guanning/arc-agi-3-schema-traces-opus48

ARC-AGI-3 Schema Gameplay Trajectories — Claude Opus 4.8

This release contains the best claude-opus-4-8 / max trajectory for each of the 25 public ARC-AGI-3 games, plus a dependency-free scoring utility. It is the Opus 4.8 counterpart of arc-agi-3-schema-traces-fable5, produced by the same agent harness (world_model_v5) and the same sanitizer, so the two collections can be compared game by game.

Each trajectory directory includes run.json, a streamed events.jsonl event log, sanitized session data, snapshots, and the shareable text/image files produced during the run.

Unlike the Fable 5 release, this one is not a clean sweep. Opus 4.8 never finished 3 of the 25 games, and those trajectories are published as they ended rather than being dropped or replaced. That is the honest ceiling of this model on this harness, and the incomplete runs are often the more interesting ones.

Layout

arc-agi-3-schema-traces-opus48/
├── README.md
├── baseline_actions.csv
├── score_trajectories.py
└── claude_opus_4_8_max/
    ├── evaluation_results.csv
    └── <25 trajectory directories>/

baseline_actions.csv contains the human action baselines (identical to the other releases in this family). evaluation_results.csv is a compact manifest of the 25 trajectories; 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 25 events.jsonl files, reconstructs per-level action counts, recomputes every RHAE score, and prints one 25-row table followed by a summary table. By default it also verifies that the event-derived actions and scores match evaluation_results.csv.

Useful options:

# Narrower terminal output
python3 score_trajectories.py --compact

# Score a copy located elsewhere
python3 score_trajectories.py --root ~/agent-dataset/arc-agi-3-schema-traces-opus48

# Allow a trajectory count other than 25
python3 score_trajectories.py --expected 0

# Recompute from events without checking the CSV manifest
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 collection does not contain exactly 25 trajectories, or a recomputed result differs from the 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%.

In the terminal table, Level actions lists only completed-level action counts in order.

Verified release summary

Running the scorer on the included data produces:

Collection	Trajectories	Wins	Levels	Mean RHAE
claude_opus_4_8_max	25	22	175/183	86.07%

Per game, highest first:

Game	State	Levels	Actions	Human baseline	RHAE
ar25	WIN	8/8	269	748	100.00%
cd82	WIN	6/6	121	171	100.00%
cn04	WIN	6/6	479	789	100.00%
ft09	WIN	6/6	94	208	100.00%
g50t	WIN	7/7	544	879	100.00%
ka59	WIN	7/7	431	730	100.00%
lp85	WIN	8/8	134	388	100.00%
ls20	WIN	7/7	642	776	100.00%
m0r0	WIN	6/6	221	1107	100.00%
r11l	WIN	6/6	83	233	100.00%
re86	WIN	8/8	615	1255	100.00%
tr87	WIN	6/6	138	414	100.00%
tu93	WIN	9/9	243	462	100.00%
wa30	WIN	9/9	956	1843	100.00%
sc25	WIN	6/6	363	350	93.88%
s5i5	WIN	8/8	643	638	89.87%
sk48	WIN	8/8	801	1070	85.05%
sb26	WIN	8/8	347	213	82.51%
vc33	WIN	7/7	507	447	81.84%
tn36	WIN	7/7	348	317	75.26%
bp35	WIN	9/9	1265	651	62.91%
su15	WIN	9/9	444	361	61.51%
sp80	ended	5/6	3019	518	56.34%
dc22	stopped	4/6	1539	1228	38.86%
lf52	stopped	5/10	739	1339	23.76%

The two non-WIN states distinguish how a run ended, and neither means the agent hit a level it could not pass:

• ended (sp80) — the harness closed the run itself after the agent exhausted its action budget on level 5.
• stopped (dc22, lf52) — the run was still playing when its wall-clock budget expired, so there is no closing event.

sk48 was in that state too — one level short, at 77.78%. It was later continued with --resume: the harness replays the recorded actions into a fresh engine to restore the live game, then plays on. It won the last level in 49 more actions. The trajectory published here is that continued run; its events.jsonl carries both sessions, 801 actions end to end, and replays exactly.

Provenance and integrity

• Every RHAE above was recomputed from the sanitized events.jsonl with the official arc_agi.scorecard.EnvironmentScoreCalculator; no value was copied from a result CSV.
• All 25 sanitized events.jsonl files were independently replayed against the trusted offline engine (arcengine, seed=0, ONLY_RESET_LEVELS=true): 14985/14985 transitions exact (grid, state, level).
• 14 of the trajectories are byte-identical to the claude-opus-4-8 subset of arc-agi-3-schema-gameplay; the other 11 were sanitized for this release.

A note on timestamps

Timestamps were shifted by one offset per source batch, and this release mixes two batches (the 14 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 bp35

An Opus 4.8 run scored higher on bp35 (79.63%) using a different agent architecture. The trajectory published here is the best run from the same world_model_v5 harness as the other 24, so that every number in this release is comparable.