artifact-auditor

name	artifact-auditor
description	Audit experiment artifact folders for completeness, schema consistency, and promotion traceability across single and grid runs.

Artifact Auditor

Use this skill when validating whether research runs are publishable/reviewable by another quant without hidden notebook state.

What this skill does

1. Verifies required files exist for each run.
2. Checks expected columns in metrics.parquet and leaderboard.parquet.
3. Confirms decision traceability (decision.md + run_manifest.json).
4. Reports missing files/fields with actionable remediation.

Required files (single run)

• run_manifest.json
• metrics.parquet
• events.parquet
• signals.parquet
• weights.parquet
• backtest.parquet
• decision.md

Required files (grid run)

• top-level leaderboard.parquet
• top-level run_manifest.json
• sub-run directories each containing the single-run file set

Suggested audit commands

# quick schema check (example)
uv run --python .venv/bin/python - <<'PY'
import polars as pl
from pathlib import Path
p = Path("experiments/artifacts")
for lf in p.rglob("leaderboard.parquet"):
    df = pl.read_parquet(lf)
    required = {"run_name", "net_sharpe", "promote_candidate"}
    missing = required - set(df.columns)
    if missing:
        print(f"{lf}: missing {sorted(missing)}")
PY