@Centific_CAIR reports Jev matched Claude Sonnet 4 spotting harmful AI replies (F1 0.758 vs 0.717) at 176× lower cost on 498 samples; newer Claude models often refused to judge.

Benchmark note. Centific Applied AI Research says switching the AI judge cut evaluation costs more than trimming prompts. They tested TypeSafe's Jev, which returns verdicts without written reasoning, against three Claude models on 498 AI replies. Jev matched Claude Sonnet 4 at spotting harmful ones (F1 0.758 vs 0.717) at 176 times lower cost. Claude Sonnet 5.5 and Opus 5.5 often refused to judge and called more than half of the harmful replies they did judge safe. The post links a Centific blog write-up and has no attached media.