What success criteria and review questions should I use on a set-up task pack?

**What counts as success on a task?** The episode ends in the expected state, or in one of the listed acceptable variants, with no forbidden writes. A text reply claiming success does not count by itself.

**Can I give partial credit?** Yes. Grade required records, correct links, rule compliance and unintended writes as separate components, then combine them as suits your training or reporting needs.

**How do I tell whether tasks are too easy?** Run a naive baseline that creates records straight from the approval without checking anything. If it passes most tasks, the exceptions are not biting. Tasks where the naive path leads to a downstream consequence, such as a wrong invoice or a contested claim, are harder to pass by accident.

**What if my agent finds a valid path the grader rejects?** Treat it as a grader defect, not an agent failure. Ask how acceptable variants are expressed and whether you can add them. Keep a record of such cases for the next release.

**How do I check grounding and disclosure risk?** Ask how the structure and exceptions were informed, then review samples for names, identifiers or combinations that look like they might belong to a real party. Check that rare attributes are not unusually specific. This is a review you carry out; generation does not replace it.

**Which failure modes should I report separately?** Missing actions, wrong values, duplicate creation, unintended writes and ignored conditions.