Organization repository on GitHub
Experiment to trick Typesafe’s Jev, aka “the language model that won’t talk” into actually talking.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Demos & Experiments
Let Jev Speak
Coaxes free text out of TypeSafe's classification endpoint by decoding one word at a time: every word is a separate choice question over a vocabulary, and the loop feeds its output back as the prefix. The README itself calls the approach a demonstration, not a good idea.
- Stars
- 2
- Language
- JavaScript
- License
- MIT
- Reading time
- 2 min
- Published by
- Community
- Author
- suidouble
- Last verified
- Oct 2026
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- JavaScript
- Platforms
- CLI
- Capabilities
- Open sourceRequires API key
Highlights
- Decodes answers word by word: each word is a separate choice question over a vocabulary, and each pick is fed back as the prefix.
- One routing call picks among 28 domain packs; the README reports routing scores 60/60 on a labelled set and roughly doubles answer quality.
- README sample: 'Why is the sky blue?' took 14 calls (1 route + 3 prior + 10 decode) over a 199-word vocabulary expanded to 207 options.
- CLI flags: --domain skips routing, --domains lists the packs, --vocab twitter dumps a pack's word list; Node 18+ and no dependencies.
- The README states a single well-posed choice question answers the same question better in one call, with calibrated confidence attached.
Quickstart
bash
git clone [email protected]:suidouble/let-jev-speak.git
cd let-jev-speak
node ask.js "Should I add an index to this column?"Watch out
Copy reviewed against the truncated repo README on 2026-10-15; the excerpt stops inside the 'just the API client' section. The 60/60 routing score and doubled quality are the project's own README claims, not external measurements.
Questions about this entry
Short answers drawn from the linked project or page.
An experiment that makes TypeSafe's classification endpoint produce prose. Every word of the answer is a separate choice question over a vocabulary, and the loop feeds the previous output back in as the prefix until the answer ends.
One choice call routes the question to one of 28 domain packs, and the decode runs over that pack's words. The README's sample run cost 14 calls: 1 route, 3 prior and 10 decode calls.
Yes, a TypeSafe key from console.typesafe.ai/keys, passed as a string, an options object, a configured client, or via process.env.TYPESAFE_API_KEY; without one the constructor throws immediately. TypeSafe's own billing applies per call.
Yes, there is a Telegram bot at t.me/let_jev_speak_bot built on Telegram's serverless platform, with source in the tg_bot directory; send it a question and watch the answer assemble word by word.
max (default 10 words), min (5, suppresses end below this length), alpha (0.45, how much per-option prior to divide out), penalty (1.5 repetition damping), blend (mix close domains), priorMode and maxRetries of 3 for 429, 5xx and network errors.
Export TYPESAFE_API_KEY or pass --key, then run node ask.js with your question. Add --domain finance to skip routing, --domains to list the domain packs, or --vocab twitter to dump a pack's word list.
No, and the README says so directly: a single well-posed choice question answers better, in one call, with calibrated confidence attached. The project is a working demonstration of what the endpoint can be pushed into doing and where it stops.
More Jev resources like this
Text generation with a model that never generates text: jev-gpt asks Jev one typed question per choice — word type, WordNet category, a group of about 200 words split by embeddings, then the word — about 400 calls and 2 cents per prompt.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
florian-hoenicke/jev-gptPython
florian-hoenicke/jev-gptPythonA workshop of small, inspectable Python experiments for TypeSafe System One models: ten critical bays, from fraud screening to emergency dispatch, each giving Jev a realistic state while deterministic policy keeps authority over the next action.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
jgalego/jevs-garagePython
jgalego/jevs-garagePythonA ChatGPT-style chat UI for Jev where answers are decisions, not prose. Pick Yes/No or styles like Yes/No/Maybe, Pirate or Tabloid, and expand any reply to see Jev's calibrated probability for every option. Runs on Deno.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
kt3k/jevchatJavaScript
kt3k/jevchatJavaScriptOpen-source Codex plugin that consults TypeSafe Jev about tools, models, tasks, skills, context, and strategies, diagnoses failed commands, and checks completion claims against evidence. The current package is version 0.4.0.
Jev Agents & MCP#community#catalog-import#logicrw
Vercel Labs example that routes three forms by context with Jev and AI SDK: Jev's team choice is accepted when its confidence is at least 95%, otherwise openai/gpt-6-luna-fast independently evaluates the submission and its choice becomes final.
Jev Demos & Experiments#community#catalog-import#logicrw
A Codex plugin that prepares traceable context handoffs: Jev answers one narrow classification question per old, verified read-only tool pair, and code validates the answers, preserves protected content and builds a JSON or Markdown packet.
Jev Agents & MCP#community#catalog-import#logicrw
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.