AI and PHI data flow
The current AI provider abstraction is backed by Anthropic when configured.
The runtime:
- resolves a scoped security context;
- resolves an enabled AI capability;
- enforces capability limits;
- requires minimum-necessary context;
- creates an authorization/context manifest;
- calls the configured provider;
- validates the output;
- records only control metadata and hashes in the context manifest.
Field classification and minimization
Context is not redacted by matching a list of names. Every field offered to a capability is classified before it can be sent:
| Class | Treatment |
|---|---|
| Direct identifier | dropped before provider access |
| Free text | scanned for known identifiers, which are removed |
| Financial | sent only where the capability's policy allows it |
| Coded clinical | permitted as coded data, not free text |
| Non-PHI | permitted |
Fields are classified explicitly rather than by keyword guessing. Patient names, subscriber and member identifiers, medical record numbers, dates of birth, contact details, and the account or guarantor number that identifies the person responsible for the bill are all treated as direct identifiers. Claim number, payer, and service date are classified explicitly instead of being redacted blindly.
Each AI capability carries its own policy, so the minimum-necessary context for one capability is not automatically available to another.
Important boundary
Redaction is not de-identification.
The repository does not establish provider BAA status, retention terms, or external telemetry behavior. Those are contractual/deployment prerequisites.