Safety preflight before an agent runs a DESTRUCTIVE, costly or IRREVERSIBLE action (delete, deploy, transfer funds, publish, overwrite): risk assessment + reversibility check + timing + CLEAR/LIMIT/REVIEW/BLOCK verdict + a cryptographically SIGNED ClearancePacket (decision, approved amount, policy version, evidence hash, expiry, rationale) the agent keeps as proof of authorization. Agent action guardrail / clearance / human-in-the-loop gate; stops a mistake costlier than the call price.