Is the race fair, or is Jev winning on a technicality?

The obvious rigging is handled: order is shuffled every game, your tap is timed on pointerdown rather than click so the finger lift does not count, and the verdict stays hidden until you choose, so you cannot read the answer off the card. What remains asymmetric is instrumentation. latencyMs covers only the Jev call and the edge-to-TypeSafe hop as measured inside the Worker, not your trip to the Worker, so the comparison is a network round trip against your own reaction plus upload.

What happens to the scoreboard when Jev times out?

There is a 3 s timeout and no retries, on the reasoning that a retry would quietly add to Jev's time. On timeout the route returns 504 with { error: "timeout" }, the round shows that Jev did not respond in time, and the round counts as wrong for Jev while being left out of its average. That is defensible for a game, and it is also a reason to read the win rate with the timeout count next to it.

Can I point it at my own spam instead of the built-in messages?

No, by design. text must be one of the 25 messages defined in public/messages.js; anything else is refused. The reasoning is on the page: the endpoint is public and spends your key, so it will not classify arbitrary input. If you want your own corpus you would need to fork the Worker, and you would want to think about rate limiting before exposing it.

Can I run it without an API key, and would the score mean anything?

You can run it: with no key the Worker falls back to a keyword MOCK with a fake delay, and the page shows a MOCK badge. That badge is the important part, because the README states plainly that the mock is not Jev. A MOCK result tells you the frontend, the Worker route and the scoring work, and tells you nothing about the model.

How does BeatJev decide each spam question?

One Noul question, is this spam, a scam or deceptive marketing; isSpam is probability >= 0.5, and confidence is derived as |probability - 0.5| x 2 since a Noul answer only carries the probability.

What does BeatJev do on a Jev timeout?

The Jev call has a 3 s timeout and no retries; on timeout the route returns 504 and the round shows Jev did not respond in time, counting as wrong for Jev and excluded from its average.

How is BeatJev deployed and where does the key live?

It is a Cloudflare Worker: npx wrangler secret put TYPESAFE_API_KEY then npx wrangler deploy; the API key stays server-side because src/worker.js calls Jev through the SDK.

CommunityJev Demos & Experiments

Beatjev

A Cloudflare Worker that races you against Jev over 25 spam-classification rounds: a countdown reveals a message, your tap and Jev's Noul question fire on the same tick, and the results screen compares speed and accuracy.

See highlights

Open github.com

Stars: 0
Language: JavaScript
Published by: Community
Author: lambertsj
Last verified: Oct 2026

lambertsj/beatjev

User repository on GitHub

View on GitHub

try to beat jev

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A human vs. TypeSafe's Jev in a 25-round spam-or-not reaction race. Each round has a 3-2-1 countdown. When it hits zero, the message appears, your timer starts, and the page asks Jev the same question, all on the same tick. After 25 rounds, a results screen compares speed and accuracy and gives you a share link for X and a result image.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

One Noul question asks whether a message is spam, a scam or deceptive marketing; isSpam is probability >= 0.5, and confidence is |p - 0.5| x 2.
The key stays server-side: src/worker.js handles POST /api/classify through @typesafe-ai/sdk, and public/ is plain static HTML, CSS and JS with no build step.
Fairness is deliberate: order shuffles per game, your tap is timed on pointerdown, and the verdict stays hidden until you pick.
Without a key the Worker answers from a keyword MOCK with a fake delay and the page shows a MOCK badge, which the README is explicit is not Jev.

Watch out

The endpoint is public and spends your key, so it deliberately refuses to be a general classifier: text must be one of the 25 messages in public/messages.js, and anything else is rejected. That is the right default, but it does mean the Worker is an open endpoint on your account if you deploy it. The timing is not symmetric in what it measures: latencyMs is taken in the Worker around client.systemOne() only, so it is Jev's decision time plus the edge-to-TypeSafe hop and excludes your own trip to the Worker. Timeouts flatter the model too: with a 3 s limit and no retries, a timeout returns 504, the round reads as Jev not responding in time, counts as wrong for Jev, and is excluded from its average. Confidence is arithmetic, not a model statistic: a Noul answer carries only a probability, so the reported confidence is |probability - 0.5| x 2. Deployed with wrangler, you need wrangler secret put TYPESAFE_API_KEY before the first deploy. 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.

More Jev resources like this

laya-vs-jev-arena

Side-by-side game harness where Laya running locally and Jev over API answer the same typed questions for snake, kombat, runner, flappy, tetris, and human-versus-model matches.