- In short
- Designing and calibrating an LLM-as-judge rubric means building a rigorous judge prompt with a detailed rubric and constrained verdicts, a small fixed set of labels rather than free-form scores, grading with a different model than the one being evaluated to avoid self-preference bias, and calibrating the judge against human-labelled examples before trusting its verdicts. An uncalibrated judge produces confident-looking scores that may not reflect quality at all, which is worse than having no automated grade because the scores appear trustworthy.
A judge is a system that can be wrong
When you climb the grading ladder to LLM-as-judge, you gain the ability to score interpretive behaviours at scale, and you take on a new liability: the judge is itself a model, and a model can grade badly. A judge that looks authoritative while producing scores unrelated to real quality is more dangerous than no automated grade at all, because the team acts on numbers it should not trust. The apply-level skill here is building the judge so its verdicts are actually reliable, through rubric design, cross-model grading, and, above all, calibration against human labels.
- Calibrated LLM-as-judge
- A model-based eval built to produce trustworthy verdicts through three disciplines: a detailed rubric with constrained verdicts (a small fixed label set rather than free-form scores), grading with a different model than the one under evaluation to avoid self-preference bias, and calibration against human-labelled examples to confirm agreement is high enough to rely on before the judge scores unlabelled data.
Constrained verdicts and a detailed rubric
A rigorous judge prompt does not ask the model for a free-form score out of ten. It supplies a detailed rubric, spelling out exactly what each quality level means, and constrains the verdict to a small fixed set of labels, for example acceptable / needs-revision / unacceptable, rather than an open numeric range. Constrained verdicts make the judge far more consistent, because there are fewer ways to interpret the task, and they make the results easy to aggregate and audit.
It also helps to require the judge to produce its reasoning alongside the label. A judge is inconsistent precisely on borderline cases, and without visible reasoning that inconsistency is invisible, you see a confident label and no way to tell it was a coin-flip. Reasoning-plus-label surfaces the shaky verdicts so you can catch them.
Grade with a different model
The second discipline is cross-model grading: the judge should be a different model from the one whose outputs are being evaluated. Using a model to grade its own output introduces self-preference bias, a tendency to rate its own style and choices more favourably, which inflates the scores in a way that has nothing to do with actual quality. Swapping in a different model as the grader removes that specific bias and makes the verdicts a cleaner signal.
This is cheap to arrange and easy to forget, which is exactly why it is a favourite exam trap. The self-preference distortion does not announce itself; the scores just come back a little too kind.
Calibrate before you trust
The third discipline, and the one teams most often skip, is calibration. Before you rely on the judge's verdicts on unlabelled data, run it against a set of outputs that humans have already labelled, and measure how often the judge agrees with the humans. If agreement is high enough, the judge has earned trust for that task. If it is not, you revise the rubric, adjust the verdicts, and re-check, before the judge grades anything for real.
Skipping calibration is worse than having no automated grade, because an uncalibrated judge still produces confident-looking scores, and those scores get believed. No grade at all keeps the team honest about its uncertainty; an uncalibrated grade manufactures a false certainty that the outputs are good. Calibration is what converts the judge from a plausible-sounding oracle into a validated instrument.
What the exam trips candidates on
The first trap is using the same model to judge its own outputs, introducing self-preference bias. A scenario will have a team grade a model's outputs with that same model; the credited answer swaps in a different grader and names the bias being avoided.
The second trap is deploying a judge rubric into production without calibrating it against a human-labelled sample. A scenario will present a judge producing confident scores that the team trusts at face value; the correct reading insists on calibration first, and points out that an uncalibrated judge is worse than no automated grade because its confidence is unearned. Both traps reward treating the judge as a system that must be validated, not a truth machine.
Worked example
A team building on one Claude model wires up an LLM judge, using the same model, to score its outputs for reasoning quality on a free-form 1-to-10 scale, and starts making ship decisions from the average score. What is wrong, and how should the judge be rebuilt?
Three of the disciplines are violated, and each undermines the score in a different way.
First, the judge is the same model that produced the outputs, so its verdicts carry self-preference bias, it rates its own style favourably, and the average is inflated by an amount unrelated to real quality. Rebuild by grading with a different model, removing that bias from the signal.
Second, the verdict is a free-form 1-to-10 score, which invites inconsistency, the model's sense of what a 7 versus an 8 means drifts across items. Replace it with a detailed rubric and a small fixed set of constrained verdicts, and require the judge to give its reasoning alongside the label so borderline calls are visible rather than hidden inside a confident number.
Third, and most seriously, the judge was never calibrated. The team is making ship decisions from scores that have never been checked against human judgment, so there is no evidence the judge's "good" corresponds to actually-good output. Before trusting it, run the judge against a human-labelled sample and confirm agreement is high enough; if not, revise the rubric and re-check. Until that passes, the confident averages are worse than no grade, because they are being believed.
The rebuilt judge uses a different grader model, a rubric with constrained verdicts and required reasoning, and a calibration step that gates trust. Only then is the average score a basis for a decision.
Common misreadings to avoid
Misconception
It is fine to grade a model's outputs with the same model, since it understands the task best.
What's actually true
Misconception
If an LLM judge returns confident, consistent-looking scores, its verdicts can be trusted.
What's actually true
How this shows up on the exam
Domain 4 questions on this knowledge point present an LLM judge that is either grading its own model's output or being trusted without calibration, and ask what is wrong. The reliable fixes are cross-model grading, a rubric with constrained verdicts, and calibration against human labels before any verdict is trusted.
This is what makes the second rung of the grading ladder safe to climb to, and it depends on knowing when a behaviour needs a model-based rather than code-based eval in the first place. The calibrated judge then plugs into the five-stage eval workflow at the judge-scoring stage, powers multi-turn eval design, and connects to production discernment, where human judgment feeds back to keep automated grading honest.
A team wants to trust an LLM judge's reasoning-quality scores for release decisions. Which set of practices best justifies that trust?
People also ask
How do you calibrate an LLM judge?
Why should the judge model differ from the model being evaluated?
What are constrained verdicts in an eval rubric?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.