What checks should I run before trusting an approval-authority task pack?

Run these checks on a sample before you build an evaluation on top of it, and keep the scripts so you can repeat them on later releases.

- **Single applicable authority:** each request resolves to one authority record at its decision date, or the ambiguity is documented. - **Independent arithmetic:** recompute limits, aggregates and delegation windows and compare them with gold outcomes. - **Boundary spread:** amounts fall below, at, just above and far above limits. - **Trivial baselines:** an always-approve and an always-escalate policy both score poorly. - **Template variety:** wording, evidence and routing history vary enough that a model cannot match on surface form. - **Policy agreement:** written rules and structured limits state the same thing. - **Practitioner read:** someone who has worked in the domain skims a sample of tasks for plausibility. - **Disclosure review:** scan for anything that looks like it identifies a real organisation or person.

Treat a failure on any item as a question for the supplier, not necessarily a reason to reject the pack. Some inconsistencies are deliberate edge cases, and the pack should say which.

Because the data is generated, passing these checks shows internal quality, not that the pack matches your production environment. Decide separately how well the domain and role structures fit your target use.