# Interventions

Response types are semantic capabilities, not an ordinal punishment score.

| Response | Effect | Typical use |
|---|---|---|
| `observe` | Record only | aligned or insufficient evidence |
| `ask` | Request clarification/evidence | ambiguous, low confidence, human-related reading |
| `suggest` | Offer an optional adjustment | small drift |
| `feedback` | Explain evidence, target, confidence, and choices | Source-facing or governed Realm feedback |
| `modulate` | Add a temporary bounded context directive | Ally/Actor, reversible, sufficient confidence |
| `correct` | Alter permitted AI/Actor behavior toward target | intervention zone under explicit standing policy |
| `mediate` | Facilitate participants and negotiate temporary targets | interaction mismatch |
| `repair` | Run a structured accountable recovery process | repeated/significant harm or failed intervention |
| `escalate` | Route a minimum-necessary case to accountable Authority | severe drift, unresolved conflict, or required human/governance decision |
| `constrain` | Deny, pause, narrow, or quarantine a prohibited Action | deterministic or high-confidence boundary hit |

`modulate` and `correct` are not permitted for a human Source. On Avatars they may alter generated representation, never the Source. Realm-level automatic action is limited to reversible process controls authorized by governance.

Every intervention records subject, scope, assessment, policy rule, Authority, reversibility, intended movement, prohibited effects, expiry, status, outcome, and appeal/release path. A Sentinel uses the minimum effective intervention and stops when the exit condition is met.

Profiles may map zones to a subset of response types. Policies may narrow that set. Neither may add a response beyond the Sentinel's Permission.