What Is Jev? TypeSafe's System One Decision Model, Explained
Jev is TypeSafe AI's System One model: it returns typed choices, scores and yes/no answers with probabilities instead of text. Here is what that means and who is building on it.
Most language models answer by writing. Jev answers by deciding. TypeSafe AI describes it as a "System One" model: fast, intuitive judgment that returns a typed result (a selected choice, a score, or a yes/no answer) together with probabilities, rather than generated prose. Projects built on it describe it the same way: jev-mcp by BYK says Jev "returns typed judgments (noul, choice, score) with probabilities instead of generated text," and jevchat calls it a model that "never generates text, it only decides."
That sounds like a narrow trick until you look at how agents actually spend their time. A large share of agent steps are not writing at all. They are picking: which tool, which model, which button, is this safe, is this urgent. A decision-only model is built for exactly those steps. This article explains the idea, then maps the open-source ecosystem that formed around it, with honest notes on how much evidence sits behind each project.
Key Takeaways
- Jev returns types, not text. Choice, score and boolean ("noul") results come back with probabilities you can threshold.
- It targets the "picking" steps of agents. Routing, gating, triage and ranking are the natural fit; drafting prose is not.
- The ecosystem is brand new. The projects in this series were mostly created in mid-September 2026, so most have single-digit stars and thin evidence.
- Claims vary widely in quality. Some come with published methodology and caveats; many are README assertions nobody has retested.
- Typed output is not correct output. Probabilities need calibration before you trust a threshold.
What "typed decision" means in practice
Ask a chat model "is this email about an invoice?" and you get a sentence you must parse. Ask Jev the same and you get a structured result. In typesafe-jev-workflow, a small LangGraph demo, a mocked email goes to Jev and comes back as a typed Choice (invoice or general) with confidence scores, and the graph routes on that value. The author is upfront that the demo's ten labeled examples are "a smoke check, not an accuracy benchmark."
The BYK server lists three question types: noul (boolean yes/no), choice (pick from options with criteria) and score (rank severity or importance on named criteria). Every answer carries a distribution over the options, which is what makes threshold logic possible: act automatically above one confidence, escalate to a bigger model below it.
For a deeper contrast with free-text agents, see Typed Decisions vs Free Text.
What people are building
The public radar at awesome-jev-projects groups the projects by category. The recurring patterns:
| Pattern | Example | What Jev decides |
|---|---|---|
| Routing | Jev Codex Router, JevRouter | Which model, tool or subagent handles a turn |
| Browser action | jev-ultrafast | Next action and target element |
| MCP access | typesafe-mcp | Choice / Score / Noul on demand for agents |
| Calibration | jevcal | Where to set confidence thresholds |
Later articles in this series cover each of these in turn, starting with routing by difficulty and browser decisions.
There is also an index of the index. jevsome-projects says it lists open-source projects that "provably call Jev," with each entry linking to the line of code that proves it. Its stated method includes confirming actual API calls, requiring the project to be built around Jev, and a five-star minimum. As of its September 18, 2026 refresh it reports 58 verified projects, with a documented list of rejected repositories. That filtering is useful precisely because the space is noisy.
Reading the evidence honestly
Timing matters here. As of September 2026, the most visible project, jev-ultrafast, has about 4,673 stars, but it is an outlier. Most others sit far lower: typesafe-mcp about 59, jkudish's jev-mcp about 67, Jev Codex Router about 26, and many at 0-5. Stars are a weak signal at any age, and these repos are days old.
Two more cautions:
- Author-run numbers are not independent numbers. Latency and cost figures in these repos come from their own authors' tests. Where they come with methodology, this series cites them with their limits.
- Probabilities are not truth. jev-mcp's own notes are that typed output does not guarantee factual accuracy and that thresholds need calibration. jevcal exists for exactly that reason.
What to do next
- Browse the awesome-jev-projects radar and the jevsome-projects list; note which entries link evidence.
- Look up the imported projects on aiskill.market via /browse.
- Pick one decision step in your own agent (a tool choice or a safety gate) and write down what a typed answer would look like.
- Read Typed Decisions vs Free Text next.