User repository on GitHub
Research notes and an interactive Cloudflare Worker demo for Jev, TypeSafe AI's structured decision model
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Jev — zeke
Research notes on Jev plus a Cloudflare Worker demo you can paste a support ticket into. Useful for the failure modes rather than the pitch: no arithmetic, unreliable date comparison, scores that are not measurements, and speed and cost figures from TypeSafe's own evaluations.
- Stars
- 3
- Language
- TypeScript
- Published by
- Community
- Author
- zeke
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Jev — mayank953.
About this project
My research notes on Jev, TypeSafe AI's new structured decision model, plus a working demo running through Cloudflare AI Gateway. See AGENTS.md for the technical details on the example project. This file is the human-readable version: what Jev is, what it's good for, and where it falls apart. Try the live demo: Paste a support message, and Jev triages it (department, urgency, frustration) in one call. The page renders the probability bars, the confidence scores, and a routing decision, live.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- Three primitives answered in one parallel call: Noul for a probability, Choice over up to 255 labels, Score across 2 to 10 levels.
- Documents where Jev breaks: no arithmetic, unreliable date comparison, scores that are not measurements, indirection and context rot.
- Flags that a zero structural type-error rate only constrains the answer's shape; a wrong choice from the allowed set is still possible.
- Treats TypeSafe's speed and cost claims as a ceiling: they come from the company's own workflow evaluations, not an independent benchmark.
Watch out
These are one person's research notes plus a demo, not a benchmark. The speed range, the price per million input tokens and the large multiples quoted all originate from TypeSafe's own evaluations of its internal workflows against other models' structured-output paths, with newer frontier models left out, and the notes themselves say to treat them as a ceiling. The live demo is a single Worker that triages a pasted support message into department, urgency and frustration; it does not demonstrate the games, browser automation or game-playing results described elsewhere in the notes. Ecosystem examples are gathered from TypeSafe docs, third-party write-ups and posts, so their numbers are reported rather than reproduced. The MIND news anecdote, where judged usefulness correlated negatively with early click-through, is offered as an illustration of the model's limits rather than a controlled evaluation. The notes also state Jev is not yet hardened against adversarial input. 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.
Treat them as a ceiling, which is what the notes themselves advise. The end-to-end latency range, the price per million input tokens and the multiples quoted as faster and cheaper all come from TypeSafe's own evaluations of a handful of internal workflows against other models' structured-output paths, with no newer frontier models included and no independent benchmark offered.
Arithmetic, date and time comparison, and interpolation between Score levels: a score of 1.3 does not mean 40% of the way from level 1 to 2, because each level is judged on its own. The notes also flag that one-hop indirection such as double negatives hurts, that irrelevant context degrades answers, and that the model is not hardened against adversarial input yet.
No, and the notes draw that distinction explicitly. What the guarantee constrains is the shape of the answer, since it cannot contain a value outside the schema you supplied. Jev can still pick the wrong option from the set you allowed, and the calibrated-confidence claim is about many predictions on average, not any single answer.
Paste a support message into the Worker demo and it triages department, urgency and frustration in one call, rendering probability bars, confidence scores and a routing decision live. The site itself is plain HTML, CSS and JavaScript on Workers with Cloudflare AI Gateway in the path, while the technical details of the example project live in AGENTS.md rather than the README.
The author's phrase is that Jev is a smart if statement: ordinary code branches on values it can compute, and Jev fills the gap when the condition is a judgment call. The name nods two ways: System One is Kahneman's fast, intuitive thinking, and Jev shortens William Stanley Jevons, the economist behind the Jevons paradox, a bet that cheap decisions get used in places you would never call an LLM.
Barely, according to the notes. Every question in a request runs against the same shared state at the same time, so 13 questions cost about the same latency as 1, which TypeSafe calls speculative fan-out: ask everything you might need in one call and let your code sort out which answers matter. The model also carries a 32,000 token context window and roughly a 64,000 token budget shared across state and questions.
The MIND news anecdote. Someone ran Jev over 3,448 articles giving only the title and abstract, asking whether each seemed useful; the correlation between judged useful and actual early click-through came back negative, about -0.129. The notes read it as a reminder that Jev will confidently answer a question the input cannot actually answer, which is why the failure modes matter more than the pitch.
More Jev resources like this
98GitHub stars
Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
1Panel-dev/laya-serverPython
1Panel-dev/laya-serverPython20GitHub stars
General-purpose System One agent where Jev chooses tools and arguments while Jot executes them and loops until it can reply.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
runta-dev/jotTypeScript
runta-dev/jotTypeScript23GitHub stars
Browser drone-fleet lab where TypeSafe Jev steers each drone as fast reflexes and an optional planner offers one-use strategy advice when confidence drops, without pausing control.
Jev Demos & Experiments#community#jevbest#catalog-import
23GitHub stars
Chrome extension that sends each X post to Jev with six typed questions in one call near the viewport, then renders a behavior chip under the post plus a session cost and latency panel.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
oso95/x-scannerTypeScript
oso95/x-scannerTypeScript14GitHub stars
Manifest V3 Chrome extension in TypeScript that sends each LinkedIn post to Jev and collapses bait, self-promo, off-topic, and sponsored posts into a one-line bar with reasons and scores.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
stefw/lkcleanTypeScript
stefw/lkcleanTypeScript11GitHub stars
Standalone Flue example that routes support chat through Cloudflare Workers AI, using gpt-4o-mini for chat and Jev for typed intent and urgency routing.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
matthewp/flue-jev-demoTypeScript
matthewp/flue-jev-demoTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.