jev-lm (uhhfeef)
Character-level language model built from Jev Choice over 28 options, where each option is the full hypothesis text.
Overview
jev-lm by uhhfeef builds a character-level language model out of a classifier. The vocabulary becomes 28 Choice options (a to z, space and stop), and generation is an ordinary autoregressive loop. The README shows it answering 'what is the capital of france' with 'paris.' character by character, with probabilities at each step.
The author explains what makes it work: options are hypotheses rather than symbols (each option is the full text you would have if you picked it), every request carries the question, stopping is the 28th option whose hypothesis is the unchanged answer, and each step averages three shuffled orderings in one request to cancel Jev's bias toward whichever option comes first. Flags control maximum characters, the visible window, ensemble size and beam width.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which next-character hypothesis (or stop) best answers the question? | — |
How it works
- Take the question and current answer
- Build 28 options as full hypothesis texts
- Ask Choice over three shuffled orderings and average
- Append the character or stop
Strengths
- Shows how option framing affects Choice
- Order-bias mitigation by shuffling
Caveats
- Experimental; short answers only
- One call per character
Getting started
`uv venv`, `uv pip install -e ".[dev,experiments]"`, set TYPESAFE_API_KEY, then `jev-lm --question "what is the capital of france"`.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
A character-level language model built on Jev, a System One classifier.
hellogumbo/awesome-jev
Frequently asked questions
Why make options full hypotheses?
So candidates differ in meaning, which a classifier can judge.
How is position bias handled?
Each step averages three shuffled orderings in one request.
Is it practical?
No, it is an experiment.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Decision-forensics library that explains a Jev answer by ablating parts of the input and measuring how far the probability moves.
ProjectbetaPythonAdapter that turns existing LLMs into Jev-style decision functions with option scores, optional calibration and review thresholds.
ProjectprototypePythonDecision sidecar and Copilot CLI pre-turn hook that mimics Jev choice, noul and score using Azure Foundry Cohere Rerank.
HookprototypePythonToy chat that makes Jev talk by repeatedly choosing its next word from a tiny caveman vocabulary.
ProjectprototypeTypeScript