Which checks should I run on an approval sample before using it?

Before training on or evaluating with a sample, run these checks on a small batch and keep the results for the discussion with the supplier.

- Recompute each approval's validity from the role, grant, limit table and policy version, and compare it with the supplied label. - Confirm every identifier in an event resolves to an existing request, role and delegation, with dates in a sensible order. - Count near-limit requests, lapsed delegations, second-approver cases, rejections and escalations to see that the set is not all clean approvals. - Perturb an amount or date on a few tasks and check that the expected outcome changes correctly. - Have an approvals practitioner read a handful of notes and cases for plausibility. - Check that the held-out split differs by policy variant or organisation, not only by row. - Search samples for text that looks too specific to be generated, and record any grounding and disclosure concerns.