How we test the agent

The agent is tested like any other release-critical component: fixed scenarios with expected verdicts, adversarial inputs, and grounding checks on every answer.