Why build a chat decoder on a model that does not emit tokens?

One-shot selection of a complete reply is the TypeSafe-native route. Autoregression here loops greetings, copies operands, and is measured in the paper mainly to show why selection fits a decision model.

How does a 255-option Choice cover an open vocabulary?

A Choice holds at most 255 options mixing phrases, copy spans, function words and frequent words plus other. Only when other wins does a parallel letter-then-word pass cover the long tail.

What setup is needed before the demo runs?

You need Python 3.11+, a TypeSafe key in .env, and uv sync. The CLI, local demo server and pytest suite are the documented paths; matplotlib is only needed for the bench script.

Does Jev generate any of the reply text?

No. Jev never emits the reply; the string is assembled in code from sampled units, with copy n-grams so names and numbers are carried over rather than spelled out.

What do speculative fan-out and adaptive unit size actually do?

Independent questions share one state, and each call asks for the next unit and for hypothetical follow-ups of the form assume the prefix was just extended by u; code then stitches an accepted path. Adaptive unit size picks phrases when they fit, words otherwise, and characters only to finish a partial word. It is a research decoder, so the point is the trace, not the output quality.

What does it fail at, according to the README?

Naive autoregression loops greetings and copies operands, so 1+1 becomes 1. Greedy decoding recovers short facts and still emits ungrammatical strings. Select is grammatical when the answer is in the candidate list and silent otherwise, which is the honest summary of the approach: useful where the answer is enumerable, silent where it is not.

What is checked in, and what license is it under?

The paper PDF, a method note at paper/METHOD.md, run traces at experiments/results/runs.json, the decoder, codebook, TypeSafe client and demo server under src/jevchat, and tests for codebook, sampling and a scripted decoder. The demo UI streams sampled units with the per-call Choice mass in a side trace, and the README is explicit that there is no hidden generator behind it. MIT, and not affiliated with TypeSafe AI.

CommunityJev Practices & Patterns

Jev Chat — adhyaay-karnwal

Research decoder that treats Jev as if it were a language model, assembling replies in code from Choice distributions over a hierarchical codebook instead of generated tokens.

See highlights

Open github.com

Stars: 4
Language: Python
License: MIT
Published by: Community
Author: adhyaay-karnwal
Last verified: Oct 2026

adhyaay-karnwal/jev-chat

User repository on GitHub

View on GitHub

A chatbot from typed Jev decisions: hierarchical speculative decoding over System One probabilities.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Jev Chat — w3cj.

About this project

A research decoder that treats Jev (a System One decision model) as if it were a language model. Jev returns Choice / Score / Noul distributions. It does not emit tokens. This repo asks for the next surface unit anyway, samples, and appends. Autoregression over that interface is the wrong use of the model; one-shot selection of a complete reply is the TypeSafe-native one. The paper measures both. Documented in paper/METHOD.md and the paper:

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Hierarchical codebook mixes phrases, copy spans and frequent words, with adaptive phrase, word and character units.
Speculative fan-out asks for the next unit and hypothetical follow-ups sharing one state.
Demo UI streams sampled units with per-call Choice mass in a side trace.
Paper, method note and run traces are checked in alongside the decoder.

Watch out

Autoregression is presented as the wrong use of the model: greedy decoding recovers short facts but emits ungrammatical strings, and Select stays silent when the answer is not in the candidate list. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.