Agent Action Risk Checker

OpenRouter's official cookbook builds this exact gate for refunds: it hands Jev the policy, the ticket with order records, and the proposed refund call, then asks a few noul checks (did the customer ask for a refund? does the order match? does policy cover this amount?). Plain Python keeps control — if every check is ≥0.9 the call runs, ≤0.1 it's blocked, anything in between goes to a human. In their runs a decision cost under $0.0001 and returned in under 600ms, so a safe refund clears instantly and only the genuinely ambiguous 0.41 case reaches a person.