sheepy928/digital-maxima-checkpoints

Digital MAXIMA agent run checkpoints (gated)

Private or gated with manual approval only. The run records hold scores on the benchmark's hidden splits. Approve only people who run no agents on this benchmark, never an account or token an agent run can use. See DECISIONS D-063 in the project repository.

• checkpoints.bundle (3.02 GB): the code history of 144 finished agent runs in one git bundle. Each run's refs are under refs/runs/<run id>/: heads/main (the agent's branch), harness/snapshots (the harness's periodic snapshots) and HEAD (the commit the run ended on).
• index.jsonl: per run, the final commit, every ref, the commit count and the audit outcome.
• runs/: 13732 record files from those runs and 13 RSI lineages: manifests, configs, audits, transcripts, egress logs, hidden-split scores, per-checkpoint and ablation scores. Prediction arrays, workspace tarballs and exported trees are not included; they are regenerable from the code.
• SHA256SUMS: checksums of every file above.

Restore one run's workspace

hf download <repo> --repo-type dataset --local-dir <dir>
uv run python -m mtb.harness.code_bundle verify <dir>
uv run python -m mtb.harness.code_bundle restore <dir> <run id> runs/<run id>/workspace

Without the project code: git init ws && git -C ws fetch --update-head-ok <dir>/checkpoints.bundle '+refs/runs/<run id>/*:refs/*' fetches the run's refs (its HEAD ref lands as refs/HEAD).