Researchers Find Jev Can Detect AI Alignment Failures at 63x Lower Judge Cost

A new arXiv study reports that Jev, TypeSafe AI's decision-focused model, can serve as a zero-shot detector for several AI alignment failures while sharply cutting evaluation costs. The researchers introduce RLCDAlignBench, a benchmark of 7,193 detection instances across 44 benchmarks and ten failure types, and report that a single generic Jev question reached a median AUROC of 0.886 while costing 63x less than LLM-judge scorers on the measured judge-scored subset.