What it is
OpenJev is an open-source System One decision server. Send a state and typed questions (yes/no, choice, score); it returns calibrated probabilities and confidence in tens of milliseconds. It runs DiffusionGemma 26B-A4B via vLLM on NVIDIA GPU or MLX on Apple silicon.
How it uses Jev
Jev reads answers directly from DiffusionGemma's probabilities: it builds a canvas with one masked answer slot per question, runs one read-only pass, and uses the probability distribution over label tokens as the answer. No text is generated or parsed. It returns noul, choice, and score answers with confidence.
Primitives:choicescorenoul
Technique worth stealing
Read answers from a diffusion model's masked-token probabilities instead of generating text.
Try it
pip install typesafe-sdk; set TYPESAFE_BASE_URL and TYPESAFE_API_KEY; call client.system_one().
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-centricyes0.94
- Shows a System One patternyes0.73
- Handles uncertaintyno0.05
- Measuredno0.22
- Runnableyes0.96
- Worth recommendingno0.31
- Model replicayes0.94
- Problem scopescore on a 0–2 scale1.02
- About Jevyes0.98
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 21 Sept 2026.