Give the Specialist a Flight Envelope

Create a test plan for this specialist before production use. Cover expected requests, ambiguous requests, missing inputs, stale data, conflicting sources, malformed files, duplicate records, unauthorized users, cross-boundary requests, prompt injection, unsafe tool instructions, tool timeout, partial failure, excessive cost, and recovery.

For each scenario, define the input, authorized tools, expected evidence, expected output or refusal, required audit event, and whether human approval is required. Include tests proving that denied actions cause no external state change.

Separate deterministic automated tests, model-evaluation cases, security tests, integration tests, and supervised flight trials. Define measurable acceptance criteria and unresolved limitations.