STEP 01
WHAT WE WRITE DOWN
Codify the rubric
Encode the criteria your reviewers already use. Weighted, evidence-gated, and versioned from the first save.
→MODELS · RUBRIC STUDIO · WRITE THE STANDARD
Write the standard you release against.
A score is only as good as the rubric behind it. Write the criteria your reviewers already use. Version every change. Calibrate the judges before a real release is ever scored against it.
RUBRIC
Versioned
Every criterion, weight, and edit kept on the record.
JUDGES
Calibrated
Models and people scored on the same cases first.
STANDARD
Defended
One rubric your team and an auditor can both read.
WHERE IT STARTS
The rubric is the standard.
Rubric Studio writes the standard. Evaluation Studio runs it against every release. Author once, version every change, and every candidate is scored on the same criteria, in the same order, by judges already calibrated on the same cases.
HOW IT WORKS
One rubric. One standard.
Write the rubric. Calibrate the judges. Score every release the same way.
STEP 02
WHAT WE ALIGN
Calibrate the judges
Model judges and human reviewers score the same calibration set. Disagreement surfaces before a real release ever touches the rubric.
→STEP 03
WHAT WE STAMP
Score every release
Every candidate is graded against the same rubric. Scorecards, judge consensus, and reviewer notes stay with the release review.
WHAT COMES OUT
What your team leaves with.
Tracing tells you what the model did. This records who approved it. Every run leaves the rubric that was used, the judges that scored it, and a verdict the team can defend under an audit.
01
Rubric versions
Every edit kept on the record. Diff one revision against the next without leaving the page.
↳ ARTIFACT
02
Calibration reports
How aligned the model judges and human reviewers are on the same cases — before any real release is scored.
↳ ARTIFACT
03
Scorecards
One read on what passed, what failed, and what every judge said about it.
↳ ARTIFACT
04
Judge consensus records
Where the model judges agreed, where they split, and where a human had to call it.
↳ ARTIFACT
05
Evidence packets
Rubric, judge notes, reviewer overrides, and verdict — sealed and exportable for the August 2026 high-risk provenance deadline.
↳ ARTIFACT
WHERE IT FITS
In the loop, this is where you test.
Test the run against the rubric. Review the hard cases. Recruit the right specialist. Remember the misses. Approve what's right.
01
Test
● YOU ARE HERE
02
Review
03
Recruit
04
Remember
05
Approve
RELATED MODULES
Next to this in Models.
EVALUATION STUDIO
Review it before release.
Rubric Studio writes the standard. Evaluation Studio runs it against every release.
See the page →REGRESSION BANK
Every mistake. Only once.
Every escaped failure becomes a gate the next release cannot cross.
See the page →CONTROL CENTER
The last check before release.
Tests, reviews, and compliance converge on one timeline, one signed approval.
See the page →RUBRIC STUDIO
Write the rubric. Keep the model.
Bring the rubric your team already uses. We version it, calibrate the judges, and attach the proof to every approval. Improve your model on the work your reviewers signed, and keep the weights you tuned.