Client-Side Anonymization & Privacy

3-layer PII pipeline (D-074)

1. Layer 1 — SequentialAnonymizer (default) — deterministic placeholders with coreference: EMAIL_1, PHONE_1, PAN_1, AADHAAR_1, UPI_1, ABHA_1, GSTIN_1, SSN_1, SECRET_k, ADDRESS_k, ACCOUNT_k, scoped NAME_k.
2. Layer 2 — SpanDetectorProtocol backends
   • regex (default)
   • presidio via qurator[privacy]
   • openai_privacy_filter via qurator[privacy-filter] — span detector only; BIOES labels map to entity types, then SequentialAnonymizer assigns placeholders. Never use static **[PRIVATE_PERSON]** masks.
3. Hosting — Privacy Filter 1.5B weights run on qurator-scoring / VPC Gateway / Foundry corpus scrub — not on developer laptops.

private_date is opt-in (default off) so policy/business dates are not over-masked. Pre-scan existing ENTITY_N tokens to avoid double-anonymization SDK → Gateway.