# Repair

Repair is a structured response after significant drift, failed intervention, interaction harm, boundary event, or successful appeal. It is not a punishment level.

## Protocol

1. Contain ongoing effect with the least-forceful reversible action.
2. Separate deterministic facts, supported inferences, confidence, and disputed evidence.
3. Identify affected Sources/Realms, commitments, Resources, Records, and responsible Authority.
4. Provide correction, representation/support, privacy, and appeal choices; participation in direct reconciliation is optional.
5. Create authorized steps: reverse/stop effect, correct output/Record, restore access/resource, acknowledge, mediate, change process, rehearse, update AI profile/Presence, or compensate through a separate Action.
6. Assign owner, deadline, Permission, completion evidence, failure route, and data handling to every step.
7. Define independent release/re-entry criteria. The Sentinel or constrained entity cannot release itself.
8. Reassess outcome, calibrate the measurement/policy if it erred, and close, extend, or escalate.

Repair never silently deletes the original Record, requires forgiveness/confession, expands surveillance, or uses a Coherence target to decide human worth. Low-confidence responsibility claims remain unresolved and cannot support sanction.