OpenJev
Run System One-style typed decisions on any Hugging Face instruct model, with a typesafe-sdk-compatible server and calibration tools.
Overview
OpenJev (GitHub30) implements System One-style decisions on any Hugging Face instruct model, such as Qwen, Llama, Gemma or SmolLM. Each question is answered in one forward pass over a closed candidate set (yes/no, option numbers or level numbers) without sampling, so off-schema values cannot occur. All questions and candidates in a request are scored in one batch, and the shared state's KV cache is computed once, keeping time nearly constant in the number of questions. The README is in Japanese.
Probabilities are a softmax over the candidates, and confidence is derived from the distribution's shape. Calibration can be improved with temperature scaling (openjev calibrate) and LoRA fine-tuning with a proper scoring rule. The POST /v1/systemone and GET /v1/models formats match TypeSafe's OpenAPI, so the official typesafe-sdk works unchanged. A Colab notebook walks through inference, evaluation, calibration, the API server and fine-tuning, and a mock backend supports tests.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this message express urgency? | yes · no |
| Choice | Which team should handle this ticket? | billing · integrations · other |
| Score | How severe is the customer's problem? | Cosmetic issue · Broken feature with a workaround · Blocking issue with no workaround |
How it works
- Load any instruct model as a backend.
- Score all questions and candidates in one batch with shared state cache.
- Optionally calibrate with temperature scaling or LoRA.
- Serve with openjev serve for SDK compatibility.
Strengths
- Works with any Hugging Face instruct model.
- Wire-compatible with typesafe-sdk.
- Built-in calibration tooling and mock backend.
Caveats
- Quality depends on the base model chosen.
- Documentation mainly in Japanese.
Getting started
git clone https://github.com/GitHub30/OpenJev.git && cd OpenJev && uv venv && uv pip install -e ".[hf,server,dev]", then openjev serve.At a glance
- Type
- Library
- Status
- prototype
- Language
- Python
- Runtime
- library, HTTP server
- Has code
- yes
- Has tests
- yes
- Stars
- 2
Use cases and integrations
How curated lists describe it
Open-weight System One model (Jev-compatible): calibrated noul / choice / score decisions in one forward pass, no text generation. Works with TypeSafe's typesafe-sdk unchanged.
wh000wh000/awesome-jev-live
Frequently asked questions
Which models can it use?
Any Hugging Face instruct model, such as Qwen, Llama, Gemma or SmolLM.
Does the official SDK work?
Yes, the server matches TypeSafe's OpenAPI request and response formats.
How is calibration improved?
With temperature scaling or LoRA fine-tuning using a proper scoring rule.
Related projects
Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
ProjectbetaGoGo harness for System One models: hosted Jev, open Kev in-process via llama.cpp, and Cloudflare Clef, with typed helpers and a CLI.
ServicebetaGoGo gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
ServicebetaGoTypeScript library for one-line local decisions with open models (kev-0.6b, kev-4b, open-jev), no API keys or cloud.
ProjectbetaTypeScriptPython library that finds numeric answers by bisection, asking Jev higher, lower or exactly at each guess.
ProjectbetaPythonSwift 6 bridge that runs System One decision models behind Apple's Foundation Models API, on device with Laya or via hosted endpoints.
ProjectbetaSwift