Skip to main content
The Review Copilot works with the full context: the trace, the rubric and its examples, past reviews of similar cases, the trace’s user journey, and saved memory from earlier reviews. Because it knows the rubric, its help is grounded in the same standard the reviewer is grading against.

​
What to ask it

  • Explain the rubric. Ask how a dimension should be applied, or why an example counts as good or bad, so a new reviewer grades it the way an expert would.
  • Draft a rationale. Have it write a short, clear rationale for a grade or a correction, which the reviewer edits and keeps.
  • Suggest a correction. Ask for an improved version of the model’s answer, then drop it straight into the review as the corrected response. You can point it at one part, like the query, the prompt, a single step, or the final answer.
  • Check the rubric itself. Ask it to flag gaps, ambiguities, or missing dimensions, so you can tighten the standard over time.
Every suggestion is a starting point the reviewer edits and approves, so speed doesn’t cost you judgment.

​
Next steps

Judges & Datasets

Calibrate a judge against this review data