Comparative evaluation of the Jev decision model on content safety tasks: who is stronger against traditional classifiers and LLM judges

Cisco compared the zero-shot decision model Jev (TypeSafe) with Laya (ConvAI Innovations) and a self-trained 1B-parameter encoder classifier across 24 harm categories and 6 evaluation datasets; under a 0.5% false-positive budget, the self-trained model had the highest recall on all 6 datasets, and Jev led Laya in AUC on 5 of 6. On Cisco's multi-turn test set, the self-trained model recalled 61%, Jev 34%, and Laya almost zero. Jev was also compared with Gemma 4 31B (with thinking disabled) on 8 public benchmarks, with Jev having higher AUC on 7 benchmarks and tying on 1.