benchmark-audit

Benchmark Audit Strategy

Systematic quality assessment of AI/ML benchmarks using the BetterBench 46-criterion framework, Datasheets for Datasets standards, and established psychometric evaluation principles.

Purpose

Produce a structured quality report for each target benchmark covering: documentation completeness, construct validity indicators, statistical robustness, maintenance status, and known failure modes.

Budget

State Ledger

<HARD-GATE>
| Metric | Current | Target | Status |
|--------|---------|--------|--------|
| Benchmarks audited | 0 | 5 | PENDING |
| Papers fetched | 0 | 30 | PENDING |
| Papers read | 0 | 20 | PENDING |
| Web searches | 0 | 40 | PENDING |
| Documentation audits complete | 0 | 5 | PENDING |
| Metric decompositions complete | 0 | 5 | PENDING |
| Contamination checks complete | 0 | 5 | PENDING |
| Synthesis reports produced | 0 | 5 | PENDING |
</HARD-GATE>

Cannot exit until 80% of all targets met.

Available Tactics

• artifact-detection — Probe for annotation artifacts and dataset shortcuts

Available SOPs

• benchmark-inventory — Identify target benchmarks in domain
• metric-decomposition — Decompose composite metrics into constituent signals
• contamination-audit — Detect train-test data leakage
• documentation-audit — Assess documentation completeness (BetterBench/Datasheets)
• benchmark-synthesis — Produce final structured audit report

Execution Guidance

1. Inventory Phase: Use benchmark-inventory to identify 5 benchmarks in target domain
2. Per-Benchmark Loop (repeat for each benchmark): a. Gather benchmark paper, documentation, leaderboard via web searches b. Run documentation-audit against BetterBench 46 criteria c. Run metric-decomposition on primary metric(s) d. Run contamination-audit checking known training corpora e. Run artifact-detection tactic if annotation-based benchmark f. Collect findings into per-benchmark report
3. Synthesis Phase: Run benchmark-synthesis to produce cross-benchmark comparison