| name | audit-assistance-provenance |
| description | Audit a run, strategy, tool, dataset, prompt, or report for hidden operational assistance H, knowledge assistance K, engine leakage, retries, and attribution errors. |
Audit assistance and provenance
Audit path
- Trace every input visible before the final action.
- Classify parser/rules/environment/legal-actions/model-only-search/engine involvement as H0-H7.
- Classify generic/phase/opening/retrieved/current-position knowledge as K0-K7.
- Compare declared ceiling to effective events.
- Inspect SAN
+/#, top-k, evaluations, PVs, candidate labels and tool outputs for leakage.
- Inspect SDK/runtime for hidden retry, fallback, routing or cache.
- Verify dataset/training-derived engine information is distinguished from live assistance.
- Verify report and leaderboard names match the system, not only the base model.
- Emit violations with evidence event/artifact IDs.
- Block claim/publication when attribution cannot be reconstructed.
Output
assistance_audit:
declared: {R: R3, H: H3, K: K4}
effective: {R: R3, H: H5, K: K4}
status: protocol_violation
findings: []