Keep sensitive AI content inside the partner boundary
Separate local enforcement and raw content from the allowlisted evidence needed for central oversight.
01
Raw content stays on the partner side
The architecture is designed so raw prompts, responses, retrieved documents, and sensitive tool values remain within the partner environment where enforcement occurs.
Only an explicit allowlist should cross into central oversight.
- Pseudonymous system and component identifiers.
- Policy, decision, and control version references.
- Selected risk signals, timestamps, and integrity-protected evidence references.
02
Treat the boundary as a testable control
The claim remains conditional on deployment configuration, network controls, schema enforcement, and negative telemetry tests in each target environment.