What it is
Reasoning-enhanced Jev-style decision model that improves classification accuracy and calibration, with Jev-compatible serving and speculative decoding. Serve a Jev-compatible classifier that reasons before answering yes/no, multiple-choice, and rating questions, then returns calibrated probabilities from a single 9B model.
Key results
- Scores 0.889 test accuracy vs 0.857 for Jev and 0.822 for Kev-9B
- Reaches 0.935 on JevBench overall and 0.865 on JevBench hard
- Runs in about 0.3 s without thinking; 3.3 s median with thinking on 1 H100
Shows that adding explicit reasoning to Jev-style decision models can materially improve accuracy and calibration on structured classification.