How do we check that a complaint benchmark is not revealing its answers?

**Where can answer leakage appear?**

A file name may identify the intended escalation route. A case identifier may encode the outcome. Documents may always place decisive evidence last. Internal metadata may include a final disposition that the acting role could not yet know.

**What should the buyer inspect?**

Ask for a field-level distinction between model-visible inputs, environment-only state and evaluator-only labels. Review exported prompts and tool responses, because a clean source dataset can still leak answers when an integration assembles the task.

**How can we test for shortcuts?**

Run a restricted-input baseline that receives identifiers and structural metadata without complaint evidence. If it predicts outcomes, investigate which artefacts carry the signal. Shuffle non-semantic document order where order is not itself part of the task.

**What about train and evaluation splits?**

Related variants should be grouped before splitting. Otherwise, a model may encounter the same underlying case with lightly changed wording during training. MirrorFirm could supply generated case-family identifiers for split management while excluding those identifiers from the model's observation. This supports an inspection of memorisation risk without turning family membership into another answer cue.