Are the replays genuine Jev output, or does the browser re-simulate the decisions?

Saved decisions and saved frames. The engine and experiment scripts stay in the repository but the published interface has no creation or admin screen, and replays never call TypeSafe. The race film also keeps the lower-left panel of recorded request activity and driver choices, and the README is explicit that it does not substitute local decisions for Jev's.

Can I trust the leaderboard if anyone can edit their own client?

No, and the project does not claim you should. The README calls it a spectator experiment rather than a competitive leaderboard, because visitors control the client and can submit invented race data to their own archives. Treat the leaderboard as a record of what the recorded runs did, not as a validated competition result.

How much will one race cost me in Jev calls?

There is no built-in budget, which is the honest answer. Each active driver has its own request, at most one in flight, three Choice questions per call, and each driver sends a fresh observation on the next physics step rather than on a timer, so frequency tracks API latency and can exceed two calls per second per driver. Ten cars is a lot of traffic; your TypeSafe account controls are what actually cap it.

Can I run this locally without handing my browser an API key?

Locally, yes — npm ci and npm run dev on Node 22, with the key entered in the page. But the key still travels from the browser to a same-origin server that forwards it to fixed TypeSafe endpoints, and the README says plainly that the hosting operator and TypeSafe both process the credential. It is a proxy, not end-to-end encryption.

What does the showcase actually present?

Ten driver prompts, 30 recorded races, and comparisons of the strategies tested, across a homepage, a championship page, a prompt lab, and a leaderboard.

How many circuits are recorded?

Six circuits with recorded Jev races, drawn from a 32-bit circuit generator with 4,294,967,296 seed values; the archive does not claim results for untested seeds.

What does Jev decide in a race?

Jev chooses a road lane, a pace, and a battery mode, and a shared steering and speed controller executes those targets at 40 Hz using only current local road samples.

CommunityJev Demos & Experiments

F1

JEVRACE is a public archive of a racing experiment: ten driver prompts, thirty recorded races and six circuits, watchable in the browser with no API key and no further inference calls.

See highlights

Open github.com

Stars: 1
Language: JavaScript
License: MIT
Published by: Community
Author: MartinPuli
Last verified: Oct 2026

MartinPuli/f1

User repository on GitHub

View on GitHub

JEV Prix: five AI drivers, unknown procedural circuits, Formula-inspired racing, BYOK Jev and saved replays.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

JEVRACE is a public showcase of a racing experiment with Jev: ten driver prompts, 30 recorded races, and comparisons of the strategies tested. The homepage plays recorded race footage; visitors can watch the original 3D replays without an API key or new inference calls. / — fullscreen race footage and six recorded circuits. /championship.html — recorded races, grouped into viewing seasons. /prompts.html — driver-by-driver prompt versions, race times and replay links.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

The published site only plays recordings. Replays use saved decisions and saved frames, so nothing calls TypeSafe while you watch.
Pages cover fullscreen footage, a championship grouped into seasons, per-driver prompt versions with lap times, and a leaderboard.
Jev picks a road lane, a pace and a battery mode; a shared controller at 40 Hz turns those into actual steering and speed.
Vercel plus Neon stores each browser's archive behind a signed HttpOnly cookie, with no user accounts and a 90-day expiry.

Watch out

This is a spectator experiment, not a competitive leaderboard: the client runs in the visitor's browser, so anyone can submit invented race data into their own archive. Jev does not drive the wheels — a shared steering and speed controller executes its targets at 40 Hz using current local road samples and slows to recover when a car leaves the road, and drivers never see the circuit seed or the full track. The 32-bit generator supports a huge seed space but the archive only contains six circuits with recorded races, and no claim is made about untested seeds. Keys are held in browser memory and proxied to fixed TypeSafe endpoints, which the README explicitly says is not end-to-end encryption, and prompt and race name fields are stored on purpose, so secrets must not go in them. Each active driver keeps one request in flight with three Choice questions per call and can exceed two calls per second, and there is no dollar budget — provider account controls are the only spend limit. Local development binds to loopback and should not be exposed. MIT licensed, with no affiliation to Formula 1 despite the driver numbers mirroring the published entry list. 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

computer-use-cache

Drop-in OpenAI-compatible cache proxy for repeatable agent workflows: exact hits stay local, and with a TypeSafe key Jev judges whether a cached response is semantically reusable for the current request.