jevify
Makes any logprob-capable model answer typed questions Jev-style in one pass, in Jev's API shape, with probes and scorecards.
Overview
jevify (fidecastro) makes a model you already serve answer typed questions the way Jev does, and measures how well it manages it. You bring a model behind an OpenAI-compatible endpoint with logprobs (vLLM, llama.cpp and others) or a small in-process model. jevify sends the state once with a warm cached prefix, then asks Choice, Score and yes/no questions, reading each answer straight from the model's next-token distribution with no generation or parsing, about 30 to 140 ms per question on a local GPU. Results come back in exactly the shape of TypeSafe's Jev API, so the unmodified typesafe-sdk works against it. Images can be part of the state.
jevify owns no weights and trains nothing. A recipe describes how a model is asked, a probe proves what a deployment can do, and a scorecard reports how it did, with the command, hashes and raw file behind every number. It installs from PyPI with uv or pipx, has CI, and ships recipes such as one for Gemma via llama.cpp.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this? | billing · support · sales |
| Score | How urgent is this? | low · medium · high |
| Yes / no | Does the customer threaten to cancel? | — |
How it works
- Write or copy a recipe for your model server
- Probe the deployment's capabilities
- Ask typed questions; answers read from next-token probabilities
- Produce a reproducible scorecard
Strengths
- Works with models you already run
- Jev-compatible responses for the official SDK
- Reproducible scorecards with hashes
Caveats
- Probabilities depend on the base model and are not RLCD-calibrated
- Requires logprob support
Getting started
uv tool install jevify, then `jevify probe recipes/gemma-4-e4b-it.llamacpp.yaml` and `jevify ask ...`.At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- CLI / local server
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Supersimple way to serve LLMs as a Jev-like endpoint.
hellogumbo/awesome-jevMakes a model you already serve answer typed questions in one pass the way Jev does, and measures how well it manages it.
MrJev/awesome-jev
Frequently asked questions
Does jevify train a model?
No, it owns no weights and trains nothing; it reads answers from your model's token probabilities.
Can I use the official SDK?
Yes, responses match Jev's API shape.
How fast is it?
About 30-140 ms per question on a local GPU, per the README.
Related projects
llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonLocal server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.
ProjectresearchRustEducational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.
ProjectresearchCExperimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.
ProjectresearchPythonReads typed decisions from one forward pass of PrismML's 1.75-bit Bonsai 27B under llama.cpp, using the OpenJev/SemIf logit method.
ProjectresearchPython