Check what can be checked exactly
HTTP failures, response limits, prohibited tool calls, invalid endpoint behavior, and explicit state assertions should be deterministic. These checks are fast, repeatable, and easy to audit.
Use semantic judgment carefully
Natural-language behavior sometimes needs semantic evaluation. It should compare the expected behavior with the recorded response and explain the judgment, rather than replace deterministic policy checks.
Keep evidence bounded
Store only the response and structured state required to reproduce a finding. Bound request and response sizes, redact credentials, and keep raw payloads out of public reports.