jevspeak
Chat interface with no generative model: Jev picks intent, stance, claim and tone, and a deterministic compiler writes the reply.
Overview
JevSpeak explores whether a model that cannot generate text can still hold a conversation. On each turn, Jev is asked about 13 questions in parallel about the user's message: intent, topic, emotion, which speech act fits, which way a judgment leans, which finite claim best fits, what caveat applies, what tone and how long. The answers are probability distributions over finite options plus a few scores such as confidence and emotion intensity.
Those decisions are normalised into a typed semantic representation, and a deterministic language compiler turns it into an English or Chinese sentence; the same representation always yields the same sentence. Confidence maps to hedging bands such as 'I think'. The README states that no GPT, Claude or Gemini model is used at runtime and there is no hidden rewrite step. A live site shows a 'Jev Brain' panel with the distributions and a debug view of the compiler trace. A mock mode uses a deterministic scorer, and the finite claim vocabulary limits what it can discuss.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the intent of the user's message? | — |
| Choice | Which way does the judgment lean?(inferred) | mostly_yes · mostly_no |
| Choice | Which finite claim best fits as the main answer? | — |
| Choice | What tone should the reply take? | — |
How it works
- User sends a message
- Jev answers about 13 parallel questions on meaning and response plan
- Answers are normalised into a semantic IR
- A deterministic compiler renders the IR as English or Chinese
- UI shows distributions and the compiler trace
Strengths
- Fully inspectable path from decisions to words
- Same IR always gives the same sentence
- Bilingual rendering from the same decisions
Caveats
- Finite claim vocabulary limits topics
- An MVP with supported domains only
- Live mode needs a Jev key
Getting started
Try it at https://jevspeak.org.At a glance
- Type
- Demo
- Status
- prototype
- Language
- TypeScript
- Runtime
- web app
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Jev can't generate text. So I made it talk anyway. A conversational interface built from probabilistic decisions and a deterministic language compiler — no generative LLM.
logicrw/awesome-jev-projectsJev can't generate text. So I made it talk anyway. A conversational interface built from probabilistic decisions and a deterministic language compiler — no generative LLM.
hellogumbo/awesome-jev
Frequently asked questions
How does it talk if Jev cannot generate text?
Jev chooses meaning from finite options and a deterministic compiler turns that into a sentence.
Is any LLM used at runtime?
No, the README states no generative model is used to produce replies.
What can it talk about?
Only what its finite claim vocabulary covers, mostly judgment questions in supported domains.
Related projects
Turns any local LLM into a Jev-style engine via masked-logit softmax, or proxies OpenAI-compatible APIs or Jev; includes a chat crush radar.
ServiceprototypePythonA chatbot with no language model: every word is chosen by Jev Choice questions from thousands of options, with lookahead and self-critique.
ServiceresearchPython- jev-chat-windows★ 802
Windows WeChat reply assistant: local offline OCR reads the chat, Jev judges intent and tension and ranks three drafted replies to fill in.
ProjectproductionPython AstrBot plugin where Jev decides whether the bot should join a group conversation; the main model still writes the reply.
ProjectproductionPythonDeepSeek Harness plugin that classifies each user message's emotion, urgency and department with Jev and injects it as context.
PluginbetaTypeScriptiOS custom keyboard that shows a copied message's intent, risk level and candidate replies inside any chat app, filling but never sending.
AppbetaSwift