ML Feature Pipeline Leakage Auditor

Audit an ML feature pipeline for point-in-time correctness, label leakage, and training-serving skew. Prediction target and prediction timestamp: [definition] Entity keys: [customer_id, device_id, etc.] Labels and observation window: [details] Feature definitions and source timestamps: [event time, ingestion time, update time] Offline and online stores: [systems] Training-set query or pipeline: [Paste SQL/code] Online serving path: [Describe or paste code] Backfill and late-data behavior: [details] Produce: 1. Timeline-based definition of what information is legally available at prediction time 2. Feature-by-feature leakage analysis 3. Point-in-time join strategy with valid_from/valid_to or as-of semantics 4. Handling for late events, source corrections, and mutable dimensions 5. Offline/online transformation parity assessment 6. Detection of label-derived, post-outcome, and proxy leakage 7. Tests using synthetic adversarial timelines 8. Freshness and skew monitoring for serving 9. Reproducibility requirements for datasets, code, and feature versions 10. Corrected SQL/pseudocode and a safe backfill strategy Flag any feature whose timestamp semantics are ambiguous. Do not treat ingestion time as event time without evidence.