WebGL demo of the card game Speed against a Jev-driven CPU, built to make decision latency and accuracy visible: every candidate move checked as a noul in one parallel request, a mechanical rule validator as referee, and p50/p95 latency panels.
User repository on GitHub
A WebGL demo where you play the card game Speed against a CPU whose brain is TypeSafe AI's Jev. The whole point of the app is to measure and show Jev's decision speed and decision accuracy in real time.
Repo stats from the GitHub API, cached Oct 2026.Homepage
A WebGL demo where you play the card game Speed against a CPU whose brain is TypeSafe AI's Jev. The whole point of the app is to measure and show Jev's decision speed and decision accuracy in real time. Jev is a System One model — it generates no text, and instead returns pre-defined typed decisions with probabilities in a single pass. Officially 70–500ms. This app makes that characteristic visible in the context of a game.
Quoted from the project's README. Read the full README.
Watch out
Needs Node 20 or later, and although runtime dependencies are zero, the first run needs network access for the Three.js CDN. Without TYPESAFE_API_KEY the server silently enters mock mode, returning plausible answers on an 80-250ms delay with roughly 3% wrong mixed in, so any latency or accuracy reading taken in that state is synthetic. In the Cloudflare Workers build, hitting a rate limit downgrades the request to a mock response instead of returning 429, so the game stays playable and only the billing stops, and Workers-side limits can only restrain calls made through the app. The lockout foul penalty is an original rule with no equivalent in the Speed rulesets the README cites. Copy reviewed against the repository README on 2026-10-04, not from running it.
trungdq88/jev-tetrisJavaScript
vinilana/live-jevJavaScript
bohutang/siftJavaScript
manifoldor/xtagsJavaScriptOriginal write-ups on what this entry is for.
Posts from builders shipping with Jev right now.
