What quality checks should precede listing a generated correction pack?

Before listing a generated correction pack, inspect both customer-visible inputs and evaluation material, using these checks to identify defects that could survive ordinary file validation.

- Inspect answer leakage: filenames, status labels and explanatory notes should not reveal the intended decision before the customer’s model attempts the task. - Check split contamination: related cases, near-duplicate documents and transformed versions should be identified so customers can separate training from evaluation appropriately. - Reconcile evidence with labels: the expected correction should follow from the supplied records, rather than facts available only in an undisclosed authoring note. - Review difficulty distribution: distinguish arithmetic complexity, evidence ambiguity and authority constraints so a catalogue does not describe all difficult cases as equivalent. - Inspect grounding and disclosure risk: generated output still needs review for recognisable source phrasing, distinctive combinations and unsupported claims about privacy protection. - Check negative cases: include valid reasons to reject a requested correction, ensuring the pack does not teach that every dispute deserves an adjustment. - Confirm documentation consistency: public descriptions, downloadable metadata and machine-readable manifests should agree about included tasks, exclusions and known limitations.