What success criteria should I use for an agent that judges document sufficiency?

**What should I measure?**

Use several scores instead of one.

- **Outcome:** did the agent choose `sufficient`, `insufficient` or `escalate` correctly? - **Gap match:** of the labelled gaps, how many did it find, and how many gaps did it report that are not real? - **Request quality:** does the message name every open gap, state the period or form needed, and leave out items already satisfied? - **Reply handling:** after a partial reply, does it close the right gaps and keep the others open? - **Escalation:** does it hand over cases the policy says need a person, with a usable note?

**Which error matters more?**

That depends on your process. A false proceed may cost rework or an unsound decision. A false wait costs delay and customer effort. Report both separately and weight them using your own costs, not a default.

**Do I need a pass mark?**

Set one from your own risk tolerance and test it on a held-out set. We do not publish thresholds here, and a number from one process may not transfer to another.

**What else should I ask?**

Whether the labelled answers follow the stated policy version, and whether borderline cases carry a rationale you can audit.