Refusal Boundary Test Agent

An assistant that refuses harmless questions frustrates users, and one that accepts risky ones creates liability. This agent builds two sets of cases: clearly allowed requests that sit close to a boundary, such as a nurse asking about medication doses, and clearly disallowed requests, such as asking for another person's data. It runs both sets against the prompt, classifies each outcome as correct, over-refusal or under-refusal, and proposes wording changes to the instructions. After each change it retests both sets, because tightening one side often breaks the other. It repeats until both error rates are within limits. The engineer approves the final wording. Edge case: a request is allowed for staff but not for customers, so the agent tests it under both roles.