Source + purpose + policy + mode
SOURCE
Authorized text or supported browser file content. The source is never written into a Privacy Receipt.
PURPOSE
One narrow description of what the downstream model or tool must do.
POLICY
A versioned mapping from entity class to transformation action, identified by a canonical policy hash.
MODE
Balanced, strict, maximum, or API-defined custom behavior within policy bounds.
Seven explicit actions
| ACTION | BEHAVIOR | EXAMPLE INTENT |
|---|---|---|
| keep | Preserve an approved value | A non-sensitive product category needed by the task |
| pseudonymize | Replace with a session-bound placeholder | PERSON_01 keeps relationships without a name |
| generalize | Reduce precision | Exact amount becomes a bounded magnitude |
| conceptualize | Replace with a functional concept | Named organisation becomes counterparty |
| remove | Delete the span | Unneeded contact details |
| block | Stop unsafe movement | Private key, password, seed phrase, credential |
| local_only | Keep the operation inside the trusted environment | A class disallowed from external routing |
Deterministic now; contextual ML later
The live detector uses deterministic patterns, validators, and a small gazetteer. Presidio, spaCy, and an ONNX relevance layer are researched extensions, but the production detector does not invoke them today.
Any model extension must be pinned by artifact hash and model card and evaluated by language and entity class. It will assist recognition; it will not replace secret blocking, explicit policy, or output inspection.