Prompt 19 — Clean-room contamination and solution-leakage auditor

Inputs

AUDIT_TARGETS: {{AUDIT_TARGETS_JSON}}
LEARNER_ALLOWLIST: {{LEARNER_ALLOWLIST_JSON}}
FORBIDDEN_REFERENCE_INDEX: {{FORBIDDEN_REFERENCE_INDEX_JSON}}
HIDDEN_TARGET_MAP: {{HIDDEN_TARGET_MAP_JSON}}
MODEL_AND_CORPUS_PROVENANCE: {{MODEL_AND_CORPUS_PROVENANCE_JSON}}
AUDIT_POLICY: {{AUDIT_POLICY_JSON}}

Task

Assess whether a learner-visible object, prompt, episode, or candidate may have obtained hidden modern knowledge.

Detection dimensions

Check for:

• direct forbidden object IDs or source IDs;
• modern concept and theorem aliases;
• canonical notation that was not available in the seed;
• distinctive textbook phrasing or memorized solution structure;
• historical dates, attributions, or labels that reveal the target;
• metadata, filenames, comments, evaluator messages, or tool outputs that leak a solution;
• dependency edges to inaccessible partitions;
• retrieval or model-call traces outside the allowlist;
• suspiciously exact correspondence with a hidden target despite absent developmental precursors;
• cross-session contamination from a previously revealed target.

Important limitations

Lexical similarity alone is not proof of contamination, and absence of a detected match is not proof of independence. Distinguish:

• confirmed_leak;
• probable_leak;
• possible_leak;
• structural_similarity_only;
• no_detected_leak;
• audit_inconclusive.

Output

For each target return:

• evidence and exact matched fields;
• accessible-source trace;
• developmental plausibility assessment;
• severity;
• required remediation;
• eligibility for clean-room evaluation;
• residual uncertainty.

A confirmed_leak or unresolved probable_leak blocks clean-room promotion. Return JSON matching schemas/contamination-report.schema.json.