guanning/arc3-ablation-runs

ARC-AGI-3 ablation runs

Fifty complete agent runs on ARC-AGI-3: five harness configurations across the ten games with the highest human baseline action counts. Each archive is one run — every action, every model revision, and the full LLM transcript that produced them.

Configurations

arm	what is removed
ctrl2	the full harness
nobt	--no-backtest: cannot replay history to check a model revision
nobfs	--no-planning: cannot search inside the model for a route
nompc	--no-mpc: no rolling re-planning
text	--text-worldmodeling: the world model is prose in notes.md; no code at all

Files

<game>_<arm>.tar.zst unpacks to two directories:

• the run workdir — events.jsonl (every action, tool call and model revision), history.jsonl, world_model_v5.py (the final model), notes.md, snapshots/ (the model at each level cleared), .git/ (one commit per turn, so every intermediate model is recoverable), and sessions/ (the complete LLM transcripts);
• .rt/ — the sandbox sidecar: logs/ and src/, the harness source as it stood inside the jail.

__pycache__/ and session_live/ (an in-flight duplicate of sessions/) are excluded.

manifest.csv carries the results: per run the job id, step and turn counts, level reached, status, score, per-level action counts, and a censored flag for runs cut short.