floor_pct":31.8},{"load":"verify 1","shape":"m=8 router","x":2.204,"ours_us":12.85,"base_us":28.32,"floor_pct":29.4},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.001,"ours_us":22.11,"base_us":22.13,"floor_pct":47.1},{"load":"verify 16","shape":"m=128 o","x":0.861,"ours_us":5.16,"base_us":4.45,"floor_pct":115.9},{"load":"verify 16","shape":"m=128 pooler","x":0.974,"ours_us":29.38,"base_us":28.62,"floor_pct":56.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.979,"ours_us":27.32,"base_us":26.75,"floor_pct":60.9},{"load":"verify 16","shape":"m=128 router","x":2.212,"ours_us":14.33,"base_us":31.7,"floor_pct":29.4},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.994,"ours_us":26.02,"base_us":25.87,"floor_pct":48.9},{"load":"verify 32","shape":"m=256 o","x":0.982,"ours_us":11.35,"base_us":11.14,"floor_pct":59.7},{"load":"verify 32","shape":"m=256 pooler","x":0.977,"ours_us":35.98,"base_us":35.16,"floor_pct":58.6},{"load":"verify 32","shape":"m=256 unpooler","x":0.991,"ours_us":30.54,"base_us":30.28,"floor_pct":69.0},{"load":"verify 32","shape":"m=256 router","x":3.305,"ours_us":14.82,"base_us":48.97,"floor_pct":31.6}]},{"step":2,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":2,"x":1.149978737245633,"lo":1.146202242086218,"hi":1.1531204083676896,"pct":52.85559716396586,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.03,"ours_us":18.09,"base_us":18.64,"floor_pct":54.5},{"load":"decode 1","shape":"m=1 o","x":0.853,"ours_us":10.91,"base_us":9.31,"floor_pct":50.4},{"load":"decode 1","shape":"m=1 pooler","x":1.038,"ours_us":26.83,"base_us":27.85,"floor_pct":59.6},{"load":"decode 1","shape":"m=1 unpooler","x":1.081,"ours_us":8.25,"base_us":8.92,"floor_pct":45.4},{"load":"decode 1","shape":"m=1 router","x":1.315,"ours_us":9.32,"base_us":12.26,"floor_pct":40.2},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.005,"ours_us":18.63,"base_us":18.72,"floor_pct":53.3},{"load":"decode 16","shape":"m=16 o","x":0.854,"ours_us":11.11,"base_us":9.49,"floor_pct":50.0},{"load":"decode 16","shape":"m=16 pooler","x":1.035,"ours_us":26.96,"base_us":27.9,"floor_pct":59.6},{"load":"decode 16","shape":"m=16 unpooler","x":0.907,"ours_us":27.84,"base_us":25.25,"floor_pct":57.7},{"load":"decode 16","shape":"m=16 router","x":2.975,"ours_us":9.61,"base_us":28.59,"floor_pct":39.6},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.024,"ours_us":18.54,"base_us":18.98,"floor_pct":53.9},{"load":"decode 32","shape":"m=32 o","x":0.794,"ours_us":11.95,"base_us":9.49,"floor_pct":47.0},{"load":"decode 32","shape":"m=32 pooler","x":1.051,"ours_us":27.37,"base_us":28.77,"floor_pct":59.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.887,"ours_us":28.76,"base_us":25.52,"floor_pct":56.1},{"load":"decode 32","shape":"m=32 router","x":1.773,"ours_us":12.31,"base_us":21.83,"floor_pct":31.4},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.013,"ours_us":18.45,"base_us":18.68,"floor_pct":53.7},{"load":"verify 1","shape":"m=8 o","x":0.825,"ours_us":11.21,"base_us":9.25,"floor_pct":49.3},{"load":"verify 1","shape":"m=8 pooler","x":1.035,"ours_us":26.79,"base_us":27.74,"floor_pct":59.8},{"load":"verify 1","shape":"m=8 unpooler","x":1.076,"ours_us":8.29,"base_us":8.93,"floor_pct":45.4},{"load":"verify 1","shape":"m=8 router","x":2.949,"ours_us":9.7,"base_us":28.61,"floor_pct":38.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.158,"ours_us":19.5,"base_us":22.58,"floor_pct":53.5},{"load":"verify 16","shape":"m=128 o","x":0.968,"ours_us":10.37,"base_us":10.03,"floor_pct":57.7},{"load":"verify 16","shape":"m=128 pooler","x":0.994,"ours_us":29.98,"base_us":29.81,"floor_pct":55.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.867,"ours_us":31.31,"base_us":27.14,"floor_pct":53.1},{"load":"verify 16","shape":"m=128 router","x":2.351,"ours_us":13.4,"base_us":31.51,"floor_pct":31.4},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.227,"ours_us":21.78,"base_us":26.74,"floor_pct":58.5},{"load":"verify 32","shape":"m=256 o","x":0.621,"ours_us":18.45,"base_us":11.46,"floor_pct":36.7},{"load":"verify 32","shape":"m=256 pooler","x":1.065,"ours_us":33.4,"base_us":35.58,"floor_pct":63.1},{"load":"verify 32","shape":"m=256 unpooler","x":0.836,"ours_us":36.18,"base_us":30.24,"floor_pct":58.3},{"load":"verify 32","shape":"m=256 router","x":3.39,"ours_us":14.5,"base_us":49.14,"floor_pct":32.3}]},{"step":3,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":3,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":3,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"INFRA_ERROR","note":null,"cases":[]},{"step":3,"x":0.992052897634296,"lo":0.987008635553778,"hi":0.9955320365273845,"pct":46.028001426669825,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.511,"ours_us":35.85,"base_us":18.31,"floor_pct":27.5},{"load":"decode 1","shape":"m=1 o","x":1.067,"ours_us":8.64,"base_us":9.22,"floor_pct":63.7},{"load":"decode 1","shape":"m=1 pooler","x":0.697,"ours_us":40.05,"base_us":27.91,"floor_pct":39.9},{"load":"decode 1","shape":"m=1 unpooler","x":0.512,"ours_us":16.61,"base_us":8.5,"floor_pct":22.6},{"load":"decode 1","shape":"m=1 router","x":1.657,"ours_us":7.22,"base_us":11.97,"floor_pct":51.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.337,"ours_us":53.85,"base_us":18.14,"floor_pct":18.4},{"load":"decode 16","shape":"m=16 o","x":0.8,"ours_us":11.77,"base_us":9.42,"floor_pct":47.2},{"load":"decode 16","shape":"m=16 pooler","x":0.477,"ours_us":59.17,"base_us":28.25,"floor_pct":27.1},{"load":"decode 16","shape":"m=16 unpooler","x":0.62,"ours_us":40.63,"base_us":25.2,"floor_pct":39.5},{"load":"decode 16","shape":"m=16 router","x":2.956,"ours_us":9.55,"base_us":28.23,"floor_pct":39.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.01,"ours_us":18.19,"base_us":18.38,"floor_pct":55.0},{"load":"decode 32","shape":"m=32 o","x":0.997,"ours_us":9.38,"base_us":9.34,"floor_pct":59.9},{"load":"decode 32","shape":"m=32 pooler","x":1.046,"ours_us":27.22,"base_us":28.46,"floor_pct":59.3},{"load":"decode 32","shape":"m=32 unpooler","x":0.941,"ours_us":27.58,"base_us":25.96,"floor_pct":58.5},{"load":"decode 32","shape":"m=32 router","x":1.73,"ours_us":12.49,"base_us":21.6,"floor_pct":30.9},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.004,"ours_us":18.23,"base_us":18.31,"floor_pct":54.3},{"load":"verify 1","shape":"m=8 o","x":0.987,"ours_us":9.4,"base_us":9.27,"floor_pct":58.8},{"load":"verify 1","shape":"m=8 pooler","x":1.028,"ours_us":27.09,"base_us":27.85,"floor_pct":59.1},{"load":"verify 1","shape":"m=8 unpooler","x":1.019,"ours_us":8.46,"base_us":8.62,"floor_pct":44.5},{"load":"verify 1","shape":"m=8 router","x":2.919,"ours_us":9.84,"base_us":28.73,"floor_pct":38.4},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.093,"ours_us":19.81,"base_us":21.66,"floor_pct":52.6},{"load":"verify 16","shape":"m=128 o","x":0.974,"ours_us":10.18,"base_us":9.91,"floor_pct":58.8},{"load":"verify 16","shape":"m=128 pooler","x":1.007,"ours_us":29.84,"base_us":30.03,"floor_pct":55.8},{"load":"verify 16","shape":"m=128 unpooler","x":0.897,"ours_us":29.79,"base_us":26.73,"floor_pct":55.9},{"load":"verify 16","shape":"m=128 router","x":2.332,"ours_us":13.3,"base_us":31.02,"floor_pct":31.7},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.169,"ours_us":21.96,"base_us":25.68,"floor_pct":58.0},{"load":"verify 32","shape":"m=256 o","x":0.987,"ours_us":11.27,"base_us":11.13,"floor_pct":60.1},{"load":"verify 32","shape":"m=256 pooler","x":1.064,"ours_us":33.46,"base_us":35.61,"floor_pct":63.0},{"load":"verify 32","shape":"m=256 unpooler","x":0.98,"ours_us":31.16,"base_us":30.55,"floor_pct":67.7},{"load":"verify 32","shape":"m=256 router","x":3.396,"ours_us":14.38,"base_us":48.84,"floor_pct":32.5}]},{"step":4,"x":1.2222079051187993,"lo":1.2164036465044776,"hi":1.2299398970769224,"pct":56.18046785535075,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.018,"ours_us":18.38,"base_us":18.71,"floor_pct":53.7},{"load":"decode 1","shape":"m=1 o","x":1.043,"ours_us":8.99,"base_us":9.38,"floor_pct":61.1},{"load":"decode 1","shape":"m=1 pooler","x":1.008,"ours_us":27.59,"base_us":27.82,"floor_pct":57.9},{"load":"decode 1","shape":"m=1 unpooler","x":1.049,"ours_us":8.38,"base_us":8.79,"floor_pct":44.7},{"load":"decode 1","shape":"m=1 router","x":2.168,"ours_us":5.75,"base_us":12.46,"floor_pct":65.3},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.008,"ours_us":18.41,"base_us":18.55,"floor_pct":54.0},{"load":"decode 16","shape":"m=16 o","x":1.033,"ours_us":9.3,"base_us":9.61,"floor_pct":59.7},{"load":"decode 16","shape":"m=16 pooler","x":0.995,"ours_us":28.55,"base_us":28.4,"floor_pct":56.3},{"load":"decode 16","shape":"m=16 unpooler","x":0.998,"ours_us":25.54,"base_us":25.48,"floor_pct":62.9},{"load":"decode 16","shape":"m=16 router","x":3.133,"ours_us":8.95,"base_us":28.05,"floor_pct":42.5},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.994,"ours_us":19.08,"base_us":18.97,"floor_pct":52.4},{"load":"decode 32","shape":"m=32 o","x":1.05,"ours_us":9.27,"base_us":9.73,"floor_pct":60.6},{"load":"decode 32","shape":"m=32 pooler","x":1.021,"ours_us":28.21,"base_us":28.8,"floor_pct":57.2},{"load":"decode 32","shape":"m=32 unpooler","x":1.012,"ours_us":25.55,"base_us":25.85,"floor_pct":63.2},{"load":"decode 32","shape":"m=32 router","x":2.218,"ours_us":9.68,"base_us":21.47,"floor_pct":39.9},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.022,"ours_us":18.34,"base_us":18.74,"floor_pct":54.0},{"load":"verify 1","shape":"m=8 o","x":1.032,"ours_us":9.5,"base_us":9.8,"floor_pct":58.2},{"load":"verify 1","shape":"m=8 pooler","x":1.011,"ours_us":27.49,"base_us":27.8,"floor_pct":58.3},{"load":"verify 1","shape":"m=8 unpooler","x":1.077,"ours_us":8.44,"base_us":9.09,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 router","x":3.156,"ours_us":9.0,"base_us":28.4,"floor_pct":42.0},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.055,"ours_us":20.75,"base_us":21.89,"floor_pct":50.2},{"load":"verify 16","shape":"m=128 o","x":1.024,"ours_us":9.97,"base_us":10.21,"floor_pct":60.1},{"load":"verify 16","shape":"m=128 pooler","x":0.994,"ours_us":29.77,"base_us":29.6,"floor_pct":55.9},{"load":"verify 16","shape":"m=128 unpooler","x":1.015,"ours_us":26.79,"base_us":27.19,"floor_pct":62.1},{"load":"verify 16","shape":"m=128 router","x":2.812,"ours_us":11.1,"base_us":31.21,"floor_pct":38.0},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.187,"ours_us":21.83,"base_us":25.9,"floor_pct":58.4},{"load":"verify 32","shape":"m=256 o","x":0.991,"ours_us":11.43,"base_us":11.33,"floor_pct":59.2},{"load":"verify 32","shape":"m=256 pooler","x":1.064,"ours_us":33.73,"base_us":35.89,"floor_pct":62.5},{"load":"verify 32","shape":"m=256 unpooler","x":1.019,"ours_us":29.98,"base_us":30.54,"floor_pct":70.3},{"load":"verify 32","shape":"m=256 router","x":3.387,"ours_us":14.4,"base_us":48.78,"floor_pct":32.5}]},{"step":5,"x":1.2473285278319683,"lo":1.2433581254375645,"hi":1.252906872212432,"pct":57.765131220258354,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.176,"ours_us":15.57,"base_us":18.32,"floor_pct":63.4},{"load":"decode 1","shape":"m=1 o","x":1.085,"ours_us":8.72,"base_us":9.46,"floor_pct":63.0},{"load":"decode 1","shape":"m=1 pooler","x":1.057,"ours_us":26.27,"base_us":27.78,"floor_pct":60.9},{"load":"decode 1","shape":"m=1 unpooler","x":1.317,"ours_us":6.83,"base_us":9.0,"floor_pct":54.9},{"load":"decode 1","shape":"m=1 router","x":2.055,"ours_us":5.96,"base_us":12.25,"floor_pct":62.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.025,"ours_us":17.95,"base_us":18.4,"floor_pct":55.3},{"load":"decode 16","shape":"m=16 o","x":0.974,"ours_us":9.75,"base_us":9.5,"floor_pct":57.0},{"load":"decode 16","shape":"m=16 pooler","x":1.045,"ours_us":26.32,"base_us":27.5,"floor_pct":61.0},{"load":"decode 16","shape":"m=16 unpooler","x":0.992,"ours_us":25.19,"base_us":25.0,"floor_pct":63.8},{"load":"decode 16","shape":"m=16 router","x":3.132,"ours_us":9.18,"base_us":28.75,"floor_pct":41.5},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.036,"ours_us":18.05,"base_us":18.71,"floor_pct":55.4},{"load":"decode 32","shape":"m=32 o","x":1.001,"ours_us":9.58,"base_us":9.59,"floor_pct":58.6},{"load":"decode 32","shape":"m=32 pooler","x":1.052,"ours_us":27.01,"base_us":28.42,"floor_pct":59.8},{"load":"decode 32","shape":"m=32 unpooler","x":1.001,"ours_us":25.49,"base_us":25.52,"floor_pct":63.3},{"load":"decode 32","shape":"m=32 router","x":2.214,"ours_us":9.83,"base_us":21.78,"floor_pct":39.3},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.01,"ours_us":18.11,"base_us":18.3,"floor_pct":54.7},{"load":"verify 1","shape":"m=8 o","x":0.997,"ours_us":9.52,"base_us":9.5,"floor_pct":58.0},{"load":"verify 1","shape":"m=8 pooler","x":1.045,"ours_us":26.51,"base_us":27.7,"floor_pct":60.4},{"load":"verify 1","shape":"m=8 unpooler","x":1.088,"ours_us":8.4,"base_us":9.14,"floor_pct":44.8},{"load":"verify 1","shape":"m=8 router","x":3.144,"ours_us":9.08,"base_us":28.54,"floor_pct":41.6},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.112,"ours_us":19.67,"base_us":21.86,"floor_pct":53.0},{"load":"verify 16","shape":"m=128 o","x":0.997,"ours_us":10.05,"base_us":10.02,"floor_pct":59.6},{"load":"verify 16","shape":"m=128 pooler","x":0.974,"ours_us":29.37,"base_us":28.6,"floor_pct":56.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.988,"ours_us":26.31,"base_us":25.98,"floor_pct":63.2},{"load":"verify 16","shape":"m=128 router","x":2.757,"ours_us":11.46,"base_us":31.58,"floor_pct":36.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.199,"ours_us":21.6,"base_us":25.91,"floor_pct":59.0},{"load":"verify 32","shape":"m=256 o","x":1.012,"ours_us":11.1,"base_us":11.23,"floor_pct":61.0},{"load":"verify 32","shape":"m=256 pooler","x":1.068,"ours_us":33.33,"base_us":35.61,"floor_pct":63.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.997,"ours_us":30.48,"base_us":30.38,"floor_pct":69.2},{"load":"verify 32","shape":"m=256 router","x":3.385,"ours_us":14.57,"base_us":49.31,"floor_pct":32.1}]},{"step":6,"x":1.2655598166140394,"lo":1.2557470956679733,"hi":1.273046112129908,"pct":61.202200050705095,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.183,"ours_us":15.76,"base_us":18.65,"floor_pct":62.6},{"load":"decode 1","shape":"m=1 o","x":1.091,"ours_us":8.62,"base_us":9.41,"floor_pct":63.8},{"load":"decode 1","shape":"m=1 pooler","x":1.053,"ours_us":26.42,"base_us":27.82,"floor_pct":60.5},{"load":"decode 1","shape":"m=1 unpooler","x":1.296,"ours_us":6.68,"base_us":8.66,"floor_pct":56.1},{"load":"decode 1","shape":"m=1 router","x":2.068,"ours_us":5.78,"base_us":11.96,"floor_pct":64.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.013,"ours_us":18.21,"base_us":18.44,"floor_pct":54.6},{"load":"decode 16","shape":"m=16 o","x":1.009,"ours_us":9.31,"base_us":9.39,"floor_pct":59.7},{"load":"decode 16","shape":"m=16 pooler","x":1.032,"ours_us":26.76,"base_us":27.62,"floor_pct":60.0},{"load":"decode 16","shape":"m=16 unpooler","x":0.998,"ours_us":25.26,"base_us":25.22,"floor_pct":63.6},{"load":"decode 16","shape":"m=16 router","x":3.121,"ours_us":9.08,"base_us":28.35,"floor_pct":41.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.033,"ours_us":18.01,"base_us":18.61,"floor_pct":55.5},{"load":"decode 32","shape":"m=32 o","x":0.977,"ours_us":9.59,"base_us":9.38,"floor_pct":58.6},{"load":"decode 32","shape":"m=32 pooler","x":1.053,"ours_us":27.19,"base_us":28.62,"floor_pct":59.4},{"load":"decode 32","shape":"m=32 unpooler","x":1.004,"ours_us":25.33,"base_us":25.42,"floor_pct":63.7},{"load":"decode 32","shape":"m=32 router","x":2.214,"ours_us":9.82,"base_us":21.75,"floor_pct":39.3},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.029,"ours_us":17.88,"base_us":18.39,"floor_pct":55.4},{"load":"verify 1","shape":"m=8 o","x":1.004,"ours_us":9.32,"base_us":9.35,"floor_pct":59.3},{"load":"verify 1","shape":"m=8 pooler","x":1.08,"ours_us":15.71,"base_us":16.97,"floor_pct":102.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.147,"ours_us":8.25,"base_us":9.46,"floor_pct":45.6},{"load":"verify 1","shape":"m=8 router","x":3.16,"ours_us":8.96,"base_us":28.3,"floor_pct":42.2},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.089,"ours_us":19.08,"base_us":20.78,"floor_pct":54.6},{"load":"verify 16","shape":"m=128 o","x":1.053,"ours_us":7.61,"base_us":8.02,"floor_pct":78.6},{"load":"verify 16","shape":"m=128 pooler","x":0.966,"ours_us":29.98,"base_us":28.97,"floor_pct":55.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.985,"ours_us":26.82,"base_us":26.42,"floor_pct":62.0},{"load":"verify 16","shape":"m=128 router","x":2.907,"ours_us":10.91,"base_us":31.72,"floor_pct":38.6},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.18,"ours_us":21.69,"base_us":25.59,"floor_pct":58.7},{"load":"verify 32","shape":"m=256 o","x":0.981,"ours_us":11.36,"base_us":11.14,"floor_pct":59.6},{"load":"verify 32","shape":"m=256 pooler","x":1.072,"ours_us":26.4,"base_us":28.3,"floor_pct":79.9},{"load":"verify 32","shape":"m=256 unpooler","x":1.0,"ours_us":30.27,"base_us":30.28,"floor_pct":69.7},{"load":"verify 32","shape":"m=256 router","x":6.645,"ours_us":6.17,"base_us":40.99,"floor_pct":75.8}]}],"confirm":{"step":null,"x":1.2463126771790565,"lo":1.2394879065403317,"hi":1.2511950920311545,"pct":57.94576814575671,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.2,"ours_us":15.4,"base_us":18.49,"floor_pct":64.1},{"load":"decode 1","shape":"m=1 o","x":1.066,"ours_us":8.83,"base_us":9.41,"floor_pct":62.3},{"load":"decode 1","shape":"m=1 pooler","x":1.04,"ours_us":26.4,"base_us":27.45,"floor_pct":60.5},{"load":"decode 1","shape":"m=1 unpooler","x":1.307,"ours_us":6.61,"base_us":8.64,"floor_pct":56.7},{"load":"decode 1","shape":"m=1 router","x":2.0,"ours_us":6.03,"base_us":12.06,"floor_pct":62.2},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.998,"ours_us":18.3,"base_us":18.27,"floor_pct":54.3},{"load":"decode 16","shape":"m=16 o","x":0.997,"ours_us":9.38,"base_us":9.35,"floor_pct":59.2},{"load":"decode 16","shape":"m=16 pooler","x":1.043,"ours_us":26.25,"base_us":27.38,"floor_pct":61.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.01,"ours_us":24.62,"base_us":24.87,"floor_pct":65.2},{"load":"decode 16","shape":"m=16 router","x":3.126,"ours_us":9.22,"base_us":28.81,"floor_pct":41.3},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.034,"ours_us":18.1,"base_us":18.72,"floor_pct":55.3},{"load":"decode 32","shape":"m=32 o","x":0.988,"ours_us":9.51,"base_us":9.39,"floor_pct":59.1},{"load":"decode 32","shape":"m=32 pooler","x":1.036,"ours_us":27.1,"base_us":28.08,"floor_pct":59.6},{"load":"decode 32","shape":"m=32 unpooler","x":1.008,"ours_us":25.24,"base_us":25.44,"floor_pct":64.0},{"load":"decode 32","shape":"m=32 router","x":2.191,"ours_us":9.9,"base_us":21.7,"floor_pct":39.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.033,"ours_us":18.12,"base_us":18.71,"floor_pct":54.6},{"load":"verify 1","shape":"m=8 o","x":0.995,"ours_us":9.38,"base_us":9.33,"floor_pct":58.9},{"load":"verify 1","shape":"m=8 pooler","x":1.061,"ours_us":26.25,"base_us":27.86,"floor_pct":61.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.075,"ours_us":8.2,"base_us":8.82,"floor_pct":45.9},{"load":"verify 1","shape":"m=8 router","x":3.146,"ours_us":9.07,"base_us":28.54,"floor_pct":41.6},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.106,"ours_us":19.56,"base_us":21.64,"floor_pct":53.3},{"load":"verify 16","shape":"m=128 o","x":0.988,"ours_us":10.14,"base_us":10.02,"floor_pct":59.0},{"load":"verify 16","shape":"m=128 pooler","x":0.963,"ours_us":29.91,"base_us":28.79,"floor_pct":55.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.968,"ours_us":26.98,"base_us":26.13,"floor_pct":61.7},{"load":"verify 16","shape":"m=128 router","x":2.829,"ours_us":11.15,"base_us":31.54,"floor_pct":37.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.217,"ours_us":21.5,"base_us":26.17,"floor_pct":59.2},{"load":"verify 32","shape":"m=256 o","x":1.002,"ours_us":11.09,"base_us":11.11,"floor_pct":61.1},{"load":"verify 32","shape":"m=256 pooler","x":1.069,"ours_us":33.07,"base_us":35.34,"floor_pct":63.8},{"load":"verify 32","shape":"m=256 unpooler","x":0.996,"ours_us":30.19,"base_us":30.09,"floor_pct":69.8},{"load":"verify 32","shape":"m=256 router","x":3.466,"ours_us":14.17,"base_us":49.12,"floor_pct":33.0}]},"src":"bench/product/runs/learnability/argon-bf16-v2-20261102.jsonl"},{"id":"bf16-A","model":"Claude Opus 5.5","harness":"Claude Code","family":"dense_proj_bf16","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":50.75850516569654},"T2":{"x":1.4535347788908006,"pct":67.07204780258934},"T3":{"x":1.9070695577816008,"pct":88.0},"T4":{"x":1.9504120477311826,"pct":90.0}},"steps":[{"step":1,"x":1.0039254597515677,"lo":1.0004475550768672,"hi":1.0067555416800575,"pct":46.137904419574646,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.969,"ours_us":19.1,"base_us":18.51,"floor_pct":51.7},{"load":"decode 1","shape":"m=1 o","x":0.79,"ours_us":12.01,"base_us":9.49,"floor_pct":45.8},{"load":"decode 1","shape":"m=1 pooler","x":1.009,"ours_us":27.79,"base_us":28.04,"floor_pct":57.5},{"load":"decode 1","shape":"m=1 unpooler","x":0.972,"ours_us":9.09,"base_us":8.83,"floor_pct":41.3},{"load":"decode 1","shape":"m=1 router","x":1.607,"ours_us":7.63,"base_us":12.26,"floor_pct":49.1},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.969,"ours_us":19.48,"base_us":18.88,"floor_pct":51.0},{"load":"decode 16","shape":"m=16 o","x":0.775,"ours_us":12.2,"base_us":9.46,"floor_pct":45.5},{"load":"decode 16","shape":"m=16 pooler","x":0.996,"ours_us":27.79,"base_us":27.68,"floor_pct":57.8},{"load":"decode 16","shape":"m=16 unpooler","x":0.898,"ours_us":28.3,"base_us":25.4,"floor_pct":56.8},{"load":"decode 16","shape":"m=16 router","x":3.736,"ours_us":7.68,"base_us":28.69,"floor_pct":49.6},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.669,"ours_us":28.22,"base_us":18.86,"floor_pct":35.5},{"load":"decode 32","shape":"m=32 o","x":0.625,"ours_us":15.69,"base_us":9.81,"floor_pct":35.8},{"load":"decode 32","shape":"m=32 pooler","x":0.751,"ours_us":38.36,"base_us":28.8,"floor_pct":42.1},{"load":"decode 32","shape":"m=32 unpooler","x":0.634,"ours_us":39.98,"base_us":25.36,"floor_pct":40.4},{"load":"decode 32","shape":"m=32 router","x":2.497,"ours_us":8.78,"base_us":21.91,"floor_pct":44.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.968,"ours_us":19.32,"base_us":18.7,"floor_pct":51.3},{"load":"verify 1","shape":"m=8 o","x":0.806,"ours_us":11.83,"base_us":9.54,"floor_pct":46.7},{"load":"verify 1","shape":"m=8 pooler","x":0.999,"ours_us":27.87,"base_us":27.82,"floor_pct":57.5},{"load":"verify 1","shape":"m=8 unpooler","x":0.99,"ours_us":9.09,"base_us":9.0,"floor_pct":41.4},{"load":"verify 1","shape":"m=8 router","x":3.736,"ours_us":7.69,"base_us":28.74,"floor_pct":49.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.805,"ours_us":27.61,"base_us":22.23,"floor_pct":37.8},{"load":"verify 16","shape":"m=128 o","x":0.596,"ours_us":16.91,"base_us":10.07,"floor_pct":35.4},{"load":"verify 16","shape":"m=128 pooler","x":0.875,"ours_us":33.37,"base_us":29.21,"floor_pct":49.9},{"load":"verify 16","shape":"m=128 unpooler","x":0.783,"ours_us":33.91,"base_us":26.55,"floor_pct":49.1},{"load":"verify 16","shape":"m=128 router","x":2.409,"ours_us":13.17,"base_us":31.72,"floor_pct":32.0},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.697,"ours_us":37.79,"base_us":26.32,"floor_pct":33.7},{"load":"verify 32","shape":"m=256 o","x":0.49,"ours_us":23.13,"base_us":11.33,"floor_pct":29.3},{"load":"verify 32","shape":"m=256 pooler","x":0.857,"ours_us":42.24,"base_us":36.19,"floor_pct":49.9},{"load":"verify 32","shape":"m=256 unpooler","x":0.654,"ours_us":46.25,"base_us":30.26,"floor_pct":45.6},{"load":"verify 32","shape":"m=256 router","x":2.802,"ours_us":17.52,"base_us":49.09,"floor_pct":26.7}]},{"step":2,"x":1.032741648436409,"lo":1.0294165218077636,"hi":1.036834282108699,"pct":47.95129471260756,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.628,"ours_us":29.23,"base_us":18.35,"floor_pct":33.8},{"load":"decode 1","shape":"m=1 o","x":0.665,"ours_us":14.04,"base_us":9.34,"floor_pct":39.2},{"load":"decode 1","shape":"m=1 pooler","x":0.694,"ours_us":40.13,"base_us":27.85,"floor_pct":39.8},{"load":"decode 1","shape":"m=1 unpooler","x":0.708,"ours_us":12.12,"base_us":8.58,"floor_pct":30.9},{"load":"decode 1","shape":"m=1 router","x":1.577,"ours_us":7.75,"base_us":12.22,"floor_pct":48.4},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.75,"ours_us":24.43,"base_us":18.33,"floor_pct":40.7},{"load":"decode 16","shape":"m=16 o","x":0.889,"ours_us":10.64,"base_us":9.46,"floor_pct":52.2},{"load":"decode 16","shape":"m=16 pooler","x":1.026,"ours_us":27.17,"base_us":27.86,"floor_pct":59.1},{"load":"decode 16","shape":"m=16 unpooler","x":0.938,"ours_us":26.96,"base_us":25.3,"floor_pct":59.6},{"load":"decode 16","shape":"m=16 router","x":3.673,"ours_us":7.74,"base_us":28.42,"floor_pct":49.2},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.71,"ours_us":26.08,"base_us":18.53,"floor_pct":38.4},{"load":"decode 32","shape":"m=32 o","x":0.744,"ours_us":12.88,"base_us":9.58,"floor_pct":43.6},{"load":"decode 32","shape":"m=32 pooler","x":0.998,"ours_us":28.61,"base_us":28.54,"floor_pct":56.4},{"load":"decode 32","shape":"m=32 unpooler","x":0.71,"ours_us":35.7,"base_us":25.35,"floor_pct":45.2},{"load":"decode 32","shape":"m=32 router","x":2.416,"ours_us":8.91,"base_us":21.54,"floor_pct":43.3},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.735,"ours_us":25.11,"base_us":18.45,"floor_pct":39.4},{"load":"verify 1","shape":"m=8 o","x":0.834,"ours_us":11.28,"base_us":9.4,"floor_pct":49.0},{"load":"verify 1","shape":"m=8 pooler","x":0.869,"ours_us":32.05,"base_us":27.84,"floor_pct":50.0},{"load":"verify 1","shape":"m=8 unpooler","x":0.786,"ours_us":11.08,"base_us":8.7,"floor_pct":34.0},{"load":"verify 1","shape":"m=8 router","x":3.633,"ours_us":7.83,"base_us":28.44,"floor_pct":48.2},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.961,"ours_us":22.38,"base_us":21.51,"floor_pct":46.6},{"load":"verify 16","shape":"m=128 o","x":0.991,"ours_us":10.1,"base_us":10.0,"floor_pct":59.3},{"load":"verify 16","shape":"m=128 pooler","x":0.933,"ours_us":30.03,"base_us":28.02,"floor_pct":55.4},{"load":"verify 16","shape":"m=128 unpooler","x":1.008,"ours_us":26.76,"base_us":26.97,"floor_pct":62.2},{"load":"verify 16","shape":"m=128 router","x":2.384,"ours_us":13.12,"base_us":31.28,"floor_pct":32.1},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.995,"ours_us":25.91,"base_us":25.78,"floor_pct":49.2},{"load":"verify 32","shape":"m=256 o","x":1.003,"ours_us":11.12,"base_us":11.15,"floor_pct":60.9},{"load":"verify 32","shape":"m=256 pooler","x":1.002,"ours_us":35.78,"base_us":35.84,"floor_pct":58.9},{"load":"verify 32","shape":"m=256 unpooler","x":0.998,"ours_us":30.54,"base_us":30.48,"floor_pct":69.0},{"load":"verify 32","shape":"m=256 router","x":3.321,"ours_us":14.78,"base_us":49.09,"floor_pct":31.6}]},{"step":3,"x":1.1026278366906546,"lo":1.0994195915613383,"hi":1.1069443931236551,"pct":50.864453670131084,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.011,"ours_us":18.26,"base_us":18.46,"floor_pct":54.0},{"load":"decode 1","shape":"m=1 o","x":0.781,"ours_us":12.02,"base_us":9.39,"floor_pct":45.7},{"load":"decode 1","shape":"m=1 pooler","x":0.959,"ours_us":29.01,"base_us":27.81,"floor_pct":55.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.028,"ours_us":8.58,"base_us":8.82,"floor_pct":43.7},{"load":"decode 1","shape":"m=1 router","x":1.525,"ours_us":7.98,"base_us":12.17,"floor_pct":47.0},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.997,"ours_us":18.7,"base_us":18.64,"floor_pct":53.1},{"load":"decode 16","shape":"m=16 o","x":0.821,"ours_us":11.65,"base_us":9.56,"floor_pct":47.7},{"load":"decode 16","shape":"m=16 pooler","x":0.96,"ours_us":28.77,"base_us":27.62,"floor_pct":55.8},{"load":"decode 16","shape":"m=16 unpooler","x":0.884,"ours_us":28.39,"base_us":25.09,"floor_pct":56.6},{"load":"decode 16","shape":"m=16 router","x":3.368,"ours_us":8.47,"base_us":28.54,"floor_pct":44.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.856,"ours_us":22.11,"base_us":18.93,"floor_pct":45.3},{"load":"decode 32","shape":"m=32 o","x":0.751,"ours_us":12.64,"base_us":9.5,"floor_pct":44.4},{"load":"decode 32","shape":"m=32 pooler","x":0.855,"ours_us":33.52,"base_us":28.67,"floor_pct":48.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.804,"ours_us":31.64,"base_us":25.43,"floor_pct":51.0},{"load":"decode 32","shape":"m=32 router","x":2.32,"ours_us":9.38,"base_us":21.75,"floor_pct":41.2},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.003,"ours_us":18.42,"base_us":18.47,"floor_pct":53.7},{"load":"verify 1","shape":"m=8 o","x":0.816,"ours_us":11.54,"base_us":9.42,"floor_pct":47.9},{"load":"verify 1","shape":"m=8 pooler","x":0.95,"ours_us":29.13,"base_us":27.66,"floor_pct":55.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.05,"ours_us":8.5,"base_us":8.92,"floor_pct":44.3},{"load":"verify 1","shape":"m=8 router","x":3.378,"ours_us":8.42,"base_us":28.42,"floor_pct":44.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.003,"ours_us":22.47,"base_us":22.54,"floor_pct":46.4},{"load":"verify 16","shape":"m=128 o","x":0.995,"ours_us":10.14,"base_us":10.09,"floor_pct":59.1},{"load":"verify 16","shape":"m=128 pooler","x":0.98,"ours_us":29.76,"base_us":29.18,"floor_pct":55.9},{"load":"verify 16","shape":"m=128 unpooler","x":0.989,"ours_us":27.36,"base_us":27.05,"floor_pct":60.8},{"load":"verify 16","shape":"m=128 router","x":1.518,"ours_us":20.89,"base_us":31.71,"floor_pct":20.2},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.005,"ours_us":26.15,"base_us":26.3,"floor_pct":48.7},{"load":"verify 32","shape":"m=256 o","x":1.01,"ours_us":11.05,"base_us":11.17,"floor_pct":61.3},{"load":"verify 32","shape":"m=256 pooler","x":1.01,"ours_us":35.46,"base_us":35.82,"floor_pct":59.5},{"load":"verify 32","shape":"m=256 unpooler","x":1.008,"ours_us":30.34,"base_us":30.58,"floor_pct":69.5},{"load":"verify 32","shape":"m=256 router","x":2.273,"ours_us":21.53,"base_us":48.92,"floor_pct":21.7}]},{"step":4,"x":1.189094753555012,"lo":1.1836846440175015,"hi":1.1941468249856573,"pct":54.8355568534774,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.992,"ours_us":18.78,"base_us":18.62,"floor_pct":52.5},{"load":"decode 1","shape":"m=1 o","x":0.989,"ours_us":9.54,"base_us":9.44,"floor_pct":57.6},{"load":"decode 1","shape":"m=1 pooler","x":0.996,"ours_us":27.97,"base_us":27.85,"floor_pct":57.2},{"load":"decode 1","shape":"m=1 unpooler","x":1.001,"ours_us":8.87,"base_us":8.88,"floor_pct":42.3},{"load":"decode 1","shape":"m=1 router","x":1.548,"ours_us":7.86,"base_us":12.17,"floor_pct":47.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.993,"ours_us":18.58,"base_us":18.45,"floor_pct":53.5},{"load":"decode 16","shape":"m=16 o","x":0.974,"ours_us":9.75,"base_us":9.5,"floor_pct":57.0},{"load":"decode 16","shape":"m=16 pooler","x":0.977,"ours_us":28.6,"base_us":27.95,"floor_pct":56.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.009,"ours_us":25.0,"base_us":25.22,"floor_pct":64.3},{"load":"decode 16","shape":"m=16 router","x":3.587,"ours_us":7.97,"base_us":28.58,"floor_pct":47.8},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.006,"ours_us":18.81,"base_us":18.93,"floor_pct":53.2},{"load":"decode 32","shape":"m=32 o","x":1.005,"ours_us":9.54,"base_us":9.59,"floor_pct":58.9},{"load":"decode 32","shape":"m=32 pooler","x":0.999,"ours_us":28.73,"base_us":28.7,"floor_pct":56.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.996,"ours_us":25.25,"base_us":25.15,"floor_pct":63.9},{"load":"decode 32","shape":"m=32 router","x":2.467,"ours_us":8.84,"base_us":21.81,"floor_pct":43.7},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.99,"ours_us":18.61,"base_us":18.42,"floor_pct":53.2},{"load":"verify 1","shape":"m=8 o","x":0.974,"ours_us":9.7,"base_us":9.45,"floor_pct":57.0},{"load":"verify 1","shape":"m=8 pooler","x":0.995,"ours_us":28.06,"base_us":27.92,"floor_pct":57.1},{"load":"verify 1","shape":"m=8 unpooler","x":0.999,"ours_us":8.97,"base_us":8.96,"floor_pct":42.0},{"load":"verify 1","shape":"m=8 router","x":3.607,"ours_us":7.85,"base_us":28.31,"floor_pct":48.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.0,"ours_us":22.3,"base_us":22.3,"floor_pct":46.8},{"load":"verify 16","shape":"m=128 o","x":0.976,"ours_us":10.32,"base_us":10.06,"floor_pct":58.0},{"load":"verify 16","shape":"m=128 pooler","x":1.005,"ours_us":29.48,"base_us":29.64,"floor_pct":56.4},{"load":"verify 16","shape":"m=128 unpooler","x":0.978,"ours_us":27.24,"base_us":26.65,"floor_pct":61.1},{"load":"verify 16","shape":"m=128 router","x":2.392,"ours_us":13.27,"base_us":31.74,"floor_pct":31.7},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.996,"ours_us":25.99,"base_us":25.9,"floor_pct":49.0},{"load":"verify 32","shape":"m=256 o","x":1.02,"ours_us":11.07,"base_us":11.3,"floor_pct":61.2},{"load":"verify 32","shape":"m=256 pooler","x":0.988,"ours_us":36.0,"base_us":35.57,"floor_pct":58.6},{"load":"verify 32","shape":"m=256 unpooler","x":1.004,"ours_us":30.19,"base_us":30.31,"floor_pct":69.8},{"load":"verify 32","shape":"m=256 router","x":3.315,"ours_us":14.82,"base_us":49.14,"floor_pct":31.5}]},{"step":5,"x":1.1281870899630662,"lo":1.1245361878382996,"hi":1.1322792636447223,"pct":51.98746847442128,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.0,"ours_us":18.63,"base_us":18.62,"floor_pct":53.0},{"load":"decode 1","shape":"m=1 o","x":0.738,"ours_us":12.71,"base_us":9.38,"floor_pct":43.3},{"load":"decode 1","shape":"m=1 pooler","x":0.973,"ours_us":28.54,"base_us":27.77,"floor_pct":56.0},{"load":"decode 1","shape":"m=1 unpooler","x":1.016,"ours_us":8.76,"base_us":8.9,"floor_pct":42.8},{"load":"decode 1","shape":"m=1 router","x":1.605,"ours_us":7.55,"base_us":12.11,"floor_pct":49.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.001,"ours_us":18.42,"base_us":18.44,"floor_pct":53.9},{"load":"decode 16","shape":"m=16 o","x":0.825,"ours_us":11.71,"base_us":9.66,"floor_pct":47.5},{"load":"decode 16","shape":"m=16 pooler","x":0.976,"ours_us":28.5,"base_us":27.82,"floor_pct":56.4},{"load":"decode 16","shape":"m=16 unpooler","x":0.914,"ours_us":27.7,"base_us":25.3,"floor_pct":58.0},{"load":"decode 16","shape":"m=16 router","x":3.702,"ours_us":7.82,"base_us":28.96,"floor_pct":48.6},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.84,"ours_us":22.38,"base_us":18.79,"floor_pct":44.7},{"load":"decode 32","shape":"m=32 o","x":0.597,"ours_us":16.16,"base_us":9.64,"floor_pct":34.8},{"load":"decode 32","shape":"m=32 pooler","x":0.855,"ours_us":33.6,"base_us":28.74,"floor_pct":48.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.785,"ours_us":32.19,"base_us":25.26,"floor_pct":50.2},{"load":"decode 32","shape":"m=32 router","x":2.561,"ours_us":8.48,"base_us":21.7,"floor_pct":45.6},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.001,"ours_us":18.82,"base_us":18.84,"floor_pct":52.6},{"load":"verify 1","shape":"m=8 o","x":0.812,"ours_us":11.67,"base_us":9.48,"floor_pct":47.3},{"load":"verify 1","shape":"m=8 pooler","x":0.973,"ours_us":28.56,"base_us":27.8,"floor_pct":56.1},{"load":"verify 1","shape":"m=8 unpooler","x":1.042,"ours_us":8.61,"base_us":8.98,"floor_pct":43.7},{"load":"verify 1","shape":"m=8 router","x":3.7,"ours_us":7.7,"base_us":28.48,"floor_pct":49.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.006,"ours_us":22.14,"base_us":22.29,"floor_pct":47.1},{"load":"verify 16","shape":"m=128 o","x":1.036,"ours_us":9.78,"base_us":10.14,"floor_pct":61.2},{"load":"verify 16","shape":"m=128 pooler","x":0.969,"ours_us":29.92,"base_us":28.98,"floor_pct":55.6},{"load":"verify 16","shape":"m=128 unpooler","x":1.0,"ours_us":26.54,"base_us":26.54,"floor_pct":62.7},{"load":"verify 16","shape":"m=128 router","x":2.38,"ours_us":13.24,"base_us":31.51,"floor_pct":31.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.999,"ours_us":26.25,"base_us":26.22,"floor_pct":48.5},{"load":"verify 32","shape":"m=256 o","x":1.016,"ours_us":11.05,"base_us":11.22,"floor_pct":61.3},{"load":"verify 32","shape":"m=256 pooler","x":1.022,"ours_us":35.56,"base_us":36.36,"floor_pct":59.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.997,"ours_us":30.4,"base_us":30.32,"floor_pct":69.4},{"load":"verify 32","shape":"m=256 router","x":3.33,"ours_us":14.67,"base_us":48.85,"floor_pct":31.9}]},{"step":6,"x":1.186158435555303,"lo":1.181476293787015,"hi":1.1909497740034678,"pct":54.901735792568424,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.988,"ours_us":18.69,"base_us":18.46,"floor_pct":52.8},{"load":"decode 1","shape":"m=1 o","x":0.977,"ours_us":9.55,"base_us":9.34,"floor_pct":57.6},{"load":"decode 1","shape":"m=1 pooler","x":1.002,"ours_us":27.61,"base_us":27.67,"floor_pct":57.9},{"load":"decode 1","shape":"m=1 unpooler","x":0.99,"ours_us":8.85,"base_us":8.76,"floor_pct":42.4},{"load":"decode 1","shape":"m=1 router","x":1.568,"ours_us":7.67,"base_us":12.03,"floor_pct":48.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.983,"ours_us":18.75,"base_us":18.42,"floor_pct":53.0},{"load":"decode 16","shape":"m=16 o","x":0.989,"ours_us":9.6,"base_us":9.5,"floor_pct":57.9},{"load":"decode 16","shape":"m=16 pooler","x":0.991,"ours_us":27.95,"base_us":27.69,"floor_pct":57.5},{"load":"decode 16","shape":"m=16 unpooler","x":1.0,"ours_us":25.18,"base_us":25.18,"floor_pct":63.8},{"load":"decode 16","shape":"m=16 router","x":3.647,"ours_us":7.8,"base_us":28.46,"floor_pct":48.8},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.995,"ours_us":18.91,"base_us":18.82,"floor_pct":52.9},{"load":"decode 32","shape":"m=32 o","x":0.989,"ours_us":9.57,"base_us":9.46,"floor_pct":58.7},{"load":"decode 32","shape":"m=32 pooler","x":0.993,"ours_us":28.75,"base_us":28.55,"floor_pct":56.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.992,"ours_us":25.51,"base_us":25.3,"floor_pct":63.3},{"load":"decode 32","shape":"m=32 router","x":2.526,"ours_us":8.67,"base_us":21.9,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.997,"ours_us":18.66,"base_us":18.6,"floor_pct":53.0},{"load":"verify 1","shape":"m=8 o","x":0.975,"ours_us":9.63,"base_us":9.38,"floor_pct":57.4},{"load":"verify 1","shape":"m=8 pooler","x":0.991,"ours_us":27.84,"base_us":27.6,"floor_pct":57.5},{"load":"verify 1","shape":"m=8 unpooler","x":1.012,"ours_us":8.86,"base_us":8.97,"floor_pct":42.5},{"load":"verify 1","shape":"m=8 router","x":3.588,"ours_us":7.88,"base_us":28.28,"floor_pct":47.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.993,"ours_us":22.39,"base_us":22.25,"floor_pct":46.5},{"load":"verify 16","shape":"m=128 o","x":1.002,"ours_us":10.07,"base_us":10.09,"floor_pct":59.4},{"load":"verify 16","shape":"m=128 pooler","x":0.979,"ours_us":29.93,"base_us":29.31,"floor_pct":55.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.982,"ours_us":27.02,"base_us":26.54,"floor_pct":61.6},{"load":"verify 16","shape":"m=128 router","x":2.374,"ours_us":13.31,"base_us":31.59,"floor_pct":31.6},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.981,"ours_us":26.65,"base_us":26.14,"floor_pct":47.8},{"load":"verify 32","shape":"m=256 o","x":0.996,"ours_us":11.1,"base_us":11.06,"floor_pct":61.0},{"load":"verify 32","shape":"m=256 pooler","x":0.979,"ours_us":36.09,"base_us":35.34,"floor_pct":58.4},{"load":"verify 32","shape":"m=256 unpooler","x":1.007,"ours_us":30.4,"base_us":30.61,"floor_pct":69.4},{"load":"verify 32","shape":"m=256 router","x":3.295,"ours_us":14.9,"base_us":49.1,"floor_pct":31.4}]}],"confirm":{"step":null,"x":1.1960171254275809,"lo":1.19129136361941,"hi":1.2000155886855388,"pct":55.18912858116126,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.966,"ours_us":19.03,"base_us":18.38,"floor_pct":51.9},{"load":"decode 1","shape":"m=1 o","x":1.016,"ours_us":9.24,"base_us":9.39,"floor_pct":59.5},{"load":"decode 1","shape":"m=1 pooler","x":1.0,"ours_us":28.02,"base_us":28.02,"floor_pct":57.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.008,"ours_us":8.82,"base_us":8.89,"floor_pct":42.5},{"load":"decode 1","shape":"m=1 router","x":1.595,"ours_us":7.57,"base_us":12.08,"floor_pct":49.5},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.991,"ours_us":18.66,"base_us":18.49,"floor_pct":53.2},{"load":"decode 16","shape":"m=16 o","x":1.003,"ours_us":9.51,"base_us":9.54,"floor_pct":58.4},{"load":"decode 16","shape":"m=16 pooler","x":0.977,"ours_us":28.41,"base_us":27.77,"floor_pct":56.5},{"load":"decode 16","shape":"m=16 unpooler","x":1.003,"ours_us":25.2,"base_us":25.28,"floor_pct":63.7},{"load":"decode 16","shape":"m=16 router","x":3.664,"ours_us":7.77,"base_us":28.46,"floor_pct":49.0},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.013,"ours_us":18.51,"base_us":18.74,"floor_pct":54.1},{"load":"decode 32","shape":"m=32 o","x":1.007,"ours_us":9.52,"base_us":9.58,"floor_pct":59.0},{"load":"decode 32","shape":"m=32 pooler","x":1.012,"ours_us":28.15,"base_us":28.5,"floor_pct":57.4},{"load":"decode 32","shape":"m=32 unpooler","x":1.006,"ours_us":25.25,"base_us":25.39,"floor_pct":63.9},{"load":"decode 32","shape":"m=32 router","x":2.523,"ours_us":8.65,"base_us":21.84,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.988,"ours_us":18.81,"base_us":18.58,"floor_pct":52.6},{"load":"verify 1","shape":"m=8 o","x":1.025,"ours_us":9.32,"base_us":9.56,"floor_pct":59.3},{"load":"verify 1","shape":"m=8 pooler","x":1.003,"ours_us":27.63,"base_us":27.73,"floor_pct":58.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.048,"ours_us":8.63,"base_us":9.05,"floor_pct":43.6},{"load":"verify 1","shape":"m=8 router","x":3.603,"ours_us":7.86,"base_us":28.3,"floor_pct":48.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.014,"ours_us":22.01,"base_us":22.33,"floor_pct":47.4},{"load":"verify 16","shape":"m=128 o","x":0.994,"ours_us":10.12,"base_us":10.06,"floor_pct":59.1},{"load":"verify 16","shape":"m=128 pooler","x":1.003,"ours_us":29.44,"base_us":29.54,"floor_pct":56.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.995,"ours_us":26.82,"base_us":26.7,"floor_pct":62.0},{"load":"verify 16","shape":"m=128 router","x":2.396,"ours_us":13.24,"base_us":31.73,"floor_pct":31.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.996,"ours_us":26.24,"base_us":26.14,"floor_pct":48.5},{"load":"verify 32","shape":"m=256 o","x":1.006,"ours_us":11.15,"base_us":11.22,"floor_pct":60.7},{"load":"verify 32","shape":"m=256 pooler","x":0.982,"ours_us":36.17,"base_us":35.51,"floor_pct":58.3},{"load":"verify 32","shape":"m=256 unpooler","x":1.0,"ours_us":30.33,"base_us":30.34,"floor_pct":69.5},{"load":"verify 32","shape":"m=256 router","x":3.336,"ours_us":14.69,"base_us":49.02,"floor_pct":31.8}]},"src":"bench/product/runs/learnability/bf16-A.jsonl"},{"id":"bf16-B","model":"Claude Opus 5.5","harness":"Claude Code","family":"dense_proj_bf16","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":50.745616920078966},"T2":{"x":1.4537769552831892,"pct":67.0661895091268},"T3":{"x":1.9075539105663786,"pct":88.0},"T4":{"x":1.9509074085337963,"pct":90.0}},"steps":[{"step":1,"x":1.030659253796212,"lo":1.0265650718073422,"hi":1.0343415945974217,"pct":47.389640936441815,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.879,"ours_us":21.6,"base_us":18.98,"floor_pct":45.7},{"load":"decode 1","shape":"m=1 o","x":0.931,"ours_us":10.51,"base_us":9.79,"floor_pct":52.3},{"load":"decode 1","shape":"m=1 pooler","x":0.934,"ours_us":30.03,"base_us":28.04,"floor_pct":53.2},{"load":"decode 1","shape":"m=1 unpooler","x":0.99,"ours_us":8.87,"base_us":8.78,"floor_pct":42.3},{"load":"decode 1","shape":"m=1 router","x":1.245,"ours_us":9.79,"base_us":12.19,"floor_pct":38.3},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.839,"ours_us":22.06,"base_us":18.5,"floor_pct":45.0},{"load":"decode 16","shape":"m=16 o","x":1.032,"ours_us":9.39,"base_us":9.69,"floor_pct":59.2},{"load":"decode 16","shape":"m=16 pooler","x":0.925,"ours_us":29.98,"base_us":27.74,"floor_pct":53.6},{"load":"decode 16","shape":"m=16 unpooler","x":0.855,"ours_us":29.41,"base_us":25.15,"floor_pct":54.6},{"load":"decode 16","shape":"m=16 router","x":2.8,"ours_us":10.23,"base_us":28.66,"floor_pct":37.2},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.889,"ours_us":21.03,"base_us":18.69,"floor_pct":47.6},{"load":"decode 32","shape":"m=32 o","x":0.787,"ours_us":12.14,"base_us":9.56,"floor_pct":46.3},{"load":"decode 32","shape":"m=32 pooler","x":0.986,"ours_us":28.96,"base_us":28.55,"floor_pct":55.8},{"load":"decode 32","shape":"m=32 unpooler","x":0.891,"ours_us":28.44,"base_us":25.34,"floor_pct":56.8},{"load":"decode 32","shape":"m=32 router","x":2.022,"ours_us":10.9,"base_us":22.04,"floor_pct":35.4},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.853,"ours_us":21.91,"base_us":18.68,"floor_pct":45.2},{"load":"verify 1","shape":"m=8 o","x":0.976,"ours_us":9.7,"base_us":9.46,"floor_pct":57.0},{"load":"verify 1","shape":"m=8 pooler","x":0.941,"ours_us":29.9,"base_us":28.14,"floor_pct":53.6},{"load":"verify 1","shape":"m=8 unpooler","x":1.015,"ours_us":8.84,"base_us":8.97,"floor_pct":42.6},{"load":"verify 1","shape":"m=8 router","x":2.792,"ours_us":10.18,"base_us":28.43,"floor_pct":37.1},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.825,"ours_us":27.47,"base_us":22.65,"floor_pct":38.0},{"load":"verify 16","shape":"m=128 o","x":0.707,"ours_us":14.28,"base_us":10.09,"floor_pct":41.9},{"load":"verify 16","shape":"m=128 pooler","x":0.9,"ours_us":32.43,"base_us":29.18,"floor_pct":51.3},{"load":"verify 16","shape":"m=128 unpooler","x":0.83,"ours_us":32.29,"base_us":26.81,"floor_pct":51.5},{"load":"verify 16","shape":"m=128 router","x":2.341,"ours_us":13.62,"base_us":31.89,"floor_pct":30.9},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.683,"ours_us":37.91,"base_us":25.91,"floor_pct":33.6},{"load":"verify 32","shape":"m=256 o","x":0.572,"ours_us":19.74,"base_us":11.28,"floor_pct":34.3},{"load":"verify 32","shape":"m=256 pooler","x":0.839,"ours_us":42.3,"base_us":35.5,"floor_pct":49.9},{"load":"verify 32","shape":"m=256 unpooler","x":0.705,"ours_us":42.77,"base_us":30.15,"floor_pct":49.3},{"load":"verify 32","shape":"m=256 router","x":3.027,"ours_us":16.21,"base_us":49.08,"floor_pct":28.8}]},{"step":2,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"GRADER_BUG_REFUNDED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.889,"ours_us":21.2,"base_us":18.85,"floor_pct":46.6},{"load":"decode 1","shape":"m=1 o","x":0.933,"ours_us":9.96,"base_us":9.3,"floor_pct":55.2},{"load":"decode 1","shape":"m=1 pooler","x":1.019,"ours_us":27.37,"base_us":27.9,"floor_pct":58.4},{"load":"decode 1","shape":"m=1 unpooler","x":1.039,"ours_us":8.5,"base_us":8.83,"floor_pct":44.1},{"load":"decode 1","shape":"m=1 router","x":1.054,"ours_us":11.47,"base_us":12.09,"floor_pct":32.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.869,"ours_us":21.34,"base_us":18.54,"floor_pct":46.5},{"load":"decode 16","shape":"m=16 o","x":1.005,"ours_us":9.45,"base_us":9.5,"floor_pct":58.8},{"load":"decode 16","shape":"m=16 pooler","x":1.042,"ours_us":27.06,"base_us":28.21,"floor_pct":59.4},{"load":"decode 16","shape":"m=16 unpooler","x":0.999,"ours_us":25.7,"base_us":25.67,"floor_pct":62.5},{"load":"decode 16","shape":"m=16 router","x":2.359,"ours_us":12.17,"base_us":28.71,"floor_pct":31.3},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.855,"ours_us":21.76,"base_us":18.61,"floor_pct":46.0},{"load":"decode 32","shape":"m=32 o","x":0.935,"ours_us":10.14,"base_us":9.48,"floor_pct":55.4},{"load":"decode 32","shape":"m=32 pooler","x":1.023,"ours_us":27.84,"base_us":28.5,"floor_pct":58.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.962,"ours_us":26.24,"base_us":25.23,"floor_pct":61.5},{"load":"decode 32","shape":"m=32 router","x":1.737,"ours_us":12.6,"base_us":21.9,"floor_pct":30.7},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.872,"ours_us":21.53,"base_us":18.77,"floor_pct":46.0},{"load":"verify 1","shape":"m=8 o","x":0.978,"ours_us":9.66,"base_us":9.45,"floor_pct":57.2},{"load":"verify 1","shape":"m=8 pooler","x":1.017,"ours_us":27.44,"base_us":27.9,"floor_pct":58.4},{"load":"verify 1","shape":"m=8 unpooler","x":1.028,"ours_us":8.65,"base_us":8.9,"floor_pct":43.5},{"load":"verify 1","shape":"m=8 router","x":2.356,"ours_us":12.1,"base_us":28.52,"floor_pct":31.2},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.868,"ours_us":25.39,"base_us":22.02,"floor_pct":41.1},{"load":"verify 16","shape":"m=128 o","x":0.899,"ours_us":11.23,"base_us":10.1,"floor_pct":53.3},{"load":"verify 16","shape":"m=128 pooler","x":0.811,"ours_us":35.3,"base_us":28.62,"floor_pct":47.1},{"load":"verify 16","shape":"m=128 unpooler","x":0.804,"ours_us":33.21,"base_us":26.7,"floor_pct":50.1},{"load":"verify 16","shape":"m=128 router","x":0.511,"ours_us":62.32,"base_us":31.85,"floor_pct":6.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.766,"ours_us":33.63,"base_us":25.76,"floor_pct":37.9},{"load":"verify 32","shape":"m=256 o","x":0.822,"ours_us":13.51,"base_us":11.1,"floor_pct":50.1},{"load":"verify 32","shape":"m=256 pooler","x":0.753,"ours_us":46.62,"base_us":35.09,"floor_pct":45.2}]},{"step":3,"x":1.0743784622576087,"lo":1.069118162973861,"hi":1.0777371380506438,"pct":49.73949983798279,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.806,"ours_us":23.03,"base_us":18.55,"floor_pct":42.9},{"load":"decode 1","shape":"m=1 o","x":0.864,"ours_us":10.76,"base_us":9.3,"floor_pct":51.1},{"load":"decode 1","shape":"m=1 pooler","x":1.013,"ours_us":27.84,"base_us":28.21,"floor_pct":57.4},{"load":"decode 1","shape":"m=1 unpooler","x":0.916,"ours_us":9.53,"base_us":8.74,"floor_pct":39.3},{"load":"decode 1","shape":"m=1 router","x":1.276,"ours_us":9.45,"base_us":12.06,"floor_pct":39.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.997,"ours_us":18.5,"base_us":18.44,"floor_pct":53.7},{"load":"decode 16","shape":"m=16 o","x":0.88,"ours_us":10.76,"base_us":9.47,"floor_pct":51.7},{"load":"decode 16","shape":"m=16 pooler","x":1.015,"ours_us":27.34,"base_us":27.74,"floor_pct":58.8},{"load":"decode 16","shape":"m=16 unpooler","x":0.94,"ours_us":26.84,"base_us":25.24,"floor_pct":59.9},{"load":"decode 16","shape":"m=16 router","x":2.912,"ours_us":9.78,"base_us":28.48,"floor_pct":38.9},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.967,"ours_us":19.12,"base_us":18.5,"floor_pct":52.3},{"load":"decode 32","shape":"m=32 o","x":0.813,"ours_us":11.69,"base_us":9.5,"floor_pct":48.1},{"load":"decode 32","shape":"m=32 pooler","x":1.017,"ours_us":27.76,"base_us":28.23,"floor_pct":58.2},{"load":"decode 32","shape":"m=32 unpooler","x":0.914,"ours_us":27.95,"base_us":25.55,"floor_pct":57.8},{"load":"decode 32","shape":"m=32 router","x":2.042,"ours_us":10.63,"base_us":21.7,"floor_pct":36.4},{"load":"verify 1","shape":"m=8 qkv_gate","x":1.016,"ours_us":18.14,"base_us":18.43,"floor_pct":54.6},{"load":"verify 1","shape":"m=8 o","x":0.859,"ours_us":10.82,"base_us":9.3,"floor_pct":51.1},{"load":"verify 1","shape":"m=8 pooler","x":1.017,"ours_us":27.31,"base_us":27.79,"floor_pct":58.7},{"load":"verify 1","shape":"m=8 unpooler","x":0.922,"ours_us":9.47,"base_us":8.73,"floor_pct":39.8},{"load":"verify 1","shape":"m=8 router","x":2.907,"ours_us":9.74,"base_us":28.32,"floor_pct":38.8},{"load":"verify 16","shape":"m=128 qkv_gate","x":0.99,"ours_us":22.46,"base_us":22.24,"floor_pct":46.4},{"load":"verify 16","shape":"m=128 o","x":0.852,"ours_us":11.65,"base_us":9.93,"floor_pct":51.4},{"load":"verify 16","shape":"m=128 pooler","x":0.738,"ours_us":39.81,"base_us":29.39,"floor_pct":41.8},{"load":"verify 16","shape":"m=128 unpooler","x":0.742,"ours_us":36.47,"base_us":27.05,"floor_pct":45.6},{"load":"verify 16","shape":"m=128 router","x":2.484,"ours_us":12.78,"base_us":31.74,"floor_pct":33.0},{"load":"verify 32","shape":"m=256 qkv_gate","x":0.833,"ours_us":31.1,"base_us":25.9,"floor_pct":41.0},{"load":"verify 32","shape":"m=256 o","x":0.855,"ours_us":12.83,"base_us":10.98,"floor_pct":52.8},{"load":"verify 32","shape":"m=256 pooler","x":0.825,"ours_us":42.81,"base_us":35.33,"floor_pct":49.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.788,"ours_us":38.12,"base_us":30.04,"floor_pct":55.3},{"load":"verify 32","shape":"m=256 router","x":3.16,"ours_us":15.49,"base_us":48.94,"floor_pct":30.2}]},{"step":3,"x":1.1825872259387007,"lo":1.1783998448096016,"hi":1.1860017601733677,"pct":54.903853032670725,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.002,"ours_us":18.42,"base_us":18.46,"floor_pct":53.6},{"load":"decode 1","shape":"m=1 o","x":0.982,"ours_us":9.39,"base_us":9.22,"floor_pct":58.5},{"load":"decode 1","shape":"m=1 pooler","x":1.031,"ours_us":27.06,"base_us":27.9,"floor_pct":59.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.039,"ours_us":8.34,"base_us":8.67,"floor_pct":44.9},{"load":"decode 1","shape":"m=1 router","x":1.52,"ours_us":8.03,"base_us":12.21,"floor_pct":46.7},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.994,"ours_us":18.43,"base_us":18.32,"floor_pct":53.9},{"load":"decode 16","shape":"m=16 o","x":0.992,"ours_us":9.43,"base_us":9.36,"floor_pct":58.9},{"load":"decode 16","shape":"m=16 pooler","x":1.026,"ours_us":27.19,"base_us":27.89,"floor_pct":59.1},{"load":"decode 16","shape":"m=16 unpooler","x":1.006,"ours_us":24.96,"base_us":25.12,"floor_pct":64.4},{"load":"decode 16","shape":"m=16 router","x":3.446,"ours_us":8.23,"base_us":28.36,"floor_pct":46.2},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.992,"ours_us":18.81,"base_us":18.66,"floor_pct":53.2},{"load":"decode 32","shape":"m=32 o","x":1.003,"ours_us":9.38,"base_us":9.42,"floor_pct":59.9},{"load":"decode 32","shape":"m=32 pooler","x":0.991,"ours_us":28.35,"base_us":28.09,"floor_pct":56.9},{"load":"decode 32","shape":"m=32 unpooler","x":1.001,"ours_us":25.25,"base_us":25.26,"floor_pct":63.9},{"load":"decode 32","shape":"m=32 router","x":2.339,"ours_us":9.24,"base_us":21.62,"floor_pct":41.8},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.993,"ours_us":18.66,"base_us":18.52,"floor_pct":53.1},{"load":"verify 1","shape":"m=8 o","x":0.979,"ours_us":9.47,"base_us":9.27,"floor_pct":58.3},{"load":"verify 1","shape":"m=8 pooler","x":1.026,"ours_us":26.95,"base_us":27.66,"floor_pct":59.4},{"load":"verify 1","shape":"m=8 unpooler","x":1.025,"ours_us":8.48,"base_us":8.69,"floor_pct":44.4},{"load":"verify 1","shape":"m=8 router","x":3.493,"ours_us":8.16,"base_us":28.5,"floor_pct":46.3},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.009,"ours_us":22.1,"base_us":22.3,"floor_pct":47.2},{"load":"verify 16","shape":"m=128 o","x":0.962,"ours_us":10.15,"base_us":9.77,"floor_pct":59.0},{"load":"verify 16","shape":"m=128 pooler","x":1.027,"ours_us":28.65,"base_us":29.42,"floor_pct":58.1},{"load":"verify 16","shape":"m=128 unpooler","x":0.99,"ours_us":26.76,"base_us":26.5,"floor_pct":62.2},{"load":"verify 16","shape":"m=128 router","x":1.75,"ours_us":17.84,"base_us":31.22,"floor_pct":23.6},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.016,"ours_us":26.26,"base_us":26.67,"floor_pct":48.5},{"load":"verify 32","shape":"m=256 o","x":0.993,"ours_us":11.34,"base_us":11.26,"floor_pct":59.7},{"load":"verify 32","shape":"m=256 pooler","x":0.993,"ours_us":35.58,"base_us":35.34,"floor_pct":59.3},{"load":"verify 32","shape":"m=256 unpooler","x":0.989,"ours_us":30.18,"base_us":29.86,"floor_pct":69.9},{"load":"verify 32","shape":"m=256 router","x":2.668,"ours_us":18.25,"base_us":48.7,"floor_pct":25.6}]},{"step":4,"x":1.0607173969012014,"lo":1.0574001903515557,"hi":1.0645411422858684,"pct":49.134311414936235,"tier":null,"verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.694,"ours_us":26.75,"base_us":18.58,"floor_pct":36.9},{"load":"decode 1","shape":"m=1 o","x":0.972,"ours_us":9.69,"base_us":9.42,"floor_pct":56.7},{"load":"decode 1","shape":"m=1 pooler","x":1.015,"ours_us":27.17,"base_us":27.58,"floor_pct":58.8},{"load":"decode 1","shape":"m=1 unpooler","x":1.045,"ours_us":8.25,"base_us":8.62,"floor_pct":45.5},{"load":"decode 1","shape":"m=1 router","x":0.888,"ours_us":13.79,"base_us":12.24,"floor_pct":27.2},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.979,"ours_us":18.78,"base_us":18.38,"floor_pct":52.9},{"load":"decode 16","shape":"m=16 o","x":0.992,"ours_us":9.42,"base_us":9.34,"floor_pct":59.0},{"load":"decode 16","shape":"m=16 pooler","x":1.033,"ours_us":27.13,"base_us":28.02,"floor_pct":59.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.008,"ours_us":25.02,"base_us":25.23,"floor_pct":64.2},{"load":"decode 16","shape":"m=16 router","x":3.021,"ours_us":9.42,"base_us":28.46,"floor_pct":40.4},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.891,"ours_us":21.06,"base_us":18.76,"floor_pct":47.5},{"load":"decode 32","shape":"m=32 o","x":0.942,"ours_us":10.04,"base_us":9.46,"floor_pct":55.9},{"load":"decode 32","shape":"m=32 pooler","x":0.999,"ours_us":28.31,"base_us":28.27,"floor_pct":57.0},{"load":"decode 32","shape":"m=32 unpooler","x":0.973,"ours_us":26.5,"base_us":25.78,"floor_pct":60.9},{"load":"decode 32","shape":"m=32 router","x":1.338,"ours_us":16.09,"base_us":21.53,"floor_pct":24.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.995,"ours_us":18.58,"base_us":18.49,"floor_pct":53.3},{"load":"verify 1","shape":"m=8 o","x":0.986,"ours_us":9.51,"base_us":9.38,"floor_pct":58.1},{"load":"verify 1","shape":"m=8 pooler","x":1.036,"ours_us":27.02,"base_us":27.99,"floor_pct":59.3},{"load":"verify 1","shape":"m=8 unpooler","x":1.039,"ours_us":8.45,"base_us":8.78,"floor_pct":44.6},{"load":"verify 1","shape":"m=8 router","x":3.391,"ours_us":8.47,"base_us":28.72,"floor_pct":44.6},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.0,"ours_us":22.3,"base_us":22.3,"floor_pct":46.8},{"load":"verify 16","shape":"m=128 o","x":0.986,"ours_us":10.11,"base_us":9.98,"floor_pct":59.2},{"load":"verify 16","shape":"m=128 pooler","x":0.955,"ours_us":29.67,"base_us":28.34,"floor_pct":56.1},{"load":"verify 16","shape":"m=128 unpooler","x":0.995,"ours_us":26.47,"base_us":26.34,"floor_pct":62.8},{"load":"verify 16","shape":"m=128 router","x":2.47,"ours_us":12.79,"base_us":31.59,"floor_pct":32.9},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.015,"ours_us":26.15,"base_us":26.54,"floor_pct":48.7},{"load":"verify 32","shape":"m=256 o","x":1.01,"ours_us":11.05,"base_us":11.16,"floor_pct":61.3},{"load":"verify 32","shape":"m=256 pooler","x":0.979,"ours_us":35.96,"base_us":35.21,"floor_pct":58.6},{"load":"verify 32","shape":"m=256 unpooler","x":0.997,"ours_us":30.42,"base_us":30.32,"floor_pct":69.3},{"load":"verify 32","shape":"m=256 router","x":0.646,"ours_us":75.5,"base_us":48.76,"floor_pct":6.2}]},{"step":5,"x":1.2052085469648746,"lo":1.2007971961070867,"hi":1.2089999867474306,"pct":55.47049897364123,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.001,"ours_us":18.7,"base_us":18.72,"floor_pct":52.8},{"load":"decode 1","shape":"m=1 o","x":1.014,"ours_us":9.35,"base_us":9.48,"floor_pct":58.8},{"load":"decode 1","shape":"m=1 pooler","x":1.018,"ours_us":27.61,"base_us":28.11,"floor_pct":57.9},{"load":"decode 1","shape":"m=1 unpooler","x":1.055,"ours_us":8.35,"base_us":8.81,"floor_pct":44.9},{"load":"decode 1","shape":"m=1 router","x":1.511,"ours_us":8.15,"base_us":12.32,"floor_pct":46.0},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.009,"ours_us":18.4,"base_us":18.57,"floor_pct":54.0},{"load":"decode 16","shape":"m=16 o","x":0.977,"ours_us":9.68,"base_us":9.46,"floor_pct":57.4},{"load":"decode 16","shape":"m=16 pooler","x":1.017,"ours_us":27.48,"base_us":27.94,"floor_pct":58.5},{"load":"decode 16","shape":"m=16 unpooler","x":1.011,"ours_us":25.04,"base_us":25.31,"floor_pct":64.2},{"load":"decode 16","shape":"m=16 router","x":3.347,"ours_us":8.55,"base_us":28.61,"floor_pct":44.5},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.998,"ours_us":18.9,"base_us":18.86,"floor_pct":52.9},{"load":"decode 32","shape":"m=32 o","x":1.014,"ours_us":9.44,"base_us":9.58,"floor_pct":59.5},{"load":"decode 32","shape":"m=32 pooler","x":0.987,"ours_us":28.6,"base_us":28.23,"floor_pct":56.5},{"load":"decode 32","shape":"m=32 unpooler","x":1.012,"ours_us":25.4,"base_us":25.7,"floor_pct":63.6},{"load":"decode 32","shape":"m=32 router","x":2.301,"ours_us":9.43,"base_us":21.69,"floor_pct":41.0},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.987,"ours_us":19.01,"base_us":18.75,"floor_pct":52.1},{"load":"verify 1","shape":"m=8 o","x":1.014,"ours_us":9.2,"base_us":9.33,"floor_pct":60.1},{"load":"verify 1","shape":"m=8 pooler","x":1.019,"ours_us":27.27,"base_us":27.78,"floor_pct":58.7},{"load":"verify 1","shape":"m=8 unpooler","x":1.038,"ours_us":8.59,"base_us":8.92,"floor_pct":43.8},{"load":"verify 1","shape":"m=8 router","x":3.398,"ours_us":8.45,"base_us":28.71,"floor_pct":44.7},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.013,"ours_us":22.19,"base_us":22.49,"floor_pct":47.0},{"load":"verify 16","shape":"m=128 o","x":1.0,"ours_us":10.03,"base_us":10.03,"floor_pct":59.7},{"load":"verify 16","shape":"m=128 pooler","x":1.041,"ours_us":28.38,"base_us":29.54,"floor_pct":58.6},{"load":"verify 16","shape":"m=128 unpooler","x":0.971,"ours_us":27.3,"base_us":26.5,"floor_pct":60.9},{"load":"verify 16","shape":"m=128 router","x":2.672,"ours_us":11.76,"base_us":31.41,"floor_pct":35.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.01,"ours_us":26.35,"base_us":26.62,"floor_pct":48.3},{"load":"verify 32","shape":"m=256 o","x":0.996,"ours_us":11.38,"base_us":11.34,"floor_pct":59.5},{"load":"verify 32","shape":"m=256 pooler","x":1.009,"ours_us":35.39,"base_us":35.72,"floor_pct":59.6},{"load":"verify 32","shape":"m=256 unpooler","x":1.006,"ours_us":29.96,"base_us":30.14,"floor_pct":70.4},{"load":"verify 32","shape":"m=256 router","x":3.777,"ours_us":12.97,"base_us":48.98,"floor_pct":36.1}]},{"step":6,"x":1.2029951274902166,"lo":1.1989752184798623,"hi":1.20686099458205,"pct":55.82624488702238,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":0.996,"ours_us":18.45,"base_us":18.38,"floor_pct":53.5},{"load":"decode 1","shape":"m=1 o","x":1.003,"ours_us":9.27,"base_us":9.3,"floor_pct":59.3},{"load":"decode 1","shape":"m=1 pooler","x":0.993,"ours_us":27.76,"base_us":27.58,"floor_pct":57.6},{"load":"decode 1","shape":"m=1 unpooler","x":1.03,"ours_us":8.41,"base_us":8.66,"floor_pct":44.6},{"load":"decode 1","shape":"m=1 router","x":1.526,"ours_us":7.99,"base_us":12.2,"floor_pct":46.9},{"load":"decode 16","shape":"m=16 qkv_gate","x":0.999,"ours_us":18.46,"base_us":18.45,"floor_pct":53.8},{"load":"decode 16","shape":"m=16 o","x":0.987,"ours_us":9.4,"base_us":9.27,"floor_pct":59.1},{"load":"decode 16","shape":"m=16 pooler","x":1.011,"ours_us":27.61,"base_us":27.91,"floor_pct":58.2},{"load":"decode 16","shape":"m=16 unpooler","x":1.0,"ours_us":25.04,"base_us":25.04,"floor_pct":64.1},{"load":"decode 16","shape":"m=16 router","x":3.441,"ours_us":8.28,"base_us":28.47,"floor_pct":46.0},{"load":"decode 32","shape":"m=32 qkv_gate","x":0.992,"ours_us":18.66,"base_us":18.51,"floor_pct":53.6},{"load":"decode 32","shape":"m=32 o","x":0.991,"ours_us":9.47,"base_us":9.38,"floor_pct":59.4},{"load":"decode 32","shape":"m=32 pooler","x":0.998,"ours_us":28.46,"base_us":28.39,"floor_pct":56.7},{"load":"decode 32","shape":"m=32 unpooler","x":1.005,"ours_us":25.03,"base_us":25.15,"floor_pct":64.5},{"load":"decode 32","shape":"m=32 router","x":2.272,"ours_us":9.46,"base_us":21.5,"floor_pct":40.8},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.992,"ours_us":18.76,"base_us":18.6,"floor_pct":52.8},{"load":"verify 1","shape":"m=8 o","x":0.998,"ours_us":9.27,"base_us":9.26,"floor_pct":59.6},{"load":"verify 1","shape":"m=8 pooler","x":1.008,"ours_us":27.14,"base_us":27.37,"floor_pct":59.0},{"load":"verify 1","shape":"m=8 unpooler","x":1.041,"ours_us":8.42,"base_us":8.76,"floor_pct":44.7},{"load":"verify 1","shape":"m=8 router","x":3.541,"ours_us":8.08,"base_us":28.62,"floor_pct":46.7},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.033,"ours_us":21.84,"base_us":22.58,"floor_pct":47.7},{"load":"verify 16","shape":"m=128 o","x":0.989,"ours_us":9.91,"base_us":9.81,"floor_pct":60.4},{"load":"verify 16","shape":"m=128 pooler","x":1.042,"ours_us":28.21,"base_us":29.41,"floor_pct":59.0},{"load":"verify 16","shape":"m=128 unpooler","x":1.001,"ours_us":26.93,"base_us":26.97,"floor_pct":61.8},{"load":"verify 16","shape":"m=128 router","x":2.695,"ours_us":11.58,"base_us":31.22,"floor_pct":36.4},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.024,"ours_us":26.07,"base_us":26.7,"floor_pct":48.9},{"load":"verify 32","shape":"m=256 o","x":1.012,"ours_us":11.02,"base_us":11.16,"floor_pct":61.4},{"load":"verify 32","shape":"m=256 pooler","x":0.986,"ours_us":35.5,"base_us":35.01,"floor_pct":59.4},{"load":"verify 32","shape":"m=256 unpooler","x":0.999,"ours_us":30.49,"base_us":30.47,"floor_pct":69.2},{"load":"verify 32","shape":"m=256 router","x":3.726,"ours_us":13.08,"base_us":48.74,"floor_pct":35.7}]}],"confirm":{"step":null,"x":1.2071974363501476,"lo":1.2026095788030222,"hi":1.2117470589752926,"pct":55.69088968356908,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"m=1 qkv_gate","x":1.0,"ours_us":18.61,"base_us":18.61,"floor_pct":53.0},{"load":"decode 1","shape":"m=1 o","x":1.012,"ours_us":9.33,"base_us":9.44,"floor_pct":59.0},{"load":"decode 1","shape":"m=1 pooler","x":1.013,"ours_us":27.52,"base_us":27.87,"floor_pct":58.1},{"load":"decode 1","shape":"m=1 unpooler","x":1.024,"ours_us":8.36,"base_us":8.56,"floor_pct":44.9},{"load":"decode 1","shape":"m=1 router","x":1.521,"ours_us":8.01,"base_us":12.18,"floor_pct":46.8},{"load":"decode 16","shape":"m=16 qkv_gate","x":1.006,"ours_us":18.51,"base_us":18.62,"floor_pct":53.7},{"load":"decode 16","shape":"m=16 o","x":0.989,"ours_us":9.47,"base_us":9.36,"floor_pct":58.7},{"load":"decode 16","shape":"m=16 pooler","x":1.023,"ours_us":27.33,"base_us":27.97,"floor_pct":58.8},{"load":"decode 16","shape":"m=16 unpooler","x":1.015,"ours_us":24.99,"base_us":25.35,"floor_pct":64.3},{"load":"decode 16","shape":"m=16 router","x":3.451,"ours_us":8.28,"base_us":28.58,"floor_pct":46.0},{"load":"decode 32","shape":"m=32 qkv_gate","x":1.011,"ours_us":18.71,"base_us":18.92,"floor_pct":53.5},{"load":"decode 32","shape":"m=32 o","x":0.999,"ours_us":9.55,"base_us":9.54,"floor_pct":58.8},{"load":"decode 32","shape":"m=32 pooler","x":0.998,"ours_us":28.52,"base_us":28.46,"floor_pct":56.6},{"load":"decode 32","shape":"m=32 unpooler","x":0.998,"ours_us":25.37,"base_us":25.31,"floor_pct":63.6},{"load":"decode 32","shape":"m=32 router","x":2.286,"ours_us":9.45,"base_us":21.62,"floor_pct":40.9},{"load":"verify 1","shape":"m=8 qkv_gate","x":0.994,"ours_us":18.77,"base_us":18.66,"floor_pct":52.7},{"load":"verify 1","shape":"m=8 o","x":0.991,"ours_us":9.51,"base_us":9.42,"floor_pct":58.1},{"load":"verify 1","shape":"m=8 pooler","x":1.017,"ours_us":27.37,"base_us":27.82,"floor_pct":58.5},{"load":"verify 1","shape":"m=8 unpooler","x":1.023,"ours_us":8.52,"base_us":8.71,"floor_pct":44.2},{"load":"verify 1","shape":"m=8 router","x":3.541,"ours_us":8.06,"base_us":28.53,"floor_pct":46.9},{"load":"verify 16","shape":"m=128 qkv_gate","x":1.012,"ours_us":22.37,"base_us":22.64,"floor_pct":46.6},{"load":"verify 16","shape":"m=128 o","x":1.001,"ours_us":9.99,"base_us":10.0,"floor_pct":59.9},{"load":"verify 16","shape":"m=128 pooler","x":1.037,"ours_us":28.42,"base_us":29.46,"floor_pct":58.5},{"load":"verify 16","shape":"m=128 unpooler","x":0.996,"ours_us":27.03,"base_us":26.94,"floor_pct":61.5},{"load":"verify 16","shape":"m=128 router","x":2.662,"ours_us":11.76,"base_us":31.31,"floor_pct":35.8},{"load":"verify 32","shape":"m=256 qkv_gate","x":1.023,"ours_us":25.96,"base_us":26.55,"floor_pct":49.1},{"load":"verify 32","shape":"m=256 o","x":0.996,"ours_us":11.2,"base_us":11.16,"floor_pct":60.5},{"load":"verify 32","shape":"m=256 pooler","x":1.025,"ours_us":34.94,"base_us":35.81,"floor_pct":60.4},{"load":"verify 32","shape":"m=256 unpooler","x":1.002,"ours_us":30.38,"base_us":30.45,"floor_pct":69.4},{"load":"verify 32","shape":"m=256 router","x":3.769,"ours_us":12.91,"base_us":48.66,"floor_pct":36.2}]},"src":"bench/product/runs/learnability/bf16-B.jsonl"},{"id":"moe-A","model":"Claude Opus 5.5","harness":"Claude Code","family":"moe_nvfp4_ep8","start_x":1.0,"contest_tier":"T1","thresholds":{"T1":{"x":1.1,"pct":0.29684302004018265},"T2":{"x":137.60950654824418,"pct":37.13492864547281},"T3":{"x":274.21901309648837,"pct":74.0},"T4":{"x":333.50961052275613,"pct":90.0}},"steps":[{"step":1,"x":null,"lo":null,"hi":null,"pct":null,"tier":null,"verdict":"GRADER_BUG_REFUNDED","note":null,"cases":[]},{"step":2,"x":67.20141620713015,"lo":67.06128338942725,"hi":67.48985670257125,"pct":18.232731168108717,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":124.403,"ours_us":461.34,"base_us":57392.36,"floor_pct":10.7},{"load":"decode 16","shape":"bs=16 dims=full","x":74.609,"ours_us":774.12,"base_us":57756.93,"floor_pct":24.6},{"load":"decode 32","shape":"bs=32 dims=full","x":107.934,"ours_us":535.36,"base_us":57783.98,"floor_pct":35.6},{"load":"verify 1","shape":"bs=8 dims=full","x":116.33,"ours_us":495.17,"base_us":57602.92,"floor_pct":38.5},{"load":"verify 16","shape":"bs=128 dims=full","x":33.584,"ours_us":1733.92,"base_us":58232.01,"floor_pct":11.0},{"load":"verify 32","shape":"bs=256 dims=full","x":23.532,"ours_us":2503.0,"base_us":58901.0,"floor_pct":9.2}]},{"step":2,"x":161.3485860462596,"lo":160.99558063679635,"hi":161.8743353641995,"pct":43.6128398435558,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":372.101,"ours_us":154.94,"base_us":57651.95,"floor_pct":31.7},{"load":"decode 16","shape":"bs=16 dims=full","x":229.732,"ours_us":252.26,"base_us":57952.75,"floor_pct":75.6},{"load":"decode 32","shape":"bs=32 dims=full","x":144.704,"ours_us":400.92,"base_us":58014.67,"floor_pct":47.6},{"load":"verify 1","shape":"bs=8 dims=full","x":305.034,"ours_us":189.45,"base_us":57787.7,"floor_pct":100.7},{"load":"verify 16","shape":"bs=128 dims=full","x":84.191,"ours_us":693.98,"base_us":58427.19,"floor_pct":27.5},{"load":"verify 32","shape":"bs=256 dims=full","x":55.541,"ours_us":1064.76,"base_us":59137.48,"floor_pct":21.7}]},{"step":3,"x":214.51761831642062,"lo":213.7071907877432,"hi":215.23277653404043,"pct":58.086168285899596,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":505.878,"ours_us":113.69,"base_us":57513.71,"floor_pct":43.3},{"load":"decode 16","shape":"bs=16 dims=full","x":258.589,"ours_us":223.81,"base_us":57874.9,"floor_pct":85.2},{"load":"decode 32","shape":"bs=32 dims=full","x":229.013,"ours_us":252.69,"base_us":57869.07,"floor_pct":75.5},{"load":"verify 1","shape":"bs=8 dims=full","x":308.305,"ours_us":187.31,"base_us":57748.39,"floor_pct":101.8},{"load":"verify 16","shape":"bs=128 dims=full","x":135.203,"ours_us":431.37,"base_us":58321.9,"floor_pct":44.3},{"load":"verify 32","shape":"bs=256 dims=full","x":78.036,"ours_us":756.48,"base_us":59033.16,"floor_pct":30.5}]},{"step":4,"x":136.15686017688427,"lo":135.7559798764701,"hi":136.59773580956085,"pct":36.84080841188893,"tier":"T1","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":245.713,"ours_us":234.29,"base_us":57568.64,"floor_pct":21.0},{"load":"decode 16","shape":"bs=16 dims=full","x":344.766,"ours_us":168.1,"base_us":57955.54,"floor_pct":113.5},{"load":"decode 32","shape":"bs=32 dims=full","x":314.529,"ours_us":184.38,"base_us":57992.71,"floor_pct":103.5},{"load":"verify 1","shape":"bs=8 dims=full","x":96.098,"ours_us":601.01,"base_us":57755.79,"floor_pct":31.7},{"load":"verify 16","shape":"bs=128 dims=full","x":61.24,"ours_us":952.06,"base_us":58303.67,"floor_pct":20.1},{"load":"verify 32","shape":"bs=256 dims=full","x":40.633,"ours_us":1453.03,"base_us":59040.61,"floor_pct":15.9}]},{"step":5,"x":241.71260644977312,"lo":241.1160306541624,"hi":242.02965497100246,"pct":65.38104479693294,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":610.88,"ours_us":94.24,"base_us":57570.01,"floor_pct":52.2},{"load":"decode 16","shape":"bs=16 dims=full","x":328.124,"ours_us":176.57,"base_us":57937.7,"floor_pct":108.1},{"load":"decode 32","shape":"bs=32 dims=full","x":248.595,"ours_us":233.38,"base_us":58017.32,"floor_pct":81.8},{"load":"verify 1","shape":"bs=8 dims=full","x":287.544,"ours_us":200.9,"base_us":57768.96,"floor_pct":95.0},{"load":"verify 16","shape":"bs=128 dims=full","x":124.986,"ours_us":466.99,"base_us":58367.55,"floor_pct":40.9},{"load":"verify 32","shape":"bs=256 dims=full","x":111.364,"ours_us":530.39,"base_us":59066.03,"floor_pct":43.6}]},{"step":6,"x":219.4748725792469,"lo":219.1459090654345,"hi":219.9396601234624,"pct":59.547828890403764,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":608.37,"ours_us":94.39,"base_us":57422.23,"floor_pct":52.1},{"load":"decode 16","shape":"bs=16 dims=full","x":327.213,"ours_us":176.49,"base_us":57748.7,"floor_pct":108.1},{"load":"decode 32","shape":"bs=32 dims=full","x":204.84,"ours_us":282.26,"base_us":57817.22,"floor_pct":67.6},{"load":"verify 1","shape":"bs=8 dims=full","x":305.151,"ours_us":188.78,"base_us":57605.1,"floor_pct":101.1},{"load":"verify 16","shape":"bs=128 dims=full","x":107.854,"ours_us":539.54,"base_us":58191.3,"floor_pct":35.4},{"load":"verify 32","shape":"bs=256 dims=full","x":83.281,"ours_us":706.98,"base_us":58877.46,"floor_pct":32.7}]}],"confirm":{"step":null,"x":218.54292306429204,"lo":217.63000920020775,"hi":219.03909564984795,"pct":58.97540117346703,"tier":"T2","verdict":"VERIFIED","note":null,"cases":[{"load":"decode 1","shape":"bs=1 dims=full","x":584.723,"ours_us":98.67,"base_us":57692.47,"floor_pct":49.9},{"load":"decode 16","shape":"bs=16 dims=full","x":321.644,"ours_us":180.66,"base_us":58109.44,"floor_pct":105.6},{"load":"decode 32","shape":"bs=32 dims=full","x":253.523,"ours_us":229.32,"base_us":58137.83,"floor_pct":83.2},{"load":"verify 1","shape":"bs=8 dims=full","x":287.833,"ours_us":201.35,"base_us":57956.6,"floor_pct":94.7},{"load":"verify 16","shape":"bs=128 dims=full","x":103.807,"ours_us":563.41,"base_us":58485.58,"floor_pct":33.9},{"load":"verify 32","shape":"bs=256 dims=full","x":76.473,"ours_us":773.65,"base_us":59163.38,"floor_pct":29.9}]},"src":"bench/product/runs/learnability/moe-A.jsonl"}],"dataset":[{"what":"GPU runs with receipts","count":58519,"src":"bench/runs.jsonl"},{"what":"GPU runs incl. remote host ledgers","count":60466,"src":"bench/remote/*/runs.jsonl + bench/runs.jsonl"},{"what":"competition-judge runs (main ledger)","count":531,"src":"bench/runs.jsonl kind"},{"what":"graded RL attempts","count":65342,"src":"bench/rl/episodes.jsonl (event=done)"},{"what":"per-case rewards","count":676844,"src":"bench/rl/episodes.jsonl cases"},{"what":"pre-registered ideas","count":14667,"src":"bench/sci/events.jsonl"},{"what":"graded idea trials","count":13638,"src":"bench/sci/events.jsonl"},{"what":"agent lane runs (pi harness)","count":64756,"src":"bench/pi/runs.jsonl"},{"what":"other lanes","count":2955,"src":"bench/oai/runs.jsonl, bench/cma/runs.jsonl"},{"what":"model turns (redacted trace shards)","count":83152,"src":"bench/traces/shards"},{"what":"tool calls (redacted trace shards)","count":107416,"src":"bench/traces/shards"},{"what":"Antigravity steps (redacted trace shards)","count":155968,"src":"bench/traces/shards"},{"what":"kernel files","count":47442,"src":"bench/runs/<problem>/, bench/rl/candidates/"},{"what":"cold-start episodes (v0.2 grader)","count":4,"src":"bench/product/runs/learnability"}],"packages":[{"name":"FloorBench-Gym","what":"All task families as RL environments. reset() gives the prompt, the starter, the cases and the floors; step() grades a file and returns ln(x) when every gate passes.","who":"For RL-environment teams."},{"name":"FloorBench-Eval","what":"We run your model on held-out seeds on our H200s and report confirmed pass@1 tiers, the share of the floor per case and the cost per episode.","who":"For eval and capability teams."},{"name":"FloorBench-Data","what":"The measured trajectories: every graded attempt with its receipt, per-case rewards, pre-registered ideas and their outcomes, and lane runs with cost and turns.","who":"For training and research teams. Model text goes only to the provider that wrote it."}],"sample":{"href":"downloads/FloorBench-H200-sample-v0.3.zip","bytes":255559,"sha256":"6c4de84b96905a780a68bfdc6860677883b0795b104bd1f69ceced7034b6c4c6","n_files":123,"featured":["dense_proj_bf16--decode-1","moe_nvfp4_ep8--decode-1","attn_decode_fp8kv--decode-1","decoder_layer_tp8--decode-1","prefill_layer--turn-1","ptx_roundtrip_surgery--vec_load"],"quickstart":"unzip FloorBench-H200-sample-v0.3.zip && cd FloorBench-H200-sample-v0.3\npython3 tests/test_cpu.py         # every gate, on any CPU\npython3 tests/test_redteam.py     # the known cheats must come back INVALID\npython3 -m floorbench.harness --task dense_proj_bf16 --file my_kernel.py   # H200: x, % of floor, tier","contents":[{"path":"floorbench/","what":"The harness, the gates, the timing code, the env API, and 6 of the 9 task families: all 5 kernel families and one compiler family."},{"path":"tasks/","what":"6 task cards in Harbor format, one per family, each with its naive starter and floors: dense_proj_bf16--decode-1, moe_nvfp4_ep8--decode-1, attn_decode_fp8kv--decode-1, decoder_layer_tp8--decode-1, prefill_layer--turn-1, ptx_roundtrip_surgery--vec_load."},{"path":"tests/redteam/","what":"Known cheats. Each must come back INVALID."},{"path":"trajectories/","what":"Four cold-start episodes: numeric receipts per step and the confirmation run."},{"path":"measurements/sample_runs.csv","what":"200 GPU runs: problem, per-case ratios, x, label, model family. Numbers only."},{"path":"DATASET_CARD.md","what":"Counts of the full dataset, with their sources."}],"tests":[{"name":"Registry: 6 runnable families, 75 tasks","result":"pass"},{"name":"Gates + starters","result":"14 of 14"},{"name":"Red team","result":"ALL CAUGHT"},{"name":"Honest-kernel false rejects","result":"false-reject rate: 0/5 = 0.0%"},{"name":"Tiers","result":"25/25"},{"name":"Compiler family (sample)","result":"36 passed in 0.60s"},{"name":"No private artifact (reference, calib.json, agent code, native stream)","result":"pass"},{"name":"No v1 calibration module","result":"pass"},{"name":"No pass rate set by threshold tuning","result":"pass"},{"name":"No warm-start incumbent targets","result":"pass"},{"name":"Every family named is runnable here or marked withheld in README.md","result":"pass"},{"name":"Canary in every doc and data file","result":"pass"},{"name":"No secret or personal-data pattern","result":"pass"}],"tests_when":"2026-11-05T20:34:55Z","excluded":[{"rule":"R-A","what":"agent-written kernel files, step notes and native agent streams of the 4 cold-start episodes (Claude Opus 5.5; Gemini 4 Argon in Antigravity)","why":"model output under provider terms that bar using it to build competing models (RISKS.md s.2, class B). Raw text and code go only to the matching provider; the numeric receipts, code sha256 and aggregate counts stay"},{"rule":"R-A","what":"kernel code and free text of the measurements rows","why":"same terms; the CSV is measurements only (class C Kimi rows keep their numbers, as in every profile)"},{"rule":"R-B","what":"private REFERENCE solutions, calib.json, reference SASS fixtures, our T3 kernels, held-out seeds","why":"CRITERIA R3.9: never in a sample pack; they are the eval's secret"},{"rule":"R-B","what":"the source competition's task files, reference, checks, seeds, judge code and our scored kernel","why":"not licensed to us (RISKS.md s.1); our references are our own re-specification"},{"rule":"R-C","what":"3 compiler/chip families (sass_static_verifier_gates, sm90a_pipe_microbench, isa_instruction_budget), their tasks, fixtures and red-team files; tests/test_isa_gpu.py; the v0.2 episode tool (needs our H200 boxes)","why":"the sample shows one family per kernel class plus one compiler family; the rest is the product"},{"rule":"R-D","what":"secrets, e-mails, names, host names, user paths","why":"scrubbed (scrub.py + literal redactions); the build fails on any residual hit"}]},"fleet":[{"model":"GPT-6 Astra","harness":"Codex CLI","lanes":85,"lane_unit":"derived GPU sessions","hours":170.4,"gpu_runs":598,"wins":24,"rate":14.08450704225352,"rlo":9.71127420504965,"rhi":19.807024726838964,"p":0.14118,"lo":0.0901,"hi":0.2144,"per_k_usd":null,"per_100_runs":4.01,"odds":{"odds_ratio":0.99,"ci90":[0.65,1.49],"n_valid_runs":538,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":1,"note":"84 of 85 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Claude Opus 5.5","harness":"Claude Code","lanes":633,"lane_unit":"derived GPU sessions","hours":1690.5,"gpu_runs":20565,"wins":107,"rate":6.329488317065957,"rlo":5.35767258104114,"rhi":7.4323654618912185,"p":0.08215,"lo":0.0659,"hi":0.10189999999999999,"per_k_usd":null,"per_100_runs":0.52,"odds":{"odds_ratio":1.0,"ci90":[1.0,1.0],"n_valid_runs":19925,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":2,"note":"596 of 633 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Gemini 4 Argon","harness":"pi","lanes":1103,"lane_unit":"mixed","hours":300.0,"gpu_runs":594,"wins":19,"rate":6.333333333333333,"rlo":4.146917576829817,"rhi":9.29250922030975,"p":0.013600000000000001,"lo":0.0089,"hi":0.0206,"per_k_usd":3.75,"per_100_runs":3.2,"odds":{"odds_ratio":1.52,"ci90":[0.96,2.28],"n_valid_runs":585,"note":"controls ln(runs so far on the problem) + problem"},"usd":4531.0,"billing":"API","rank":3,"note":"14 of 1103 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"GPT-6.1 Sol","harness":"pi","lanes":1162,"lane_unit":"ledger","hours":237.0,"gpu_runs":740,"wins":9,"rate":3.7974683544303796,"rlo":1.9800765286605277,"rhi":6.625092138984318,"p":0.00688,"lo":0.0039000000000000003,"hi":0.012199999999999999,"per_k_usd":14.88,"per_100_runs":1.22,"odds":{"odds_ratio":0.82,"ci90":[0.46,1.25],"n_valid_runs":698,"note":"controls ln(runs so far on the problem) + problem"},"usd":605.0,"billing":"API","rank":4,"note":"active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Gemini 4 Argon","harness":"Antigravity (agy)","lanes":67,"lane_unit":"derived GPU sessions","hours":55.0,"gpu_runs":259,"wins":3,"rate":5.454545454545454,"rlo":1.4730838452456676,"rhi":14.08179469120214,"p":0.04478,"lo":0.018000000000000002,"hi":0.1069,"per_k_usd":null,"per_100_runs":1.16,"odds":{"odds_ratio":0.94,"ci90":[0.35,1.85],"n_valid_runs":244,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":5,"note":"67 of 67 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses)"},{"model":"Kimi K3.1 Pro","harness":"Kimi Code CLI","lanes":350,"lane_unit":"mixed","hours":301.6,"gpu_runs":2874,"wins":4,"rate":1.326259946949602,"rlo":0.45112678929681843,"rhi":3.03261955435398,"p":0.00857,"lo":0.0034000000000000002,"hi":0.0213,"per_k_usd":null,"per_100_runs":0.14,"odds":{"odds_ratio":0.35,"ci90":[0.19,0.64],"n_valid_runs":2837,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":6,"note":"45 of 350 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active in the fresh-pool phase (from 10-22..10-24) when records were easy"},{"model":"Kimi K3","harness":"pi","lanes":15913,"lane_unit":"ledger","hours":699.5,"gpu_runs":2378,"wins":5,"rate":0.7147962830593281,"rlo":0.2809986047541031,"rhi":1.502068278039031,"p":0.00025,"lo":0.0001,"hi":0.0006,"per_k_usd":2.47,"per_100_runs":0.21,"odds":{"odds_ratio":1.93,"ci90":[0.85,3.37],"n_valid_runs":2319,"note":"controls ln(runs so far on the problem) + problem"},"usd":2027.0,"billing":"API","rank":7,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"Claude Sonnet 5.5","harness":"pi","lanes":32929,"lane_unit":"ledger","hours":471.7,"gpu_runs":5855,"wins":2,"rate":0.423998304006784,"rlo":0.07315787417252727,"rhi":1.3323563662201099,"p":6e-05,"lo":0.0,"hi":0.0002,"per_k_usd":0.99,"per_100_runs":0.03,"odds":{"odds_ratio":0.45,"ci90":[0.23,0.89],"n_valid_runs":5607,"note":"controls ln(runs so far on the problem) + problem"},"usd":2015.0,"billing":"API","rank":8,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"GPT-6.1 Sol","harness":"Codex CLI","lanes":896,"lane_unit":"mixed","hours":738.3,"gpu_runs":1006,"wins":2,"rate":0.27089259108763375,"rlo":0.04674057869047964,"rhi":0.8512427169795825,"p":0.00223,"lo":0.0007000000000000001,"hi":0.0067,"per_k_usd":null,"per_100_runs":0.2,"odds":{"odds_ratio":0.48,"ci90":[0.25,0.86],"n_valid_runs":993,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":9,"note":"32 of 896 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses)"},{"model":"GPT-6.1 Terra","harness":"Codex CLI","lanes":6110,"lane_unit":"mixed","hours":1847.6,"gpu_runs":1196,"wins":1,"rate":0.05412426932236415,"rlo":0.0021377452724046994,"rhi":0.25592077967962956,"p":0.00016,"lo":0.0,"hi":0.0007000000000000001,"per_k_usd":null,"per_100_runs":0.08,"odds":{"odds_ratio":0.64,"ci90":[0.39,1.13],"n_valid_runs":1183,"note":"controls ln(runs so far on the problem) + problem"},"usd":null,"billing":"subscription","rank":10,"note":"72 of 6110 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses)"},{"model":"Claude Haiku 5.5","harness":"pi","lanes":5147,"lane_unit":"ledger","hours":140.4,"gpu_runs":1226,"wins":0,"rate":0.0,"rlo":0.0,"rhi":2.114485264885942,"p":0.0,"lo":0.0,"hi":0.0005,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.41,"ci90":[0.37,0.47],"n_valid_runs":1221,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":63.0,"billing":"API","rank":11,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"Claude Opus 5.5","harness":"pi","lanes":2352,"lane_unit":"mixed","hours":447.1,"gpu_runs":1484,"wins":0,"rate":0.0,"rlo":0.0,"rhi":0.6639985041153796,"p":0.0,"lo":0.0,"hi":0.0011,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.51,"ci90":[0.46,0.58],"n_valid_runs":1450,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":3617.0,"billing":"API","rank":12,"note":"49 of 2352 lane runs are derived GPU sessions (span + 15 min; a derived session always has >= 1 GPU run, which inflates P per lane vs ledger harnesses); active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"GPT-6.1 Terra","harness":"pi","lanes":3502,"lane_unit":"ledger","hours":56.5,"gpu_runs":274,"wins":0,"rate":0.0,"rlo":0.0,"rhi":5.254402321946659,"p":0.0,"lo":0.0,"hi":0.0008,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.74,"ci90":[0.69,0.79],"n_valid_runs":274,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":1250.0,"billing":"API","rank":13,"note":"mostly RUNNER role after the 10-28 provenance gate: re-ran files others wrote; active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"Claude Sonnet 5.5","harness":"Claude Managed Agents","lanes":1490,"lane_unit":"ledger","hours":207.6,"gpu_runs":996,"wins":0,"rate":0.0,"rlo":0.0,"rhi":1.4300276068881803,"p":0.0,"lo":0.0,"hi":0.0018,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.6,"ci90":[0.56,0.67],"n_valid_runs":511,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":1664.0,"billing":"API","rank":14,"note":"active only from 10-28, when the problems were heavily mined (records rare for every family)"},{"model":"GPT-6.1 Terra","harness":"OpenAI Agents","lanes":1538,"lane_unit":"ledger","hours":100.4,"gpu_runs":460,"wins":0,"rate":0.0,"rlo":0.0,"rhi":2.956909673207034,"p":0.0,"lo":0.0,"hi":0.0018,"per_k_usd":0.0,"per_100_runs":0.0,"odds":{"odds_ratio":0.62,"ci90":[0.58,0.68],"n_valid_runs":459,"note":"controls ln(runs so far on the problem) + problem; 0 records: the odds is set by the ridge penalty, read it as 'no records'"},"usd":219.0,"billing":"API","rank":15,"note":"active only from 10-28, when the problems were heavily mined (records rare for every family)"}],"knowledge":{"generated":"2026-11-06T00:57:34Z","learning_files":839,"learning_entries":27216,"derived_entries":3774,"assumptions":182,"knobs":283,"questions":101,"primitives":46,"hw_sources":57,"rules":"entries = bullet or table rows in every LEARNINGS.md; derived = rows under a heading containing 'derived' (the 'Derived from first principles' section AGENTS.md asks for); assumptions/knobs/primitives = table rows; questions = headings in bench/QUESTIONS.md","sources":[{"file":"**/LEARNINGS.md","what":"per-folder notebooks: what moved the number, what did not, what was derived"},{"file":"bench/ASSUMPTIONS.md","what":"assumptions tested"},{"file":"bench/KNOBS.md","what":"tuning knobs and ranges"},{"file":"bench/QUESTIONS.md","what":"open and answered questions"},{"file":"bench/PRIMITIVES.md","what":"reusable kernel building blocks"},{"file":"bench/hw/sources/","what":"hardware documents read"}]},"examples":{"problems":[{"key":"muse_decode_layer","short":"dec","label":"Decode layer","slot":1,"prompt":{"head":"Write kernel_fn(data, state) for one whole decoder layer on one H200 of a TP8/EP8 deployment of a 256-expert NVFP4 MoE model, at decode (q_len 1) and 8-token verify shapes, ~80K-token contexts, up to 32 requests:","full":"Write kernel_fn(data, state) for one whole decoder layer on one H200 of a TP8/EP8 deployment of a\n256-expert NVFP4 MoE model, at decode (q_len 1) and 8-token verify shapes, ~80K-token contexts, up to 32 requests:\nthe fused qkv/gate projection, GQA attention over a paged FP8 cache (QK-norm, RoPE on sliding layers, write the\ncache), the o projection, the residual norms, the FP32 router (sigmoid, bias, top 8), the pooler, this GPU's 32\nlocal NVFP4 experts, the unpooler, and the final norm (the exact formula is in the task's reference description).\nReturn (BF16 out [T, 8192], FP32 residual [T, 8192], int32 experts [T, 8]). Weights may be repacked losslessly in\nprepare(), never at a lower precision; activations stay BF16 where the reference keeps BF16; FP32 softmax sums.\nGraph-replayed: no host syncs, no value-dependent allocation or control flow, nothing cached from data.\nCorrectness: out rel L2 <= 2%, residual <= 1%, experts equal to the reference up to 1e-4 score ties, the cache rows\nwritten. Score: geomean speedup over loads and % of the serial-stage physical floor.\nWrite **`/app/submission.py`** (a starter that passes every gate is already there). It must define:\n```python\ndef prepare(weights):        # once, untimed, up to 300 s; return any state (repack losslessly, keep precision)\n    ...\ndef kernel_fn(data, state):  # the timed call; it is replayed as a CUDA graph\n    ...\n```\nThe reference implementation (our own spec, in plain PyTorch) is at `/app/reference/decoder_layer.py`. Read it; do\nnot import it (the static scan rejects imports of `floorbench` and `reference`). The NVFP4 codec may be imported as\n`from floorbench_nvfp4 import dequantize, quantize`.\n| field | dtype | shape (self-test case; scored shapes in the table below) |\n|---|---|---|\n| data[\"positions\"] | int32 | [2] |\n| data[\"cache_loc\"] | int32 | [2] |\n| data[\"page_table\"] | int32 | [2, 200] |\n| data[\"seq_lens\"] | int32 | [2] |\n| data[\"k_cache\"] | float8_e4m3fn | [404, 2, 64] |\n| data[\"v_cache\"] | float8_e4m3fn | [404, 2, 64] |\n| data[\"x\"] | bfloat16 | [2, 512] |\n| data[\"s\"] | float32 | [2, 512] |\n| data[\"config\"] | dict | keys: batch, dims, hidden, kv_scale, layer, max_seq_len, page_size, q_len, unpool_rows, window |\n| weights[\"qkv_gate\"] | bfloat16 | [2304, 512] |\n| weights[\"o\"] | bfloat16 | [512, 1024] |\n| weights[\"pooler\"] | bfloat16 | [256, 512] |\n| weights[\"unpooler\"] | bfloat16 | [512, 256] |\n| weights[\"router\"] | float32 | [64, 512] |\n| weights[\"router_bias\"] | float32 | [64] |\n| weights[\"w_pre\"] | bfloat16 | [512] |\n| weights[\"w_pool\"] | bfloat16 | [256] |\n| weights[\"w_post\"] | bfloat16 | [512] |","src":"bench/product/harbor/tasks/decoder_layer_tp8--all/instruction.md"},"run_x":2.4513545799748853,"run_src":"bench/runs/muse_decode_layer/2f09d275e8ea4096.json (best passed judge run); floors from bench/tools/gap.py phys()","cases":[{"load":"decode 1","col":"sliding","variant":"b1 ctx80000 sliding repeats=60","w":46,"ours_us":59.937,"sglang_us":172.4,"floor_us":54.596,"ours_pct":91.1,"sglang_pct":31.7},{"load":"decode 1","col":"full","variant":"b1 ctx80000 full repeats=60","w":16,"ours_us":64.779,"sglang_us":182.32,"floor_us":58.851,"ours_pct":90.8,"sglang_pct":32.3},{"load":"decode 16","col":"sliding","variant":"b16 ctx80000 sliding","w":46,"ours_us":112.771,"sglang_us":259.488,"floor_us":99.881,"ours_pct":88.6,"sglang_pct":38.5},{"load":"decode 16","col":"full","variant":"b16 ctx80000 full","w":16,"ours_us":189.068,"sglang_us":434.976,"floor_us":161.768,"ours_pct":85.6,"sglang_pct":37.2},{"load":"decode 32","col":"sliding","variant":"b32 ctx80000 sliding","w":46,"ours_us":137.788,"sglang_us":265.856,"floor_us":109.839,"ours_pct":79.7,"sglang_pct":41.3},{"load":"decode 32","col":"full","variant":"b32 ctx80000 full","w":16,"ours_us":283.291,"sglang_us":623.136,"floor_us":236.51,"ours_pct":83.5,"sglang_pct":38.0},{"load":"verify 1","col":"sliding","variant":"b1 ctx80000 sliding q8","w":46,"ours_us":96.188,"sglang_us":259.728,"floor_us":84.926,"ours_pct":88.3,"sglang_pct":32.7},{"load":"verify 1","col":"full","variant":"b1 ctx80000 full q8","w":16,"ours_us":107.841,"sglang_us":273.856,"floor_us":89.181,"ours_pct":82.7,"sglang_pct":32.6},{"load":"verify 16","col":"sliding","variant":"b16 ctx80000 sliding q8","w":46,"ours_us":224.041,"sglang_us":541.696,"floor_us":167.182,"ours_pct":74.6,"sglang_pct":30.9},{"load":"verify 16","col":"full","variant":"b16 ctx80000 full q8","w":16,"ours_us":302.212,"sglang_us":712.432,"floor_us":230.88,"ours_pct":76.4,"sglang_pct":32.4},{"load":"verify 32","col":"sliding","variant":"b32 ctx80000 sliding q8","w":46,"ours_us":286.915,"sglang_us":737.52,"floor_us":189.207,"ours_pct":65.9,"sglang_pct":25.7},{"load":"verify 32","col":"full","variant":"b32 ctx80000 full q8","w":16,"ours_us":442.587,"sglang_us":1093.968,"floor_us":315.576,"ours_pct":71.3,"sglang_pct":28.8}],"lost":[{"load":"verify 32","col":"sliding","variant":"b32 ctx80000 sliding q8","w":46,"ours_us":286.915,"sglang_us":737.52,"floor_us":189.207,"ours_pct":65.9,"sglang_pct":25.7},{"load":"verify 16","col":"sliding","variant":"b16 ctx80000 sliding q8","w":46,"ours_us":224.041,"sglang_us":541.696,"floor_us":167.182,"ours_pct":74.6,"sglang_pct":30.9},{"load":"verify 32","col":"full","variant":"b32 ctx80000 full q8","w":16,"ours_us":442.587,"sglang_us":1093.968,"floor_us":315.576,"ours_pct":71.3,"sglang_pct":28.8}],"loads":["decode 1","decode 16","decode 32","verify 1","verify 16","verify 32"],"cols":["sliding","full"],"top_class":"PORT","top_class_share":0.3265616845912329,"first_x":1.1682,"final":{"t":"2026-11-05T14:38:51Z","x":2.4513545799748853,"rank":25,"entries":1300,"leader":"pjmullin3","leader_x":2.614352808274751,"gap_pct":-6.23,"pct_ceiling":81.2},"ceiling":3.02},{"key":"muse_moe_w4a16","short":"moe","label":"MoE W4A16","slot":2,"prompt":{"head":"Write kernel_fn(data, state) for the routed experts of a top-8-of-256 MoE layer on one H200 (this GPU holds experts 0-31 of 256, 8-way expert parallel), faster than the baseline, for every case below. Weights are NVFP4","full":"Write kernel_fn(data, state) for the routed experts of a top-8-of-256 MoE layer on one H200 (this GPU holds\nexperts 0-31 of 256, 8-way expert parallel), faster than the baseline, for every case below. Weights are NVFP4\n(E2M1 + block-16 E4M3 scales + FP32 tensor scale) and must stay NVFP4 (dequantize inside the kernel). Per local route:\nSwiGLU expert, then y * (1 + post_norm_w), RMSNorm, times the route weight, summed per token. Return BF16 [bs, 4096].\nGraph-replayed: no host syncs, no value-dependent allocation or control flow, nothing cached from data. Correctness:\nrel L2 <= 1.5% vs an FP32 reference, each row <= 5x that. Score: geomean speedup over loads and % of physical floor.\nWrite **`/app/submission.py`** (a starter that passes every gate is already there). It must define:\n```python\ndef prepare(weights):        # once, untimed, up to 300 s; return any state (repack losslessly, keep precision)\n    ...\ndef kernel_fn(data, state):  # the timed call; it is replayed as a CUDA graph\n    ...\n```\nThe reference implementation (our own spec, in plain PyTorch) is at `/app/reference/moe_w4a16.py`. Read it; do\nnot import it (the static scan rejects imports of `floorbench` and `reference`). The NVFP4 codec may be imported as\n`from floorbench_nvfp4 import dequantize, quantize`.\n| field | dtype | shape (self-test case; scored shapes in the table below) |\n|---|---|---|\n| data[\"hidden_states\"] | bfloat16 | [1, 256] |\n| data[\"topk_ids\"] | int32 | [1, 8] |\n| data[\"topk_weights\"] | float32 | [1, 8] |\n| data[\"config\"] | dict | keys: bs, dims, sharding |\n| weights[\"w13\"] | uint8 | [8, 256, 128] |\n| weights[\"w13_sf\"] | float8_e4m3fn | [8, 256, 16] |\n| weights[\"w13_gs\"] | float32 | [8] |\n| weights[\"w2\"] | uint8 | [8, 256, 64] |\n| weights[\"w2_sf\"] | float8_e4m3fn | [8, 256, 8] |\n| weights[\"w2_gs\"] | float32 | [8] |\n| weights[\"post_norm_w\"] | bfloat16 | [256] |\nFloors: the time to move every byte at 4.8 TB/s or do every FLOP at 989 TFLOP/s, whichever is longer; the physical\nfloor uses the power-capped tensor rate (~900 TFLOP/s), adds launch (and, for whole layers, per-stage) costs.\n\"source\" columns are measurements from the public H200 competition this task was re-specified from (SGLang's kernel,\nand the best ratio our agent fleet reached there): context, not targets, and not this harness's baseline.\n| load | case | weight | floor (us) | physical floor (us) | source SGLang (us) | source best fleet ratio |\n|---|---|---|---|---|---|---|\n| decode 1 | bs 1 | 60 | 47.19 | 49.19 | 79.7 | 2.671x |\n| decode 16 | bs 16 | 60 | 188.80 | 190.80 | 188.0 | 2.016x |\n| decode 32 | bs 32 | 60 | 188.85 | 190.85 | 207.0 | 1.452x |\n| verify 1 | bs 8 | 60 | 188.77 | 190.77 | 141.8 | 2.113x |\n| verify 16 | bs 128 | 60 | 189.18 | 191.18 | 461.3 | 2.356x |","src":"bench/product/harbor/tasks/moe_nvfp4_ep8--all/instruction.md"},"run_x":2.5393063115219796,"run_src":"bench/runs/muse_moe_w4a16/8443d31fa577403b.json (best passed judge run); floors from bench/tools/gap.py phys()","cases":[{"load":"decode 1","col":"all","variant":"bs1","w":60,"ours_us":22.736,"sglang_us":80.224,"floor_us":18.438,"ours_pct":81.1,"sglang_pct":23.0},{"load":"decode 16","col":"all","variant":"bs16","w":60,"ours_us":86.204,"sglang_us":188.288,"floor_us":74.51,"ours_pct":86.4,"sglang_pct":39.6},{"load":"decode 32","col":"all","variant":"bs32","w":60,"ours_us":120.704,"sglang_us":212.4,"floor_us":99.376,"ours_pct":82.3,"sglang_pct":46.8},{"load":"verify 1","col":"all","variant":"bs8","w":60,"ours_us":61.612,"sglang_us":142.72,"floor_us":53.776,"ours_pct":87.3,"sglang_pct":37.7},{"load":"verify 16","col":"all","variant":"bs128","w":60,"ours_us":172.689,"sglang_us":472.352,"floor_us":141.002,"ours_pct":81.7,"sglang_pct":29.9},{"load":"verify 32","col":"all","variant":"bs256","w":60,"ours_us":203.364,"sglang_us":634.496,"floor_us":159.282,"ours_pct":78.3,"sglang_pct":25.1}],"lost":[{"load":"verify 32","col":"all","variant":"bs256","w":60,"ours_us":203.364,"sglang_us":634.496,"floor_us":159.282,"ours_pct":78.3,"sglang_pct":25.1},{"load":"verify 16","col":"all","variant":"bs128","w":60,"ours_us":172.689,"sglang_us":472.352,"floor_us":141.002,"ours_pct":81.7,"sglang_pct":29.9},{"load":"decode 32","col":"all","variant":"bs32","w":60,"ours_us":120.704,"sglang_us":212.4,"floor_us":99.376,"ours_pct":82.3,"sglang_pct":46.8}],"loads":["decode 1","decode 16","decode 32","verify 1","verify 16","verify 32"],"cols":["all"],"top_class":"STRUCTURAL-REWRITE","top_class_share":0.39155019747746433,"first_x":1.2712,"final":{"t":"2026-11-05T08:14:44Z","x":2.4292289478838875,"rank":122,"entries":1426,"leader":"pjmullin3","leader_x":2.843799916672147,"gap_pct":-14.58,"pct_ceiling":79.1},"ceiling":3.07},{"key":"muse_attention_decode","short":"attn","label":"Attention decode","slot":3,"prompt":{"head":"Write kernel_fn(data, state) for one decode step (q_len 1) or one 8-token verify block of grouped-query attention on one H200: 16 query heads, 2 KV heads, head dim 64, over a paged FP8 E4M3 KV cache (~80K tokens per","full":"Write kernel_fn(data, state) for one decode step (q_len 1) or one 8-token verify block of grouped-query\nattention on one H200: 16 query heads, 2 KV heads, head dim 64, over a paged FP8 E4M3 KV cache (~80K tokens per\nrequest, up to 32 requests). Per new token: split the fused qkv row, QK-norm (q scaled by 5.473), RoPE on sliding\nlayers, WRITE fp8(k / kv_scale) and fp8(v / kv_scale) into the cache at cache_loc, then attend causally over the\nrequest's keys (the last 2,048 positions on sliding layers, the whole context on full layers), softmax scale 1/8,\nand return rmsnorm(a) * sigmoid(gate) as BF16 [B, L, 1024]. Exact attention only: every key in range, no top-k,\nno skipping. Keep FP32 softmax sums. Graph-replayed: no host syncs, no value-dependent allocation or control\nflow (read lengths on the device), nothing cached from data. Correctness: rel L2 <= 2% vs an FP32 reference, rows\n<= 5x that, the written cache rows within 3%. Score: geomean speedup over loads and % of the physical floor.\nWrite **`/app/submission.py`** (a starter that passes every gate is already there). It must define:\n```python\ndef prepare(weights):        # once, untimed, up to 300 s; return any state (repack losslessly, keep precision)\n    ...\ndef kernel_fn(data, state):  # the timed call; it is replayed as a CUDA graph\n    ...\n```\nThe reference implementation (our own spec, in plain PyTorch) is at `/app/reference/attn_decode.py`. Read it; do\nnot import it (the static scan rejects imports of `floorbench` and `reference`). The NVFP4 codec may be imported as\n`from floorbench_nvfp4 import dequantize, quantize`.\n| field | dtype | shape (self-test case; scored shapes in the table below) |\n|---|---|---|\n| data[\"qkv\"] | bfloat16 | [2, 2304] |\n| data[\"positions\"] | int32 | [2] |\n| data[\"cache_loc\"] | int32 | [2] |\n| data[\"page_table\"] | int32 | [2, 300] |\n| data[\"seq_lens\"] | int32 | [2] |\n| data[\"k_cache\"] | float8_e4m3fn | [604, 2, 64] |\n| data[\"v_cache\"] | float8_e4m3fn | [604, 2, 64] |\n| data[\"config\"] | dict | keys: batch, kv_scale, layer, max_seq_len, page_size, q_len, window |\nFloors: the time to move every byte at 4.8 TB/s or do every FLOP at 989 TFLOP/s, whichever is longer; the physical\nfloor uses the power-capped tensor rate (~900 TFLOP/s), adds launch (and, for whole layers, per-stage) costs.\n\"source\" columns are measurements from the public H200 competition this task was re-specified from (SGLang's kernel,\nand the best ratio our agent fleet reached there): context, not targets, and not this harness's baseline.\n| load | case | weight | floor (us) | physical floor (us) | source SGLang (us) | source best fleet ratio |\n|---|---|---|---|---|---|---|\n| decode 1 | batch 1, q_len 1, ctx 80000, layer sliding, page 1, kv_scale 1.0, spread fixed, dist soft | 46 | 0.11 | 2.11 | 21.0 | 4.013x |\n| decode 1 | batch 1, q_len 1, ctx 80000, layer full, page 1, kv_scale 1.0, spread fixed, dist coherent | 16 | 4.27 | 6.27 | 31.4 | 2.424x |\n| decode 16 | batch 16, q_len 1, ctx 80000, layer sliding, page 1, kv_scale 1.0, spread agentic, dist soft | 46 | 1.77 | 3.77 | 26.9 | 3.010x |\n| decode 16 | batch 16, q_len 1, ctx 80000, layer full, page 1, kv_scale 1.0, spread agentic, dist coherent | 16 | 68.29 | 70.29 | 201.1 | 2.233x |\n| decode 32 | batch 32, q_len 1, ctx 80000, layer sliding, page 1, kv_scale 1.0, spread agentic, dist soft | 46 | 3.54 | 5.54 | 30.9 | 2.764x |","src":"bench/product/harbor/tasks/attn_decode_fp8kv--all/instruction.md"},"run_x":2.9097920417274863,"run_src":"bench/runs/muse_attention_decode/652c08c4dc2e4ab9.json (best passed judge run); floors from bench/tools/gap.py phys()","cases":[{"load":"decode 1","col":"sliding","variant":"b1 ctx80000 sliding","w":46,"ours_us":3.922,"sglang_us":20.912,"floor_us":3.3,"ours_pct":84.1,"sglang_pct":15.8},{"load":"decode 1","col":"full","variant":"b1 ctx80000 full","w":16,"ours_us":9.993,"sglang_us":31.632,"floor_us":9.91,"ours_pct":99.2,"sglang_pct":31.3},{"load":"decode 16","col":"sliding","variant":"b16 ctx80000 sliding","w":46,"ours_us":8.083,"sglang_us":26.944,"floor_us":6.82,"ours_pct":84.4,"sglang_pct":25.3},{"load":"decode 16","col":"full","variant":"b16 ctx80000 full","w":16,"ours_us":83.503,"sglang_us":201.536,"floor_us":74.721,"ours_pct":89.5,"sglang_pct":37.1},{"load":"decode 32","col":"sliding","variant":"b32 ctx80000 sliding","w":46,"ours_us":9.78,"sglang_us":31.296,"floor_us":10.1,"ours_pct":103.3,"sglang_pct":32.3},{"load":"decode 32","col":"full","variant":"b32 ctx80000 full","w":16,"ours_us":159.799,"sglang_us":387.648,"floor_us":146.581,"ours_pct":91.7,"sglang_pct":37.8},{"load":"verify 1","col":"sliding","variant":"b1 ctx80000 sliding q8","w":46,"ours_us":4.855,"sglang_us":23.312,"floor_us":4.45,"ours_pct":91.7,"sglang_pct":19.1},{"load":"verify 1","col":"full","variant":"b1 ctx80000 full q8","w":16,"ours_us":13.993,"sglang_us":36.752,"floor_us":7.731,"ours_pct":55.2,"sglang_pct":21.0},{"load":"verify 16","col":"sliding","variant":"b16 ctx80000 sliding q8","w":46,"ours_us":9.895,"sglang_us":29.664,"floor_us":8.86,"ours_pct":89.5,"sglang_pct":29.9},{"load":"verify 16","col":"full","variant":"b16 ctx80000 full q8","w":16,"ours_us":92.264,"sglang_us":206.288,"floor_us":74.965,"ours_pct":81.3,"sglang_pct":36.3},{"load":"verify 32","col":"sliding","variant":"b32 ctx80000 sliding q8","w":46,"ours_us":12.917,"sglang_us":34.72,"floor_us":13.34,"ours_pct":103.3,"sglang_pct":38.4},{"load":"verify 32","col":"full","variant":"b32 ctx80000 full q8","w":16,"ours_us":170.917,"sglang_us":391.52,"floor_us":147.07,"ours_pct":86.0,"sglang_pct":37.6}],"lost":[{"load":"verify 32","col":"full","variant":"b32 ctx80000 full q8","w":16,"ours_us":170.917,"sglang_us":391.52,"floor_us":147.07,"ours_pct":86.0,"sglang_pct":37.6},{"load":"verify 16","col":"full","variant":"b16 ctx80000 full q8","w":16,"ours_us":92.264,"sglang_us":206.288,"floor_us":74.965,"ours_pct":81.3,"sglang_pct":36.3},{"load":"decode 32","col":"full","variant":"b32 ctx80000 full","w":16,"ours_us":159.799,"sglang_us":387.648,"floor_us":146.581,"ours_pct":91.7,"sglang_pct":37.8}],"loads":["decode 1","decode 16","decode 32","verify 1","verify 16","verify 32"],"cols":["sliding","full"],"top_class":"STRUCTURAL-REWRITE","top_class_share":0.45636124594841593,"first_x":1.4804,"final":{"t":"2026-11-05T19:10:41Z","x":2.835522644274862,"rank":42,"entries":1380,"leader":"pjmullin3","leader_x":3.3389207412481454,"gap_pct":-15.08,"pct_ceiling":83.9},"ceiling":3.38},{"key":"muse_prefill_layer","short":"pre","label":"Prefill layer","slot":4,"prompt":{"head":"Write kernel_fn(data, state) for one whole decoder layer on one H200 of a TP8/EP8 deployment of a 256-expert NVFP4 MoE model, at prefill shapes: 753 new tokens on ~80K-token contexts (1 or 4 requests) and 16K-token first-turn chunks, causal within the block:","full":"Write kernel_fn(data, state) for one whole decoder layer on one H200 of a TP8/EP8 deployment of a\n256-expert NVFP4 MoE model, at prefill shapes: 753 new tokens on ~80K-token contexts (1 or 4 requests) and 16K-token first-turn chunks, causal within the block:\nthe fused qkv/gate projection, GQA attention over a paged FP8 cache (QK-norm, RoPE on sliding layers, write the\ncache), the o projection, the residual norms, the FP32 router (sigmoid, bias, top 8), the pooler, this GPU's 32\nlocal NVFP4 experts, the unpooler, and the final norm (the exact formula is in the task's reference description).\nReturn (BF16 out [T, 8192], FP32 residual [T, 8192], int32 experts [T, 8]). Weights may be repacked losslessly in\nprepare(), never at a lower precision; activations stay BF16 where the reference keeps BF16; FP32 softmax sums.\nGraph-replayed: no host syncs, no value-dependent allocation or control flow, nothing cached from data.\nCorrectness: out rel L2 <= 2%, residual <= 1%, experts equal to the reference up to 1e-4 score ties, the cache rows\nwritten. Score: geomean speedup over loads and % of the serial-stage physical floor.\nWrite **`/app/submission.py`** (a starter that passes every gate is already there). It must define:\n```python\ndef prepare(weights):        # once, untimed, up to 300 s; return any state (repack losslessly, keep precision)\n    ...\ndef kernel_fn(data, state):  # the timed call; it is replayed as a CUDA graph\n    ...\n```\nThe reference implementation (our own spec, in plain PyTorch) is at `/app/reference/prefill_layer.py`. Read it; do\nnot import it (the static scan rejects imports of `floorbench` and `reference`). The NVFP4 codec may be imported as\n`from floorbench_nvfp4 import dequantize, quantize`.\n| field | dtype | shape (self-test case; scored shapes in the table below) |\n|---|---|---|\n| data[\"positions\"] | int32 | [40] |\n| data[\"cache_loc\"] | int32 | [40] |\n| data[\"page_table\"] | int32 | [1, 300] |\n| data[\"seq_lens\"] | int32 | [1] |\n| data[\"k_cache\"] | float8_e4m3fn | [304, 2, 64] |\n| data[\"v_cache\"] | float8_e4m3fn | [304, 2, 64] |\n| data[\"x\"] | bfloat16 | [40, 512] |\n| data[\"s\"] | float32 | [40, 512] |\n| data[\"config\"] | dict | keys: batch, dims, hidden, kv_scale, layer, max_seq_len, page_size, q_len, unpool_rows, window |\n| weights[\"qkv_gate\"] | bfloat16 | [2304, 512] |\n| weights[\"o\"] | bfloat16 | [512, 1024] |\n| weights[\"pooler\"] | bfloat16 | [256, 512] |\n| weights[\"unpooler\"] | bfloat16 | [512, 256] |\n| weights[\"router\"] | float32 | [64, 512] |\n| weights[\"router_bias\"] | float32 | [64] |\n| weights[\"w_pre\"] | bfloat16 | [512] |\n| weights[\"w_pool\"] | bfloat16 | [256] |\n| weights[\"w_post\"] | bfloat16 | [512] |","src":"bench/product/harbor/tasks/prefill_layer--all/instruction.md"},"run_x":1.920697465982476,"run_src":"bench/runs/muse_prefill_layer/e4feafb0d35b4c00.json (best passed judge run); floors from bench/tools/gap.py phys()","cases":[{"load":"turn 1","col":"sliding","variant":"b1 ctx80000 sliding q753 repeats=6","w":46,"ours_us":586.537,"sglang_us":1177.968,"floor_us":378.156,"ours_pct":64.5,"sglang_pct":32.1},{"load":"turn 1","col":"full","variant":"b1 ctx80000 full q753 repeats=6","w":16,"ours_us":790.974,"sglang_us":1465.2,"floor_us":721.786,"ours_pct":91.3,"sglang_pct":49.3},{"load":"turn 4","col":"sliding","variant":"b4 ctx80000 sliding q753 repeats=6","w":46,"ours_us":1768.198,"sglang_us":3367.44,"floor_us":1495.191,"ours_pct":84.6,"sglang_pct":44.4},{"load":"turn 4","col":"full","variant":"b4 ctx80000 full q753 repeats=6","w":16,"ours_us":2701.565,"sglang_us":4943.936,"floor_us":2865.661,"ours_pct":106.1,"sglang_pct":58.0},{"load":"first turn","col":"full","variant":"b1 ctx0 full q16384 repeats=6","w":16,"ours_us":9175.349,"sglang_us":17638.495,"floor_us":7919.612,"ours_pct":86.3,"sglang_pct":44.9},{"load":"first turn","col":"sliding","variant":"b1 ctx32768 sliding q16384 repeats=","w":138,"ours_us":8971.403,"sglang_us":17361.488,"floor_us":8127.334,"ours_pct":90.6,"sglang_pct":46.8},{"load":"first turn","col":"full","variant":"b1 ctx32768 full q16384 repeats=6","w":16,"ours_us":11211.635,"sglang_us":21338.896,"floor_us":11060.866,"ours_pct":98.7,"sglang_pct":51.8},{"load":"first turn","col":"full","variant":"b1 ctx65536 full q16384 repeats=6","w":16,"ours_us":12772.159,"sglang_us":24148.513,"floor_us":14124.872,"ours_pct":110.6,"sglang_pct":58.5}],"lost":[{"load":"first turn","col":"sliding","variant":"b1 ctx32768 sliding q16384 repeats=","w":138,"ours_us":8971.403,"sglang_us":17361.488,"floor_us":8127.334,"ours_pct":90.6,"sglang_pct":46.8},{"load":"first turn","col":"full","variant":"b1 ctx0 full q16384 repeats=6","w":16,"ours_us":9175.349,"sglang_us":17638.495,"floor_us":7919.612,"ours_pct":86.3,"sglang_pct":44.9},{"load":"turn 4","col":"sliding","variant":"b4 ctx80000 sliding q753 repeats=6","w":46,"ours_us":1768.198,"sglang_us":3367.44,"floor_us":1495.191,"ours_pct":84.6,"sglang_pct":44.4}],"loads":["turn 1","turn 4","first turn"],"cols":["sliding","full"],"top_class":"TWEAK","top_class_share":0.5366160440141866,"first_x":1.3469,"final":{"t":"2026-11-05T09:58:09Z","x":1.920697465982476,"rank":104,"entries":899,"leader":"LeshengJin","leader_x":2.1470638260618298,"gap_pct":-10.54,"pct_ceiling":84.6},"ceiling":2.27},{"key":"muse_bf16_proj","short":"bf16","label":"BF16 projections","slot":5,"prompt":{"head":"Write kernel_fn(data, state) for out = x @ W.T on one NVIDIA H200, faster than cuBLAS, for every case below. x: [m, k] BF16; W = weights[layer]: [n, k], BF16 (router: FP32, output FP32). prepare(weights) runs once, untimed, and","full":"Write kernel_fn(data, state) for out = x @ W.T on one NVIDIA H200, faster than cuBLAS, for every case below.\nx: [m, k] BF16; W = weights[layer]: [n, k], BF16 (router: FP32, output FP32). prepare(weights) runs once, untimed, and\nmay repack the weights losslessly but not change their precision (the router may use a BF16 hi + lo split with FP32\naccumulation). kernel_fn is replayed as a CUDA graph: no host syncs, no allocation or branching on values, no caching\nof anything derived from data. Correctness: rel L2 <= 1% vs an FP64 reference (router 0.01%), each row <= 5x that,\nprobe rows (one power-of-two entry) <= 0.1%. Score: geomean over loads of weighted baseline/ours time, and % of the\nphysical floor per case.\nWrite **`/app/submission.py`** (a starter that passes every gate is already there). It must define:\n```python\ndef prepare(weights):        # once, untimed, up to 300 s; return any state (repack losslessly, keep precision)\n    ...\ndef kernel_fn(data, state):  # the timed call; it is replayed as a CUDA graph\n    ...\n```\nThe reference implementation (our own spec, in plain PyTorch) is at `/app/reference/bf16_proj.py`. Read it; do\nnot import it (the static scan rejects imports of `floorbench` and `reference`). The NVFP4 codec may be imported as\n`from floorbench_nvfp4 import dequantize, quantize`.\n| field | dtype | shape (self-test case; scored shapes in the table below) |\n|---|---|---|\n| data[\"x\"] | bfloat16 | [3, 1024] |\n| data[\"config\"] | dict | keys: k, layer, m, n |\n| weights[\"qkv_gate\"] | bfloat16 | [2304, 8192] |\n| weights[\"o\"] | bfloat16 | [8192, 1024] |\n| weights[\"pooler\"] | bfloat16 | [4096, 8192] |\n| weights[\"unpooler\"] | bfloat16 | [8192, 4096] |\n| weights[\"router\"] | float32 | [256, 8192] |\nFloors: the time to move every byte at 4.8 TB/s or do every FLOP at 989 TFLOP/s, whichever is longer; the physical\nfloor uses the power-capped tensor rate (~900 TFLOP/s), adds launch (and, for whole layers, per-stage) costs.\n\"source\" columns are measurements from the public H200 competition this task was re-specified from (SGLang's kernel,\nand the best ratio our agent fleet reached there): context, not targets, and not this harness's baseline.\n| load | case | weight | floor (us) | physical floor (us) | source SGLang (us) | source best fleet ratio |\n|---|---|---|---|---|---|---|\n| decode 1 | m 1, layer qkv_gate | 62 | 7.87 | 9.87 | 14.0 | 1.410x |\n| decode 1 | m 1, layer o | 62 | 3.50 | 5.50 | 6.8 | 1.189x |\n| decode 1 | m 1, layer pooler | 62 | 13.99 | 15.99 | 21.4 | 1.262x |\n| decode 1 | m 1, layer unpooler | 62 | 1.75 | 3.75 | 7.1 | 1.754x |\n| decode 1 | m 1, layer router | 62 | 1.75 | 3.75 | 10.4 | 2.610x |\n| decode 16 | m 16, layer qkv_gate | 62 | 7.93 | 9.93 | 14.1 | 1.231x |\n| decode 16 | m 16, layer o | 62 | 3.56 | 5.56 | 6.8 | 1.172x |\n| decode 16 | m 16, layer pooler | 62 | 14.06 | 16.06 | 21.4 | 1.198x |","src":"bench/product/harbor/tasks/dense_proj_bf16--all/instruction.md"},"run_x":1.604270887718952,"run_src":"bench/runs/muse_bf16_proj/2c207d7414734787.json (best passed judge run); floors from bench/tools/gap.py phys()","cases":[{"load":"decode 1","col":"qkv_gate","variant":"qkv_gate m1","w":62,"ours_us":8.563,"sglang_us":14.336,"floor_us":7.554,"ours_pct":88.2,"sglang_pct":52.7},{"load":"decode 1","col":"o","variant":"o m1","w":62,"ours_us":5.207,"sglang_us":7.072,"floor_us":4.392,"ours_pct":84.3,"sglang_pct":62.1},{"load":"decode 1","col":"pooler","variant":"pooler m1","w":62,"ours_us":13.373,"sglang_us":21.616,"floor_us":11.98,"ours_pct":89.6,"sglang_pct":55.4},{"load":"decode 1","col":"unpooler","variant":"unpooler m1","w":62,"ours_us":3.306,"sglang_us":6.816,"floor_us":3.126,"ours_pct":94.6,"sglang_pct":45.9},{"load":"decode 1","col":"router","variant":"router m1","w":62,"ours_us":3.694,"sglang_us":10.464,"floor_us":3.127,"ours_pct":84.7,"sglang_pct":29.9},{"load":"decode 16","col":"qkv_gate","variant":"qkv_gate m16","w":62,"ours_us":9.488,"sglang_us":14.144,"floor_us":7.601,"ours_pct":80.1,"sglang_pct":53.7},{"load":"decode 16","col":"o","variant":"o m16","w":62,"ours_us":5.523,"sglang_us":6.816,"floor_us":4.434,"ours_pct":80.3,"sglang_pct":65.1},{"load":"decode 16","col":"pooler","variant":"pooler m16","w":62,"ours_us":13.533,"sglang_us":21.44,"floor_us":12.036,"ours_pct":88.9,"sglang_pct":56.1},{"load":"decode 16","col":"unpooler","variant":"unpooler m16","w":62,"ours_us":13.648,"sglang_us":18.72,"floor_us":12.036,"ours_pct":88.2,"sglang_pct":64.3},{"load":"decode 16","col":"router","variant":"router m16","w":62,"ours_us":4.305,"sglang_us":14.976,"floor_us":3.167,"ours_pct":73.6,"sglang_pct":21.1},{"load":"decode 32","col":"qkv_gate","variant":"qkv_gate m32","w":62,"ours_us":10.272,"sglang_us":14.464,"floor_us":7.652,"ours_pct":74.5,"sglang_pct":52.9},{"load":"decode 32","col":"o","variant":"o m32","w":62,"ours_us":5.815,"sglang_us":6.992,"floor_us":4.478,"ours_pct":77.0,"sglang_pct":64.0},{"load":"decode 32","col":"pooler","variant":"pooler m32","w":62,"ours_us":14.837,"sglang_us":21.92,"floor_us":12.095,"ours_pct":81.5,"sglang_pct":55.2},{"load":"decode 32","col":"unpooler","variant":"unpooler m32","w":62,"ours_us":14.49,"sglang_us":18.688,"floor_us":12.095,"ours_pct":83.5,"sglang_pct":64.7},{"load":"decode 32","col":"router","variant":"router m32","w":62,"ours_us":4.857,"sglang_us":15.072,"floor_us":3.209,"ours_pct":66.1,"sglang_pct":21.3},{"load":"verify 1","col":"qkv_gate","variant":"qkv_gate m8","w":62,"ours_us":9.141,"sglang_us":14.384,"floor_us":7.576,"ours_pct":82.9,"sglang_pct":52.7},{"load":"verify 1","col":"o","variant":"o m8","w":62,"ours_us":5.225,"sglang_us":6.864,"floor_us":4.411,"ours_pct":84.4,"sglang_pct":64.3},{"load":"verify 1","col":"pooler","variant":"pooler m8","w":62,"ours_us":13.314,"sglang_us":21.152,"floor_us":12.006,"ours_pct":90.2,"sglang_pct":56.8},{"load":"verify 1","col":"unpooler","variant":"unpooler m8","w":62,"ours_us":4.015,"sglang_us":6.912,"floor_us":3.137,"ours_pct":78.1,"sglang_pct":45.4},{"load":"verify 1","col":"router","variant":"router m8","w":62,"ours_us":4.262,"sglang_us":26.24,"floor_us":3.146,"ours_pct":73.8,"sglang_pct":12.0},{"load":"verify 16","col":"qkv_gate","variant":"qkv_gate m128","w":62,"ours_us":11.804,"sglang_us":16.832,"floor_us":7.956,"ours_pct":67.4,"sglang_pct":47.3},{"load":"verify 16","col":"o","variant":"o m128","w":62,"ours_us":7.146,"sglang_us":7.52,"floor_us":4.745,"ours_pct":66.4,"sglang_pct":63.1},{"load":"verify 16","col":"pooler","variant":"pooler m128","w":62,"ours_us":17.315,"sglang_us":25.28,"floor_us":12.451,"ours_pct":71.9,"sglang_pct":49.3},{"load":"verify 16","col":"unpooler","variant":"unpooler m128","w":62,"ours_us":17.497,"sglang_us":19.872,"floor_us":12.451,"ours_pct":71.2,"sglang_pct":62.7},{"load":"verify 16","col":"router","variant":"router m128","w":62,"ours_us":5.854,"sglang_us":16.288,"floor_us":3.46,"ours_pct":59.1,"sglang_pct":21.2},{"load":"verify 32","col":"qkv_gate","variant":"qkv_gate m256","w":62,"ours_us":17.35,"sglang_us":19.728,"floor_us":12.597,"ours_pct":72.6,"sglang_pct":63.9},{"load":"verify 32","col":"o","variant":"o m256","w":62,"ours_us":8.846,"sglang_us":9.488,"floor_us":6.632,"ours_pct":75.0,"sglang_pct":69.9},{"load":"verify 32","col":"pooler","variant":"pooler m256","w":62,"ours_us":23.99,"sglang_us":30.928,"floor_us":20.949,"ours_pct":87.3,"sglang_pct":67.7},{"load":"verify 32","col":"unpooler","variant":"unpooler m256","w":62,"ours_us":22.938,"sglang_us":24.288,"floor_us":20.949,"ours_pct":91.3,"sglang_pct":86.3},{"load":"verify 32","col":"router","variant":"router m256","w":62,"ours_us":7.016,"sglang_us":17.44,"floor_us":3.796,"ours_pct":54.1,"sglang_pct":21.8}],"lost":[{"load":"verify 16","col":"unpooler","variant":"unpooler m128","w":62,"ours_us":17.497,"sglang_us":19.872,"floor_us":12.451,"ours_pct":71.2,"sglang_pct":62.7},{"load":"verify 16","col":"pooler","variant":"pooler m128","w":62,"ours_us":17.315,"sglang_us":25.28,"floor_us":12.451,"ours_pct":71.9,"sglang_pct":49.3},{"load":"verify 32","col":"qkv_gate","variant":"qkv_gate m256","w":62,"ours_us":17.35,"sglang_us":19.728,"floor_us":12.597,"ours_pct":72.6,"sglang_pct":63.9}],"loads":["decode 1","decode 16","decode 32","verify 1","verify 16","verify 32"],"cols":["qkv_gate","o","pooler","unpooler","router"],"top_class":"STRUCTURAL-REWRITE","top_class_share":0.513552449373005,"first_x":1.1769,"final":{"t":"2026-11-05T16:46:28Z","x":1.5375097993769167,"rank":313,"entries":1747,"leader":"gogowhy","leader_x":1.7276592188096174,"gap_pct":-11.01,"pct_ceiling":77.3},"ceiling":1.99}],"derive":{"problem":"Attention decode","case":{"load":"verify 1","col":"full","variant":"b1 ctx80000 full q8","w":16,"ours_us":13.993,"sglang_us":36.752,"floor_us":7.731,"ours_pct":55.2,"sglang_pct":21.0}},"test":{"late_reruns":62,"late_with_sha":72,"records":209,"gain_lt_1pct":130},"improve":{"problem":"Decode layer","first":1.1682,"final":2.4513545799748853,"ceiling":3.02}},"methods":[{"name":"First principles / floor","p":0.1465,"k":29,"n":198,"lo":0.1099,"hi":0.1926},{"name":"Rival / peer code","p":0.1316,"k":10,"n":76,"lo":0.0803,"hi":0.2082},{"name":"Measurement / oracle / profiling","p":0.1105,"k":20,"n":181,"lo":0.0778,"hi":0.1547},{"name":"No source given","p":0.1079,"k":30,"n":278,"lo":0.081,"hi":0.1424},{"name":"Port / amplify a win","p":0.1047,"k":9,"n":86,"lo":0.0619,"hi":0.1715},{"name":"Combination engine","p":0.0701,"k":11,"n":157,"lo":0.0433,"hi":0.1114},{"name":"Lane's own work (origin not stated)","p":0.0658,"k":101,"n":1535,"lo":0.0561,"hi":0.077},{"name":"Evolution search","p":0.0408,"k":2,"n":49,"lo":0.0136,"hi":0.1161},{"name":"Knob sweep","p":0.0163,"k":2,"n":123,"lo":0.0054,"hi":0.0479},{"name":"Codex island","p":0.0157,"k":14,"n":890,"lo":0.0102,"hi":0.0242}],"papers":[0,22],"start":{"ideas":14936,"gpu_runs":58618,"proven":236,"first_x":1.1682,"final_x":2.4513545799748853},"ladder":[{"family":"dense_proj_bf16","T1":50.3,"T2":77.3,"T3":86.8,"T4":90.0},{"family":"moe_nvfp4_ep8","T1":32.6,"T2":79.1,"T3":92.6,"T4":95.6},{"family":"attn_decode_fp8kv","T1":29.6,"T2":83.9,"T3":98.8,"T4":101.8},{"family":"decoder_layer_tp8","T1":33.1,"T2":81.2,"T3":86.6,"T4":90.0},{"family":"prefill_layer","T1":44.1,"T2":84.6,"T3":94.6,"T4":97.6}],"hard_set":[{"family":"decoder_layer_tp8","load":"verify 32","harbor_task":"decoder_layer_tp8--verify-32","ours_pct_of_floor":67.8,"sglang_pct_of_floor":26.7,"floor_us":221.82,"cases":2},{"family":"dense_proj_bf16","load":"verify 16","harbor_task":"dense_proj_bf16--verify-16","ours_pct_of_floor":68.9,"sglang_pct_of_floor":47.9,"floor_us":8.21,"cases":5},{"family":"attn_decode_fp8kv","load":"verify 1","harbor_task":"attn_decode_fp8kv--verify-1","ours_pct_of_floor":73.4,"sglang_pct_of_floor":19.8,"floor_us":5.3,"cases":2},{"family":"decoder_layer_tp8","load":"verify 16","harbor_task":"decoder_layer_tp8--verify-16","ours_pct_of_floor":75.2,"sglang_pct_of_floor":31.3,"floor_us":183.62,"cases":2},{"family":"moe_nvfp4_ep8","load":"verify 32","harbor_task":"moe_nvfp4_ep8--verify-32","ours_pct_of_floor":78.3,"sglang_pct_of_floor":25.1,"floor_us":159.28,"cases":1},{"family":"dense_proj_bf16","load":"decode 32","harbor_task":"dense_proj_bf16--decode-32","ours_pct_of_floor":78.6,"sglang_pct_of_floor":51.2,"floor_us":7.91,"cases":5}],"hard_rule":"loads where our best judge run stayed below 80% of the physical floor; loads with an uncertain floor binding ('F?') or a floor above 100% excluded","fleet_title":"24-hour records per 100 lane-hours, by model and harness","fleet_caption":"Lane-hours are summed session time. Claude Code, Codex and Antigravity sessions are rebuilt from GPU activity, which favours them against pi lanes. Associations from 14 days of mixed roles and phases, not a controlled comparison: early lanes found easy wins, late lanes worked a mined problem, and cheap models mostly re-ran other models' files. Whiskers: 90% Poisson interval. The cold-start tab is the controlled protocol.","stats":{"families":9,"tasks":85,"gpu_runs":58519,"models":9}}}