Why not treat sufficiency as a complete or incomplete classification task?

**Why not label each file complete or incomplete?** Because a binary label hides the work. Real decisions depend on which requirement set applies, which items satisfy which requirements and what to ask for.

**What else needs to be in the answer?** A useful case has an expected requirement list, a mapping from supplied documents to requirements and a set of acceptable next actions. Two agents can both say 'incomplete', yet one names the right missing item while the other asks for five unnecessary ones.

**Can a file be sufficient with items missing?** Yes. Policies often allow substitutes, such as a utility bill instead of a bank statement for address proof, or waivers when risk is low. A case can encode these so the right answer is 'proceed'.

**Can a complete-looking file be insufficient?** Yes. A document may be present but expired, unsigned, issued for the wrong entity or inconsistent with another document. Presence is not sufficiency.

**What about partial progress?** Some cases allow work to continue on one part while another waits. A supplier onboarding team might create the vendor record but hold payment setup until bank details are verified. Scoring should reflect whether the agent split the case correctly.

**What does this mean for benchmarks?** Score at least three things: the decision (proceed, request or escalate), the identified gaps and the quality of the request. A single accuracy figure over a binary label would blend these and hide where models fail.

Generated task packs can carry these fields so each is scored on its own.