Safe content-monitoring evaluation
The evaluation model keeps classification on the device: classify locally, discard benign content, and expose only the minimum test signal needed to inspect the workflow.
This page describes an evaluation architecture. Phosra does not claim that a production content-routing rail, independent transparency authority, accredited classifier program, or customer deployment is operating today.
Responsibility boundary
| Component | Evaluation responsibility |
|---|---|
| Your app | Obtains any required authorization and presents the user experience. |
| Your classifier | Classifies locally and minimizes retained content. |
| Phosra SDK/sandbox | Exercises policy, envelope, and sample-artifact shapes with synthetic data. |
Phosra is not the classifier and the sandbox is not a content pipe. Do not send real child messages, images, credentials, or family identifiers into the public sandbox.
Suggested local design
- Run classification where plaintext already exists.
- Use a closed, documented harm taxonomy and reject unknown classes.
- Discard benign content immediately according to your reviewed retention policy.
- If your pilot needs an excerpt, minimize and encrypt it for the authorized recipient before it leaves the device.
- Log content-free counts and failure states without logging messages or secrets.
- Test consent withdrawal, key rotation, deletion, and incident-response paths.
These steps are architectural guidance, not proof that deletion occurred, that a classifier found all harm, or that the design satisfies a law.
Evidence for a design-partner pilot
Bring:
- a data-flow diagram;
- a threat model and abuse cases;
- the closed taxonomy and unknown-class behavior;
- retention and deletion rules;
- consent, revocation, and recipient-key handling;
- local test results; and
- the exact roadmap dependencies in your design.
The available harness can help exercise selected failure modes with synthetic data. A passing local or sandbox run is pilot evidence only; it does not confer accreditation, certification, or a production Trust List entry.
Roadmap dependencies
Treat all of the following as roadmap unless a later dated release explicitly says otherwise:
- hosted routing between verified participants;
- a broader production receipt rail;
- independent transparency anchoring;
- production classifier adapters; and
- independent accreditation or certification pathways.
Next step
Apply for the design-partner beta or email [email protected] with a bounded workflow and pilot success criteria.