Privacy & retention

The data inventory

Data	Where	Encrypted	Retention
ID document images	Vault (cloud → offline archive)	Per-artifact AES-256-GCM	30 days hot, then archive, until erasure
Selfie/challenge frames	Vault	Same	Same
Full AI extraction (incl. document number)	Vault	Same	Same
Minimal identity fields (names, DOB, doc type/country/expiry)	Database	At-rest platform encryption	Until erasure request
Status, decision reasons, hashes	Database	— (no PII)	Indefinite (your evidence trail)
Audit chain	Database	— (PII-light by design)	Indefinite, append-only
Biometric embeddings	Nowhere	—	Computed transiently, never persisted

The retention pipeline

capture ──► vault (encrypted, "hot")          ── 30 days ──►
        archival job: tar bundle + hash manifest ──► offline media
        hot copies deleted; DB records archive ref

• The 30-day hot window exists for disputes and manual review; it is a config value (VAULT_HOT_RETENTION_DAYS), not a magic number.
• Archives are packaging, not re-encryption — blobs are already ciphertext, so the offline media is safe to transport and store.
• Restore is symmetrical (python -m app.archival restore ver_…) for the rare regulator/dispute lookup.

How deletion actually works (and why you can trust it)

Traditional deletion fails at backups: the data you “deleted” lives on in last month’s archive. BAQSHI-KYC inverts this with cryptographic erasure:

1. Every artifact was encrypted with its own key; only the wrapped key exists, in the database.
2. DELETE /v1/verifications/{id}/data destroys the wrapped keys, deletes hot blobs, and nulls the identity fields.
3. Archived ciphertext — wherever it physically is — is now permanent noise. No media retrieval, no archive rewriting.

What remains after erasure: the verification’s status, its decision reasons, and the audit chain (including the erasure event itself). You keep the fact that you verified someone and the evidence of process; you stop holding what you saw.

Data subject requests, mapped to API calls

Request	Action
“What do you hold about me?”	GET /v1/verifications/{id} + (if needed) vault artifact retrieval by the operator
“Delete my data”	DELETE /v1/verifications/{id}/data — done, retroactively, everywhere
“Correct my details”	Reviewer re-examines artifacts and records the correction via the review endpoints (audited)

Third-party processing disclosure

One disclosure belongs in your privacy notice: document images (not selfies, not biometrics) are sent to the AI provider (Anthropic) for text extraction and tamper inspection under their API data-handling terms. If a deployment must eliminate even that, swap the document model for a self-hosted one — the pipeline interface (extract_document) is the single seam.