How can a buyer tell whether an agent uses relevant evidence rather than surface wording?

**What is a useful inspection method?**

Ask for paired tasks that hold most request details constant while changing one policy-relevant fact. In logistics, a delivery enquiry can require different destinations depending on whether customs clearance is pending or complete. The pair should differ in that fact and its necessary consequences.

**What should stay stable?**

Tone, message length and incidental wording can stay similar. A complementary pair can vary those surface features while keeping the decisive operational facts unchanged. Together, the pairs test whether the agent reacts to the right distinction.

**What would indicate a problem?**

An agent that sends every message containing “urgent” to escalation may ignore the actual escalation rule. An agent that chooses the same queue after the clearance state changes may be relying on a memorised phrase rather than the environment evidence.

**What should a buyer inspect in the generated pairs?**

Check that changing one fact has not introduced contradictory records or accidentally revealed the answer through identifiers. MirrorFirm could include pair identifiers and expected invariances so builders can report sensitivity to relevant facts separately from sensitivity to wording.