jev-numeric
JevNext: decodes numbers from Jev by walking a multiway decision tree of interval or digit Choice calls, without logits or training.
Overview
JevNext (repository jev-numeric) is a simple algorithm, with an arXiv paper and live demo, that gives Jev-like decision models numerical output. Jev natively offers choices, option probabilities and rubric scores but not numbers. JevNext treats a numeric answer as a path through a multiway decision tree: each Choice call selects a branch and the final leaf identifies a finite-precision value.
Branches can be described as numerical intervals (the default) or as decimal digits, where each step asks for the next digit given the prefix chosen so far, making digit decoding a ten-way instance of the same tree. Both use ordinary Choice calls with no training, regression head or token logits. An experimental mode builds a CDF or histogram, with calibration marked unverified. On 256 arithmetic expressions, the README reports 83.40% of answers within 5% relative error using interval decoding, against 80.47% for direct choice from a list that contained the correct answer.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which interval contains the value? | — |
| Choice | Given the selected prefix, what is the next decimal digit? | 0 · 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 · 9 |
How it works
- Define the numeric range and precision
- Ask a Choice over intervals (or digits) at the top level
- Recurse into the chosen branch with a new Choice
- Stop at a leaf and return the finite-precision value
Strengths
- Adds numeric output using only standard Choice calls
- No training or logit access required
- Outperformed direct choice in the reported arithmetic test
Caveats
- Multiple sequential calls per number increase latency and cost
- CDF/histogram mode has unverified calibration
- Custom licence
Getting started
See the Quick start section of the README; a live demo is at https://bring-ai.github.io/JevNext/At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Library and scripts
- Has code
- yes
- License
- Custom license
Use cases and integrations
How curated lists describe it
\*\*Both are multiway decision trees; decimal-digit decoding is a ten-way instance.\*\* On an aligned decimal grid, they can have identical branches and leaves, expressed through different prompts. The digit is a \*\*Choice option\*\*, not a vocabulary token; Jev returns the option probabilities directly.
logicrw/awesome-jev-projects
Frequently asked questions
How does JevNext produce numbers?
By asking a sequence of Choice questions that narrow an interval or pick digits one at a time.
Does it need token logits?
No. It uses ordinary Jev Choice calls.
How accurate is it?
On 256 arithmetic expressions the README reports 83.40% within 5% relative error versus 80.47% for direct choice.
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