The plain-English version
A concrete example of an output and its assigned rubric outcome that helps reviewers interpret a criterion consistently.
Why it matters
It makes a vague quality label easier to apply.
Let’s make it click
A citation rubric includes a supported answer, an unsupported added claim, and an appropriate insufficient-evidence response with separate expected results.
Vibe Bro’s heads-up
Anchors can encode a reviewer mistake or cover too little variation; check disagreements and revise them deliberately.
You might also hear
- graded examples
- score anchors
- rubric calibration
The docs behind it
Research sources reviewed .
Editorial review · October 28, 2026
Supported: OpenAI supports clear criteria and checking model-judge agreement against human labels. The anchor examples are original calibration guidance, not a universal scoring scale.