How can a lab tell whether a model is using evidence or dataset shortcuts?

**What shortcuts should a buyer look for?**

A generated pack can accidentally reveal the expected result through filenames, identifier prefixes, document length or the order of attachments. A model may then appear competent without interpreting the evidence.

**How can those shortcuts be tested?**

Use paired cases with controlled differences. In one pair, change a decisive fact, such as whether an approval falls within the approver's delegated authority. The expected conclusion should change. In another pair, change formatting or record order while preserving the business facts. The expected conclusion should remain stable.

**What should be kept outside model inputs?**

Scoring labels, adjudicator notes and generation metadata that disclose the answer should remain separate from the task view. A clear partition matters even when the records themselves are generated.

A buyer can ask for an inspection of outcome correlations in non-business fields and for documented counterfactual transformations. The aim is to test a specific dependency: the model should respond to changes in evidence meaning while resisting irrelevant presentation changes. Unexpected behaviour then becomes a useful diagnostic rather than an unexplained aggregate score.