How can we tell whether an agent responds to meaningful evidence rather than superficial cues?

**What does a paired evaluation look like?**

Start with a review case and create a second version that changes one decision-relevant fact. In a professional services account review, that could be whether an unresolved scope dispute has an accepted settlement recorded.

**What should change?**

The expected decision or required follow-up should change only where the supplied policy makes that fact material. The evaluator should document the causal reason for the difference before testing the agent.

**Can the same technique detect superficial shortcuts?**

Yes. Another pair can retain the business facts while changing document order, record identifiers or wording. The substantive outcome should remain stable unless the task explicitly assigns meaning to those changes.

**What makes this difficult?**

A variation can accidentally alter several facts at once. Removing a settlement record might also remove the only evidence identifying the dispute owner. A buyer should inspect the pair's change manifest and confirm that the intended comparison remains valid. Carefully controlled pairs reveal evidence sensitivity more clearly than unrelated easy and difficult cases.