Methods · Sep 9, 2026 · 6 min read
The anatomy of a good rubric
A rubric is a quality bar written down so that different people reach the same verdict. Most rubrics fail because they describe a feeling instead of a check.
Make each criterion atomic
One criterion, one thing. 'Accurate and well organised' is two checks pretending to be one. Split them so a reviewer can pass one and fail the other.
Make it binary and observable
Prefer 'States the dose adjustment for reduced kidney function: yes or no' over 'Handles dosing well, rated 1 to 5'. A reviewer should be able to point at the evidence.
Include the failure cases
List what a fail looks like, with examples. The most useful rubric items come from real mistakes the model has made.
Test the rubric itself
Give the same ten outputs to several reviewers. Wherever they disagree, the rubric is unclear. Rewrite, then test again.
Need expert judgment on your model's output?
Try SoReliable