guanning/arc-agi-3-schema-traces-gpt56

ARC-AGI-3 Schema Gameplay Trajectories — GPT-5.6 Sol

This release contains every gpt-5.6-sol gameplay trajectory produced on our cluster with the world_model_v5 agent harness — 100 runs across the 25 public ARC-AGI-3 games — plus a dependency-free scoring utility.

It is the GPT-5.6 Sol member of a family built by the same harness and the same sanitizer, so trajectories can be compared game by game:

• arc-agi-3-schema-traces-fable5 — Claude Fable 5, best per game (25)
• arc-agi-3-schema-traces-opus48 — Claude Opus 4.8, best per game (25)
• this one — GPT-5.6 Sol, the full sweep (100)

This release differs from its siblings in two ways. It is not a best-of: a game appears once per attempt, so you can see variance across repeated runs of the same game (ls20 has 6 attempts; sp80 has 10). And it is split by reasoning effort, because effort turns out to matter more than anything else here.

Layout

arc-agi-3-schema-traces-gpt56/
├── README.md
├── baseline_actions.csv
├── score_trajectories.py
├── gpt_5_6_sol_max/
│   ├── evaluation_results.csv
│   └── <39 trajectory directories>/
├── gpt_5_6_sol_xhigh/
│   ├── evaluation_results.csv
│   └── <54 trajectory directories>/
└── gpt_5_6_sol_unknown/
    ├── evaluation_results.csv
    └── <7 trajectory directories>/

Trajectory directories are named gpt-5.6-sol_<effort>_<task>_<rhae>, with a trailing _2, _3, … when repeated attempts on the same game reach the same score. Each contains run.json, a streamed events.jsonl event log, sanitized session data, snapshots, and the shareable files the agent produced.

About the unknown effort

unknown is not a guess and not a third setting. It means the reasoning effort that was actually passed to the model was not recorded anywhere we could recover: no launch manifest row, no effort tag in the run directory name, and run.json never stored the value. Seven trajectories are in that state. They are published rather than dropped, but do not pool them with either named effort.

baseline_actions.csv holds the human action baselines (identical across this dataset family). Each evaluation_results.csv is a manifest of its collection; its level0…level9 columns list the action counts of completed levels only.

Recompute all scores

Python 3.10 or newer is recommended. The scorer uses only the Python standard library, so no packages need to be installed.

python3 score_trajectories.py

The command discovers all trajectory directories, streams all 100 events.jsonl files, reconstructs per-level action counts, recomputes every RHAE score, and prints a 100-row table followed by a per-collection summary. By default it also verifies that the event-derived actions and scores match each evaluation_results.csv.

python3 score_trajectories.py --compact        # narrower output
python3 score_trajectories.py --expected 0     # allow a different count
python3 score_trajectories.py --no-manifest-check

The default command exits nonzero if a log is malformed, an action sequence is not contiguous, a baseline is missing, the collections do not hold exactly 100 trajectories, or a recomputed result differs from a manifest.

Scoring

For completed level i, with human baseline actions h_i and trajectory actions a_i, the per-level score is:

level_score_i = min(115, 100 * (h_i / a_i)^2)

Incomplete or missing levels receive zero. The raw game score is the weighted mean of the level scores, using the one-based level number as its weight. A completion cap prevents unfinished games from receiving more credit than the weighted share of levels they completed:

raw_game_score = weighted_mean(level_score_i, weight=i)
completion_cap = 100 * sum(i for completed levels) / sum(i for all levels)
RHAE           = min(raw_game_score, completion_cap)

The 115% per-level cap permits a more action-efficient trajectory to offset a less efficient level, while the final game score remains capped at 100%.

Verified release summary

Running the scorer on the included data produces:

Collection	Trajectories	Wins	Levels	Mean RHAE
gpt_5_6_sol_max	39	25	245/291	67.26%
gpt_5_6_sol_xhigh	54	46	340/387	82.26%
gpt_5_6_sol_unknown	7	2	26/49	26.67%
ALL	100	73	611/727	72.52%

xhigh is both the stronger and the better-covered setting: 85% of its runs win, it is the only effort with at least one trajectory on all 25 games, and it reaches that coverage in far fewer actions than max. Read the max column with that in mind — most max attempts were aimed at a handful of hard games.

Best RHAE per game and effort

Parenthesised numbers are how many attempts that game has at that effort. — means no trajectory at that effort.

Game	Levels	max	xhigh	unknown
ar25	8	—	100.00 (2)	—
bp35	9	60.93 (3)	2.22 (1)	40.57 (1)
cd82	6	—	100.00 (2)	—
cn04	6	—	100.00 (3)	—
dc22	6	100.00 (2)	100.00 (1)	0.00 (1)
ft09	6	—	100.00 (3)	—
g50t	7	—	100.00 (2)	—
ka59	7	65.34 (3)	0.00 (1)	63.38 (1)
lf52	10	100.00 (2)	100.00 (1)	—
lp85	8	—	100.00 (3)	—
ls20	7	—	100.00 (6)	—
m0r0	6	0.00 (1)	100.00 (2)	—
r11l	6	—	100.00 (2)	—
re86	8	0.00 (1)	100.00 (2)	—
s5i5	8	—	100.00 (2)	—
sb26	8	—	100.00 (2)	—
sc25	6	82.72 (3)	42.53 (1)	18.26 (1)
sk48	8	87.80 (2)	2.78 (1)	50.18 (1)
sp80	6	100.00 (7)	100.00 (3)	14.29 (1)
su15	9	100.00 (4)	80.31 (1)	—
tn36	7	87.02 (7)	80.27 (3)	0.00 (1)
tr87	6	—	100.00 (3)	—
tu93	9	—	100.00 (2)	—
vc33	7	—	100.00 (2)	—
wa30	9	100.00 (4)	100.00 (3)	—

Taking the best trajectory per game regardless of effort, GPT-5.6 Sol reaches 100.00 on 20 of the 25 games.

Provenance and integrity

• Every RHAE here was recomputed from the sanitized events.jsonl with the official arc_agi.scorecard.EnvironmentScoreCalculator; no value was copied from a result CSV.
• All 100 sanitized events.jsonl files were independently replayed against the trusted offline engine (arcengine, seed=0, ONLY_RESET_LEVELS=true) and reproduce exactly (grid, state, level) — see the verification note below.
• 25 of the trajectories are byte-identical to the gpt_5_6_sol collection of arc-agi-3-schema-gameplay; the other 75 were sanitized for this release.
• Runs that produced zero actions are excluded: two launches where the provider never returned a usable action and the harness's no-commit breaker ended the run. They contain no gameplay.
• Duplicate copies of the same run (the same game_id and started_at reachable through several paths on disk) were collapsed to one trajectory.

A note on timestamps

Timestamps were shifted by one offset per source batch, and this release mixes two batches (the 25 republished trajectories keep the offset of the earlier release). Timestamps are therefore comparable within a trajectory, but a delta computed between trajectories from different batches is meaningless.

A note on console.log

The harness wrote a host-side console.log next to each run. It is not included here — it is not part of any release in this family. In two wa30 trajectories the agent greps that file; the recorded tool calls and their results are kept, because they are a faithful record of what the agent did. What came back was the agent's own print() output from its world-model code, not engine internals or game source.