| name | researchstack-artifact-audit |
| description | Reproducibility and evidence-audit skill for research projects. Use when Codex should trace figures and claims back
to scripts, configs, datasets, seeds, hardware assumptions, and logs; find missing provenance; or prepare the code
and experiment pipeline for submission, open-sourcing, or artifact evaluation.
|
Artifact Audit
Audit the paper like an artifact evaluator.
Check for each reported result:
- where the number came from,
- which code path produced it,
- which config or seed was used,
- whether the environment is documented,
- whether reruns are expected to match exactly or statistically.
Produce:
- Provenance table mapping claim -> experiment -> script/config -> output artifact.
- Reproducibility gaps.
- Environment assumptions.
- Highest-risk silent failure modes.
Be especially strict about:
- cached outputs with unclear freshness,
- figures hand-edited after generation,
- hidden dataset filtering,
- mismatched train/test or warmup/steady-state conditions,
- benchmark harnesses that favor the proposed method,
- copied numbers that no longer match the current codebase.