How can customers test whether an agent actually follows the triggering policy?

A useful evaluation can compare cases that share the same operational context but differ in one policy-relevant fact. One case may be inside an allowed waiting period; its paired case may be outside it. The expected action changes while the record structure and communication style remain stable.

MirrorFirm could generate these pairs with an explicit explanation of the decisive difference. A reseller can ask for both action-changing pairs and action-preserving pairs. Changing an irrelevant note should not alter the agent's decision, while changing the triggering condition should.

This exposes a failure mode that ordinary correctness totals can hide: an agent may react to words such as urgent or overdue while ignoring the actual policy conditions. It may also overreact to harmless differences in formatting or narrative tone.

Downstream buyers can inspect whether the agent changes its action for the right reason and cites the field that controls the decision. Pair membership should remain in evaluation metadata rather than being shown as a hint in the task input. Results then describe sensitivity to a specific operational rule instead of implying broad competence from a collection of unrelated cases.