Which failure modes should a sufficiency evaluation set expose?

**What failure modes should an evaluation set expose?**

**Premature proceeding.** The agent accepts a file with a missing or invalid item. Test with cases where one requirement is unmet and everything else looks polished.

**Over-requesting.** The agent asks for items already supplied or not required. Test with fully satisfied files and files that use permitted substitutes.

**Presence-only checking.** The agent confirms a document exists but not that it is current, signed or for the right party.

**Wrong requirement set.** The agent applies rules for a different product or amount band. Test with cases near a threshold.

**Vague requests.** The message says 'more information needed' without naming the item or the reason. Grade request text against the gap list.

**Repeat requests.** After a reply, the agent asks again for what was just supplied. This needs multi-turn cases.

**Escalation errors.** The agent escalates routine gaps, or handles alone a case that policy sends to a person.

**Which evaluation questions help?** Does the cited evidence actually support each requirement marked as met? Does the request list equal the unmet set? Does behaviour change when only one policy parameter changes? The last can be tested with paired cases that differ by a single threshold, where the correct decision flips and a rule-following agent should flip with it.