evaluation-leakage-audit

• SKILL.mdcore · 10,495 bytes
• SKILL.mdcore · 10,495 bytes
• SKILL.mdcore · 10,495 bytes
• SKILL.mdcore · 10,495 bytes
• agents/openai.yamlasset · 248 bytes

# Evaluation leakage audit

## Contract

| Field | Bound contract |
|---|---|
| Trigger | Building or reviewing an evaluation, benchmark, or scoring harness, or an explicit request to check it for leakage or contamination (leakage mode); or a validation plan or result meant to show whether something worked, such as an A/B test, holdout, or score, looks too clean or self-confirming (self-audit mode). |
| Authority | Read-only: inspects supplied evidence and reports in chat; writes nothing, so no rollback is needed. No remote mutation. |
| Side effect | Return only a report naming where independent ground truth enters and where it does not, with an independence fix for each detected pattern; self-audit mode also names the audit's own non-independence and one root failure. |
| Done | Every applicable pattern below has been tested; only detected patterns are reported; each finding has evidence and an independence fix; self-audit mode also applies the auditor-bias patterns to the audit itself and names one root independence failure; if none fire, the report says no leak was found. |

## Inputs

- Mode. `leakage` or `self-audit`; when omitted, step 1 selects it: `self-audit` when the subject is a validation result meant to show whether something worked, `leakage` otherwise. Optional.