duck-harness-kaggle

⏳ Run in progress. This page was generated by the periodic save loop while the benchmark is still executing (started 2026-10-25 14:26:01, this snapshot written 2026-10-25 15:06:08). Numbers, plots, and movies reflect a partial snapshot — they will be rewritten when the run finishes.

Diagnostics for one TAAF benchmark run — a single solver played each game over a fixed number of passes. Plots track score-vs-cost across the run; the table lists every (game, pass) cell, with the movie cell linking to a video of that play-through.

Solver: duck-harness

10 games × 3 passes = 30 runs

Started: 2026-10-25 14:26:01

Status: in progress

git status
git status:
  ARC3-Inference                   aa69123     DIRTY  add-kaggle-share-flag     Add flag to use shareable notebook
  tufa-arc-agi-framework           fe9f7c4     clean  submission-share-mode-bugfix  Fix dataset mapping, spell Isaiah's name properly
  re-arc-3                         57e46d619d  pinned  master                    (non-editable)

Per-run stats

runpassesmeanσ across passesSEM
duck-harness-kaggle33.155.453.15

Score vs cumulative tokens

What this curve shows. Reading the curve at x = X tells you the expected score per game when the method is run at an average per-game budget of X tokens. Slow games may use more, fast games less; the average works out to X.

score vs tokens

What-if: per-level token caps

Each curve shows what the score-vs-tokens curve would have been if every level had been stopped at the listed per-level token cap. The cap is on cumulative tokens spent since the most recent level win: cross the cap without winning the level and the lane truncates there with no further credit (full credit is kept for previously-won levels). 100% is the largest token total any single level actually consumed in this run (won or not); the other curves are fractional cuts of that.

score vs tokens, per-level caps

Score vs wallclock

Design note: we still have to think about how to plot this exactly when we have some real benchmark data — games run in parallel during the benchmark, so wallclock per game is not just total wallclock / n_games.
score vs wallclock

Total generated tokens vs job wallclock

total tokens vs job wallclock

Per-game score vs tokens

per-game score vs tokens scatter

Per-game

initialar25-0c556536
ar25-0c556536 tokensar25-0c556536 wallclock
initialft09-0d8bbf25
ft09-0d8bbf25 tokensft09-0d8bbf25 wallclock
initiallp85-305b61c3
lp85-305b61c3 tokenslp85-305b61c3 wallclock
initialr11l-495a7899
r11l-495a7899 tokensr11l-495a7899 wallclock
initialre86-8af5384d
re86-8af5384d tokensre86-8af5384d wallclock
initialsb26-7fbdac44
sb26-7fbdac44 tokenssb26-7fbdac44 wallclock
initialsc25-635fd71a
sc25-635fd71a tokenssc25-635fd71a wallclock
initialtr87-cd924810
tr87-cd924810 tokenstr87-cd924810 wallclock
initialtu93-0768757b
tu93-0768757b tokenstu93-0768757b wallclock
initialvc33-5430563c
vc33-5430563c tokensvc33-5430563c wallclock

Per-pass per-game results

gamepassstatescorelevelactionstokensmovienote
initialar25-0c5565360playing27.784/8135105277—
initialar25-0c5565361playing0.000/800—
initialar25-0c5565362playing0.000/800—
initialft09-0d8bbf250playing0.000/600—
initialft09-0d8bbf251playing0.000/600—
initialft09-0d8bbf252playing0.000/600—
initiallp85-305b61c30gave_up0.000/800—analyzer failed 11 times consecutively
initiallp85-305b61c31playing0.000/800—
initiallp85-305b61c32playing0.000/800—
initialr11l-495a78990playing47.624/641119289—
initialr11l-495a78991playing0.000/600—
initialr11l-495a78992playing0.000/600—
initialre86-8af5384d0playing0.000/800—
initialre86-8af5384d1playing0.000/800—
initialre86-8af5384d2playing0.000/800—
initialsb26-7fbdac440playing0.000/800—
initialsb26-7fbdac441playing0.000/800—
initialsb26-7fbdac442playing0.000/800—
initialsc25-635fd71a0playing19.063/6109104933—
initialsc25-635fd71a1playing0.000/600—
initialsc25-635fd71a2playing0.000/600—
initialtr87-cd9248100playing0.000/600—
initialtr87-cd9248101playing0.000/600—
initialtr87-cd9248102playing0.000/600—
initialtu93-0768757b0gave_up0.000/900—analyzer failed 11 times consecutively
initialtu93-0768757b1playing0.000/900—
initialtu93-0768757b2playing0.000/900—
initialvc33-5430563c0gave_up0.000/710finalanalyzer failed 11 times consecutively
initialvc33-5430563c1playing0.000/700—
initialvc33-5430563c2playing0.000/700—