JEV-as-a-Judge is a decision-only LLM evaluation model designed as a cheap first-pass filter, tested against sixteen generative and reward-model judges with blinded human adjudication. It performs within three percentage points of a state-of-the-art LLM judge on ordinary preference and factuality checks at just 0.36% of that judge’s cost, with larger gaps appearing on tasks requiring derivation-checking or resisting elaborately-written wrong answers. A cascade that accepts JEV’s confident verdicts and escalates only uncertain ones retains 99% of the stronger judge’s accuracy at substantially lower cost.
