Rubrics

Your team’s written standards for a correct answer, their versions, and who used them.

A rubric is named criteria, weights and a scale — written down, owned, and versioned. This screen is where most people draft one before moving to the API for later edits.

Rubrics with their criteria, versions and usage.

What you see

Region	What it shows
List	Every rubric you can read: your own, plus public ones
Criteria	Each criterion’s name, weight and description
Version	The current version number; every edit increments it
Visibility	Private (you and administrators) or public (everyone reads, you still write)
Revisions	The full text of every past version
Usage	Which runs cited which version, and whether as a training reward or a benchmark score

What you can do

• Create a rubric, with a sample criterion to start from.
• Edit one — which creates a new version rather than overwriting the old text.
• Delete one you own.

Editing invalidates the judge’s cached scores for this rubric, because the cache key carries its version. That is deliberate: sharpening a criterion should not leave old scores standing as though nothing changed.

Where the rubric is actually used

One object, three places — which is the whole reason it is an object rather than a prompt pasted into two configs:

Used as	Configured in
A training reward	An environment verifier with kind: rubric
A benchmark	A benchmark pack with runner: judge
A safety evaluation	A pack with runner: safety

Because it is one object, the standard a model trains against and the standard it is scored against cannot quietly drift apart. See writing a rubric.