audit-data-contamination

Purpose

Audit train/test leakage, memorization artifacts, temporal leakage, and disclosure pathways that can invalidate evaluation.

Input contract

required: [dataset_partitions, training_sources, evaluation_records]
optional: [release_history, deduplication_keys, access_logs]
constraints: [each contamination claim requires a matched record, pathway, and comparison basis]

Procedure

1. Define partition, temporal, entity, and disclosure boundaries.
2. Match evaluation items against training and public-source records using declared keys.
3. Test suspected leakage pathways and compare affected with unaffected cases.
4. Classify confirmed, plausible, and unassessed contamination with impact notes.

Output contract

produces: [contamination_matches, pathway_map, affected_case_set, impact_assessment, audit_uncertainties]
delta_fields: [findings, evidence_updates, uncertainties, decisions, open_questions]

Quality gates

• Matching keys and time boundaries are recorded.
• False-positive matches are sampled and documented.

Failure and counterexamples

Do not call lexical overlap contamination without a disclosure pathway, and do not infer clean separation from missing metadata.

Provenance map

• resolved: knowledge-acquisition-contamination-audit