First answer
wins too often.
In pairwise comparison, judges favour the response presented first. Presenting both orders and averaging removes most of the effect for the cost of a second pass.
The grader
needs grading.
A judge model makes evaluation affordable at scale. It also inherits every preference its own training gave it, which is why a judge has to be measured before it is trusted.
READ THE DEFINITION ↓BIAS · VALIDATION · FINE-TUNING
01 / DEFINITION
The short answer first.What is LLM-as-a-judge?
LLM-as-a-judge is the use of a language model to score another model's output against stated criteria. It replaces human grading for open-ended tasks where no reference answer exists, and it is only as trustworthy as the agreement between its judgements and the human judgements it stands in for.
The technique is now standard because the alternative does not scale. Grading a few thousand open-ended responses by hand is a week of expert time per evaluation run, which makes iteration impossible. A judge model turns that into minutes, at the cost of introducing a second model whose own behaviour must be evaluated.
That second evaluation is the part most often skipped. A judge deployed without a measured agreement rate against human labels is an unvalidated instrument, and every downstream number inherits its errors silently.
02 / KNOWN BIASES
What judge models get wrong, reliably.First answer
Longer reads
as more thorough.
Length correlates with perceived quality independently of correctness. Controlling for length, or scoring against an explicit rubric rather than overall preference, reduces it.
A model prefers
its own style.
Judges tend to score output that resembles their own generation more highly. Using the same family for generation and judging quietly inflates results.
A fourth, less discussed failure is criteria collapse. Asked to score several dimensions at once, a judge tends to produce highly correlated scores across all of them: a response judged accurate is also judged well structured and appropriately detailed. Scoring one dimension per call costs more and produces genuinely independent signals.
General-purpose judges are also weakest exactly where expert judgement matters most. On material that requires domain knowledge — a defective clause, an unsupported citation, a procedurally wrong step — a generalist judge scores fluency confidently and substance poorly.
03 / VALIDATING A JUDGE
The measurement that has to come first.Collect human labels.
A few hundred items, labelled by the domain experts whose judgement the judge is meant to reproduce, blind to the model's output.
Measure human agreement.
Experts disagree. The agreement rate between them is the ceiling for the judge, and a judge that matches it has done all that can be asked.
Measure the judge against the labels.
Report agreement, ranking quality and calibration, not just accuracy. A judge that cannot order bad work above good work cannot gate anything.
Re-measure on a schedule.
Judges drift when the judged system changes its output style, and a hosted judge can change underneath you without notice.
Step two is regularly skipped and it changes the interpretation of everything after it. If two qualified reviewers agree on 78% of items, a judge agreeing with one reviewer 76% of the time is performing at roughly human level, not failing. Without that denominator, the judge's score is uninterpretable.
04 / WHEN A JUDGE NEEDS FINE-TUNING
First-party evidence.A specialist judge
is a different instrument.
MKC2 is a supervision model Privilege AI fine-tuned from Qwen3-8B specifically to judge legal drafting and the actions of legal AI agents. On 4,403 held-out items it ranks defective work above sound work at 0.9623, against 0.7281 for the stock base model it was fine-tuned from, scored in the same run on the same material. Its accuracy is 93.17% against 83.09%, and its calibration error is 0.0610 against 0.1363.
The gap between 0.7281 and 0.9623 is the practical argument for a specialist judge. A stock model with a ranking quality near 0.73 produces a score that orders defective and sound work only loosely, which means no threshold on that score supports a reliable routing rule. The fine-tuned judge makes thresholds possible, which is what makes the rest of the review workflow buildable.
The same record states its limits: these are benchmark results on held-out data, no attorney dispositions have been recorded, and MKC2 runs in review-only mode. Full detail is on the MKC2 page.
05 / QUESTIONS
Asked before a judge model is trusted with a gate.Can a model judge its own output?
+
It can, and it will be generous. Self-evaluation is useful as a cheap filter for obvious errors and unreliable as a quality measure, because the same blind spots that produced the error also affect the review. Use a different model family, and validate it against human labels.
Does the judge need to be larger than the model it judges?
+
No. Judging against explicit criteria is a narrower task than open-ended generation, and a smaller fine-tuned model can outperform a much larger general one on it. MKC2 is an 8B-class fine-tune used to review work from other systems.
How much human labelling does validation need?
+
A few hundred carefully labelled items is usually enough to establish agreement and detect gross bias. Thousands are needed to report per-check agreement or to detect small differences between candidate judges. Budget for re-labelling, because criteria drift over time.
Should a judge model give a score or a verdict?
+
A score, plus a reason, with the threshold applied outside the model. Keeping the threshold in configuration rather than in the prompt means it can be tuned, audited and changed per check without touching the model, and it makes calibration measurable. See AI model calibration.
Can a judge model run on-premise?
+
Yes, and for sensitive evaluation sets it should, since the judge sees every input and every output. MKC2 runs entirely offline. See private AI inference.
A judge
worth trusting.
Privilege AI develops and validates specialised supervision models, and reports the checks that do not reach the bar.