A confidence cascade over a decision model kept 99% of judge accuracy at 57% of the cost.

arXiv 2609.26550 (CMU, submitted September 22) compares Jev against sixteen generative and reward-model judges under blinded human adjudication. It costs $0.044 per 1,000 judgments at 152ms median and lands within 3 points of the strongest LLM judge on RewardBench-style preference and HaluEval factuality. It trails by 14.5 points on JudgeBench, where the judge has to check a derivation or resist a well-written wrong answer. The frozen cascade, accept confident verdicts and escalate the rest to GPT-6 Astra, is directly copyable for eval pipelines.