Jev vs Jev: Duel and Arena Projects Built to Stress-Test It

The awesome-jev-projects radar files six small repos under "Game Development," and the label is doing more work than it should. Only two of the six are actual head-to-head matches. One is a local server that reimplements Jev's own API so you can run your own arena without a TypeSafe API key. The remaining three are solo demos: a model making one discrete decision per tick against a fixed board or task, no opponent involved.

That's worth untangling before you read star counts as a scoreboard, because none of these six projects are directly comparable to each other. What they share is scale — all are new, all are small, all put Jev (or a reimplementation of it) into a tight decision loop where latency and correctness show up fast. This article sorts them into what they actually test, with the caveats each one's own documentation gives.

Key Takeaways

• Only two projects are literal duels: laya-vs-jev races Jev against a local MLX model at a browser dinosaur game; OneVOneJev pits Jev against an opponent in a 1v1 browser shooter.
• litjev is infrastructure, not a game — a local server implementing Jev's /v1/systemone endpoint on open-weights models, useful for running any of these demos without a hosted API key.
• jev-libero and RoboJEV are single-agent simulation tasks, not competitions, despite sharing the category tag.
• typesafe-snake is the simplest of the six: one legal-move decision per tick, illegal moves blocked by code, not by the model.
• Every one of these is new and small. None has independent verification of its claims; read each project's own "claims & evidence" language before repeating a number from it.

The two actual duels

laya-vs-jev (89 stars) runs a local MLX model against Jev on the same T-Rex-runner-style task, with live metrics and replay recording. The project's own documentation doesn't spell out the exact decision policy each side uses in detail, and there's no independent runtime or performance verification behind the comparison — it's a fun, inspectable setup, not a benchmark you should cite numbers from yet.

OneVOneJev (23 stars) is a browser-based 1v1 shooter where Jev reads structured match state — position, aim, firing — and answers on each decision tick, with a heuristic fallback if the API call fails. The README states a roughly 9 Hz tick rate, but that's the configured tick interval, not an independently measured sustained throughput figure, and this site has not retested it. Both projects are exactly the kind of thing you'd expect in the first weeks after a new decision API ships: fun to watch, thin on rigor, worth checking back on.

The infrastructure: litjev

litjev (42 stars, Apache-2.0) doesn't play a game at all. It's an open reproduction of Jev's decision layer running on open-weights models, serving the same /v1/systemone schema — choice, score, yes/no — by reading candidate token log-probabilities directly rather than generating text. That matters for this cluster specifically: if you want to run laya-vs-jev, OneVOneJev, or any of the tick-based demos below without a hosted TypeSafe key, litjev is the kind of local substitute you'd reach for. Its own caveat is candid — a locally calibrated probability distribution may not match hosted Jev's, so don't assume interchangeable behavior without checking.

Solo demos filed under the same tag

Three more repos share the "Game Development" category without involving an opponent at all.

jev-libero (61 stars, MIT) runs two LIBERO manipulation tasks through one control engine, each loading its own JSON task definition, with demo videos that follow simulation time rather than real wall-clock decision latency. RoboJEV (36 stars, Apache-2.0) is a small inspectable robotics lab where Jev receives structured simulator state — not images — picks an intent, then picks X/Y/Z directions and a gripper command, which a Cartesian controller executes against real MuJoCo contacts; each task has an independent physical success check the model can't talk its way past. Both are covered in more depth, alongside jev-drone, in jev-drone: putting a decision model in the flight loop.

typesafe-snake (21 stars) is the simplest: a React/TypeScript Snake autoplayer that supplies board state, fruit location, and the set of legal directions to Jev each tick, and takes whichever direction it returns. Illegal moves are blocked by code, not by the model, and the README notes tick frequency is bounded by network latency rather than model speed — a useful reminder that in a live decision loop, the round trip often matters more than the model.

For both jev-libero and RoboJEV, the dataset's own evidence note is blunt: these are checks that establish source-level Jev integration, not runtime, safety, or performance validation. Treat any number you see quoted from either repo the same way.

What to do next

1. If you're deciding whether to build a demo like this, start with what Jev is so you know what you're wiring in.
2. Want a self-hosted decision endpoint for a project like these? Read litjev's own README before assuming its calibration matches hosted Jev.
3. For the broader pattern of Jev in real-time loops — games, trading, robotics — see real-time Jev: loops, games and trading.
4. These six, plus kev, NanoJev, and jev-drone, are the full first wave of Jev game demos — see what a week of them actually proves.