How can we tell whether an agent learned the handover rule rather than a routing shortcut?

**What would a routing shortcut look like?**

An agent might send every case containing a particular phrase to a specialist team, even when the policy depends on a separate status field. Strong results on repetitive examples would not establish that it understands the rule.

**How can generated cases expose that?**

Ask for paired episodes in which the decisive condition changes while the surrounding narrative stays stable. In healthcare administration, a referral can require a missing administrative authorisation in one episode and contain a valid authorisation in the paired episode. The permitted scheduling action should change accordingly.

**Should every detail remain identical?**

No. Separate tests can vary irrelevant features, such as record identifiers or note order, while preserving the correct decision. Together, these tests check sensitivity to governing conditions and stability under harmless variation.

**What should a buyer inspect?**

Review the expected behavioural difference for each pair and confirm that it follows from the policy and state. Keep related variants together when allocating training and held-out evaluation material, so a near-identical case does not accidentally reveal the answer. Pair design should test a stated mechanism rather than merely increase case volume.