Measure how much work it creates for reviewers

Alongside whether the answer is right, I'd measure how much work each output creates for the people reviewing it. A workflow where most outputs are correct can still take too long to supervise.

For each case, record whether the reviewer accepted it, made a small edit, rebuilt it, went looking for another source, asked the user for more information, or rejected it. Capture review time if you can do it without getting in the way of the task. I'd treat that number with some caution, though: reviewers get faster as they learn the interface, difficult cases cluster together, and timing a small set can look more precise than it is. The pattern of what reviewers had to do is often more useful than an average.

Look at where reviewers' attention goes. If they have to reopen every source and rebuild the answer, the feature isn't saving them much even when its facts are right. If the interface shows the exact supporting passage and makes uncertainty visible, review can be narrower and more reliable.

Reviewer agreement matters here as well. One person might accept fluent wording while another checks every claim. Give reviewers the same instructions and ask them to note why they corrected something. Frequent disagreement can point to an unclear policy, poor presentation of sources, or an interface that invites people to trust the output without showing them anything to check it against.

Count the operational handling too: cases sent to a queue, clarifications requested, tool actions cancelled, and outputs that need support staff to dig through logs. That's all part of what the workflow costs to run.

Sometimes the review numbers point to a design change, and another prompt revision won't help. A structured answer can be easier to check than prose. Showing one source excerpt might save reviewers repeated searches. A narrower task might cut the review work down. I'd want those changes made before customers end up carrying the extra effort.