jev-chat
Research decoder that treats Jev as a language model, assembling replies from repeated choices over phrases, words and letters.
Overview
jev-chat by Adhyaay Karnwal is a research decoder that deliberately treats Jev, a decision model that returns distributions rather than tokens, as if it were a language model. It repeatedly asks Jev for the next surface unit, samples and appends, with the string assembled in code; the author notes this autoregressive use is the wrong use of the model, while one-shot selection of a complete reply is the native one, and the accompanying paper measures both.
The method uses a hierarchical codebook within the 255-option Choice limit, mixing phrases, copy spans from the user, function words and frequent content words plus an other option, with a letter-then-word fallback asked in parallel only when other wins. Speculative fan-out asks for the next unit and hypothetical follow-ups in one call and code stitches an accepted path, and unit size adapts from phrases to words to characters, copying n-grams so names and numbers are not spelled out. It has a CLI, a demo server that streams units with the per-call Choice mass in a side trace, tests and benchmarks. It needs Python 3.11+ and is MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the next phrase, word or character of the reply? | — |
How it works
- Build a hierarchical codebook for the prompt
- Ask Jev for the next unit plus speculative follow-ups
- Code samples and stitches the accepted path
- Repeat until the reply ends
Strengths
- Shows the limits of using a decision model for generation
- Transparent trace of every choice
- Paper with measured comparisons
Caveats
- Explicitly not the intended use of Jev
- Output quality limited compared with LLMs
Getting started
`uv sync --extra dev`, set TYPESAFE_API_KEY in .env, then `uv run jevchat "What is a System One model?"`.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CLI and demo server
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
A research chat decoder that repeatedly asks Jev to choose words or phrases and assembles them in code.
logicrw/awesome-jev-projectsA chatbot from typed Jev decisions: hierarchical speculative decoding over System One probabilities.
hellogumbo/awesome-jev
Frequently asked questions
Can Jev write chat replies?
Only by repeated choices assembled in code, which the author calls the wrong use; selection of a full reply is native.
How are names and numbers handled?
By copying n-grams from the user's text instead of spelling them.
Is there a paper?
Yes, in the repository's paper folder.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Reads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.
ProjectresearchPythonWord-level language model whose output layer is Jev: next-word choices, Noul-verified drafted chunks and bits-per-token evaluation.
ProjectresearchTypeScriptRunnable yardstick that tests Jev's three headline claims on a small local model: typed output, no hallucination, calibration.
ProjectresearchPythonOpen Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.
ProjectresearchPythonOpen Jev replica on Gemma 4 E2B and Gemma 3 270M served on Modal, with eight demos and a TypeSafe eval head-to-head.
ProjectresearchPython