Handbook
Prompt 29 — Clean-room contamination and solution-leakage auditor
Audit whether the learner or curriculum received information outside the declared clean-room boundary.
Updated
Inputs
FROZEN_SNAPSHOT_ID: {{FROZEN_SNAPSHOT_ID}}
AUDITED_ARTIFACTS: {{AUDITED_ARTIFACTS_JSON}}
ALLOWLISTED_SOURCE_AND_LIBRARY_IDS: {{ALLOWLISTED_SOURCE_AND_LIBRARY_IDS_JSON}}
RETRIEVAL_AND_TOOL_TRACE: {{RETRIEVAL_AND_TOOL_TRACE_JSON}}
HIDDEN_ALIAS_AND_TARGET_FINGERPRINTS: {{HIDDEN_ALIAS_AND_TARGET_FINGERPRINTS_JSON}}
MODEL_AND_CACHE_PROVENANCE: {{MODEL_AND_CACHE_PROVENANCE_JSON}}
Task
Audit whether the learner or curriculum received information outside the declared clean-room boundary.
Check for:
- modern law, theory, quantity, experiment, or constant names;
- canonical equation forms, notation, numerical values, or standard substitutions;
- solution templates and target-revealing metadata;
- unauthorized reference-library retrieval;
- simulator ground-truth or evaluator leakage;
- cross-session cache, embedding-index, tool-output, or prompt contamination;
- suspiciously exact historical phrasing or canonical derivation order;
- pretraining contamination indicators that cannot be ruled out;
- post-freeze historian outputs entering a learner-visible snapshot.
Classify findings as informational, possible, probable, or confirmed. A confirmed or policy-defined probable leak blocks clean-room promotion. Distinguish contamination from legitimate independent convergence when evidence permits, but do not assume convergence merely because the result is plausible.
Return only JSON matching schemas/contamination-report.schema.json.