Research note · October 26, 2026

Generative Judges Before Jev, and a Meta-Harness for a 4B Judge

21 generative judges · 3 rubric benchmarks · compared with the decision model Jev 1.13 · an automated harness search lifts a 4B judge on HealthBench meta-evaluation from 61.11 to 72.58 test MF1