Human review

Some judgments should be made by people: new products where no rubric exists yet, high-stakes domains, and calibrating model graders. Human review works best when it is structured like any other grader.

Set up a review round

1. Export results from a run (or sample production logs) to a spreadsheet or review tool: input, output, and the ideal answer if any.
2. Give reviewers the same rubric a model grader would get, with anchored score levels.
3. Have two people grade each item independently for at least the first 50 items.
4. Measure agreement. Below 80% on pass/fail means the rubric is ambiguous; fix it before grading more.

Use reviews to calibrate model graders

Run the model grader on the items people graded. Where the two disagree, read both judgments. Adjust the rubric until the model grader agrees with people as often as people agree with each other; then let the model grader take over the volume and keep a small human sample (5%) as a check.

Reviewing in production

For live applications, sample a fixed share of conversations each week for review, stratified by tag or outcome, and add every confirmed failure to the evaluation dataset. Production review is what keeps the dataset honest as your users change.

Reviewer guidance

• Show outputs from different variants in random order and without labels.
• Ask for a one-line reason with each failing grade; reasons are what you fix.
• Rotate reviewers to avoid one person’s taste becoming the standard.