How can labs test whether a model responds to the right review evidence?

A plausible review answer does not show which evidence drove it. Paired cases offer a more targeted check. One pair can preserve the same account history while changing a fact that the supplied policy makes decisive. Another can change presentation details while leaving the required outcome unchanged.

For a retail supplier review, the decisive fact might be whether a required corrective action has supporting completion evidence. Rephrasing the supplier description should not alter the outcome. Replacing an unsupported completion claim with an accepted verification record may need to alter it.

Evaluate the decision and the explanation together. A model that changes its answer appropriately but cites an unrelated fact has not demonstrated reliable reasoning. Equally, a model that produces polished explanations while ignoring the changed evidence has missed the task.

MirrorFirm could generate paired cases with a documented transformation and an expected relationship between outcomes. This supports evaluation of decision boundaries without requiring every case to have a single preferred wording. It also helps labs identify shortcuts based on record position, document length or stylistic cues.