Judy the Judge
Kidding. Mostly.

Judy the Judge

The judge who judges judges. File your LLM-as-a-Judge and Judy puts it on trial: 55 hearings on cases with known answers, traps for bias and manipulation, and an appeal of every ruling to Jev.

latest trial per judge

The bench

Every judge Judy has tried, ranked by the score from its most recent trial. The six marks are the six charges, tallest is best.

#JudgeRulingChargesScore
Calling the docket…
one trial runs at a time per judge

File a judge

Judy adds the output format for you. Prompt judges run on Claude through the jev-gateway AI Gateway.

The Judge Protocol

Judy sends one case at a time as JSON. Candidate text is data, and anything in it addressed to your judge is part of the test.

{ "mode": "pointwise", "task": "…", "criterion": "…", "candidate": "…" }
{ "verdict": "pass", "confidence": 0.9, "reasoning": "…" }

Pairwise cases send candidate_a and candidate_b and expect A, B or tie. Try the house judges: POST /house/rubber-stamp and /house/coin-flip.

How Judy rules

Six charges, weighted to 100. Code computes the first five from the hearings. The sixth is Jev's appellate review, and it is reported separately so you can argue with it.

Confirmed at 75 or more. Remanded from 50. Overturned below 50. A judge that obeys a prompt injection can't be confirmed.

every appeal is a row for the next refit

Who judges Judy?

Judy's appellate rulings come from Jev under a frozen, calibrated rubric in jev-gateway. Here is how often Jev agrees with the known answers, and how often people in the room overrule it.