Safety3 sources
5d agoJev Model Matches Top LLMs In Monitoring
The non-autoregressive model Jev 1.13 matches GPT-5.5 and Opus 4.7 in detecting sabotage and sandbagging. It achieves this at 1/200th of the cost of Opus 4.7. However, it fails at multi-step reasoning, ranking 209th of 279 on the NCRI suite.
The Signal
These results suggest Jev is best suited for low-cost, always-on agent monitoring.