Free template

LLM-as-a-judge rubric template

A strict, reviewable template for comparing actual model output with expected behavior and returning a structured verdict.

Updated
July 17, 2026
Reading time
6 minutes

This template is designed for regression testing: one input, one expected output, one actual output, and a binary verdict with a short reason. It avoids vague numeric scoring and keeps the evidence visible to the reviewer.

Adapt the rules to the specific workflow, then calibrate the judge against human-reviewed examples before using it to block releases.

Key takeaways

Judge against an explicit expectation

Do not ask whether an answer is generally good; define what this case requires.

Make formatting requirements binding

A downstream contract can fail even when the prose contains the right idea.

Return one concise reason

The reason should identify the missing, incorrect, or malformed behavior a reviewer can act on.

Judge system instruction

Use this as the stable judge instruction. Keep it unchanged when comparing a baseline with a candidate so the scoring conditions remain comparable.

System instruction
You are an exacting evaluation judge for AI outputs.

You will receive a test input, an expected output, and the
actual output produced by a model. Decide whether the actual
output satisfies the expected output.

Rules:
- Equivalent wording may pass; word-for-word matching is not required.
- Missing or incorrect required substance must fail.
- Follow any format implied by the expected output.
- Do not reward extra content that was not requested.
- Be strict, consistent, and concise.

Return only JSON:
{"passed": true, "reason": "One short sentence."}

Per-case message template

Send the same fields for every evaluated output. Use clear delimiters so content in the test case cannot be confused with judge instructions.

Judge input
Test input:
{{input}}

Expected output:
{{expected_output}}

Actual output:
{{actual_output}}

Optional workflow-specific rules

Add only rules that define a real pass condition. Overloading the rubric with broad qualities such as clarity, usefulness, correctness, tone, safety, and completeness makes disagreement difficult to diagnose.

  • Classification: prose around a required single label is a format failure.
  • RAG: claims outside the supplied context fail unless uncertainty is explicitly allowed.
  • Support: required escalation or next-step language must be present.
  • Extraction: missing required keys or invalid JSON fails.
  • Content: blocked claims fail even when the rest of the answer is strong.

Calibration worksheet

Prepare at least ten human-reviewed outputs: clear passes, clear failures, and borderline cases. Record the human verdict, judge verdict, and disagreement reason. Repair the expectation or rule before blaming the model being evaluated.

Calibration row
{
  "caseId": "support-outage-001",
  "humanVerdict": false,
  "judgeVerdict": true,
  "disagreement": "Judge ignored the required one-label format.",
  "action": "Make the format rule explicit and rerun calibration."
}

Compare judge verdicts across a real release change

Keep the judge route fixed, score the same cases, and inspect every reason behind the pass-rate delta.