How can a lab test whether a model learned the rule rather than the wording?

**What makes a useful pair of evaluation cases?**

Keep the business structure stable while changing one fact that should alter the decision. A task can change whether the requester is also the required independent approver, while leaving the amount, evidence and proposed correction unchanged. The expected outcome changes because the independence condition changes.

**Should every variation change the answer?**

No. Other pairs can alter irrelevant wording, record identifiers or document order while preserving the valid outcome. Together, the pairs test both appropriate sensitivity and resistance to superficial variation.

**How should training and evaluation material be separated?**

Ask for case-family identifiers and generation lineage suitable for grouping related variants. A split by individual rows can place closely related cases on both sides, making performance harder to interpret.

**What should the buyer review?**

Inspect the stated causal difference for each decision-changing pair and confirm that no unintended field reveals the expected answer. A model's result is more informative when the pack explains exactly which business fact should drive the change in behaviour.