lqf0624 Artifact Audit
Audit the paper like an artifact evaluator.
Read ../researchstack/references/evidence-contracts.md and preserve the claim-promotion states during the audit.
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.
- whether the timed and semantic boundary matches the real deployment lifecycle.
Produce:
- Provenance table mapping claim -> experiment -> script/config -> output artifact.
- Reproducibility gaps.
- Environment assumptions.
- Highest-risk silent failure modes.
- Claim-promotion downgrades where provenance or controls are insufficient.
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.
- required planning, packing, synchronization, waiting, or assembly work excluded from headline timing.