What it is
An Apache-licensed Python toolkit that turns local text and vision language models into Jev-style structured decision models, returning Choice, Score, and Noul answers via Python API or HTTP service for developers.
How it uses Jev
It reads causal-model logits during prefill to compute probabilities for Choice, Score, and Noul questions over a given state and instructions, assembling results without token-by-token decoding. Results are returned as structured JSON through the Python API or a System One-shaped HTTP endpoint.
Primitives:choicescorenoul
Technique worth stealing
Prefill-only scoring: compute probabilities from logits during prefill and assemble results directly, avoiding token-by-token decoding.
Try it
git clone, uv sync --extra sglang, then python examples/sglang_inference.py --model-path /path/to/model
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricyes0.91
- Shows a System One patternunclear0.42
- Handles uncertaintyno0.05
- Measuredno0.15
- Runnableyes0.92
- Worth recommendingno0.23
- Model replicayes0.92
- Problem scopescore on a 0–2 scale0.93
- About Jevyes0.93
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 23 Sept 2026.