New primitive binds AI agent authorization and audit into one object

A developer proposes a pattern where AI agents' actions are checked against a policy before execution and recorded in a tamper-evident log after. The verdict is synchronous while the record settles asynchronously, preventing drift between policy and proof. This approach aims to prevent unbounded harm from agent actions while maintaining auditability.