How should I score the request a model writes, not just its decision?

**Question: The decision is right, but is the request any good?**

A correct decision can come with a poor request, so grade the request on its own. Useful checks:

- **Item match.** Each requested item maps to a gold gap. Count extras and omissions. - **Specificity.** The request names the document type and the problem, such as a statement for March and April, not just more statements. - **Traceability.** Each item cites or implies the requirement it satisfies. - **Recipient.** The request goes to the right role, for instance the broker rather than the applicant when the broker holds the file. - **Single round.** All known gaps are included, so the customer is not asked twice. - **Tone and length.** Brief, polite and free of internal jargon. - **No leakage.** It does not disclose another party's details or internal notes.

The first five can be scored by script against the gold gap list. Tone and leakage usually need a rubric-based grader or human review. If you use a model as a grader, check its agreement with human reviewers on a subset first.

A task pack can include a reply simulator, so you can also test whether the request actually got the missing items back. Ask in a pilot whether reply behaviour, such as partial or ignored replies, is part of the scope.