What it is
Local decision model that reads a document once and returns calibrated probabilities for many typed questions in one forward pass. Answers many yes/no, multiple-choice, and score questions about one document in parallel, returning learned probabilities in a single prefill pass with no text decoding.
Key results
- Packed vs separate max delta is 3.7e-6, with packed 2.0× faster
- kev-4b gets 0.806 locked-test OOD accuracy and serves in ~1 s on M5
- Supports noul, choice, and score questions with 2-255 options
Shows that a small local LLM can do Jev-style multi-question decisioning in one pass with exact branch isolation.