Real-Time Jev Loops: Games, Driving Sims and Trading Bots

A chat model that writes a paragraph per decision cannot drive a car in a simulator or play Snake in real time. A model that returns a single choice with a probability can be called again and again. That is the pitch behind a small group of Jev projects built around fast loops: a driving demo, board-game experiments, a duel between two Snake players and a pair of trading bots.

Most are demos, and demos are what they should be judged as. They show that the loop works and expose the inputs and outputs for inspection. They do not show that Jev beats alternatives, and none of them claims to. This article covers what each repo actually demonstrates, then treats trading separately because the risks there are real. For the general idea, start with what Jev is.

Key Takeaways

• Demos, not benchmarks. Every project here is a showcase. Results depend on setup, model version and randomness.
• Small stars, real code. As of September 2026: jevpilot 58, Jev-Trades 7, jev-gomoku 1, jevarena 1, and two jev-trade repos at 3 and 0.
• Inspectable loops are the value. Several repos expose the exact request payload and returned probabilities per step.
• Trading is different. Paper trading and live order placement carry very different risk. One listed bot says plainly that it is real.
• This is not financial advice. Nothing here is a recommendation to trade.

jevpilot: a driving demo with a request budget

jevpilot (58 stars as of September 2026, no license listed) is described in its README as a demo showing Tesla Autopilot-like behavior. Jev receives compact tables of eligible paths, road boundaries, nearby traffic, signals and destination guidance, and picks among sampled steering-and-speed candidates. Detailed geometry and control calculations stay in local code, and a separate safety brake handles collision risk.

The cadence is stated: up to four requests per second near turns or traffic, about 1.5 per second on clear roads, with single-answer questions resolved locally. That is the honest shape of a "real-time" Jev loop: it is a few decisions per second, backed by code doing the fast work. A directory blurb talks about sub-100 ms latency, but the README documents request rates, not a latency measurement. You can inspect the exact payload through a JSON panel. Hosted play uses a login and free credit. See /skills/jevpilot-standardagents.

jev-gomoku: an experiment about input representation

jev-gomoku (1 star; README in Chinese) runs nine simultaneous 15x15 games between two Jev players. Each move is one Choice question over candidate coordinates. The point is to compare what the model is shown: board text only, tactical facts, short-term lookahead, a coordinate list or four-direction line counts.

Its stated limits are unusually good. Candidates are pre-filtered by code, so it is not an unassisted strength test. Choice probabilities are not win probabilities. Nine games cannot prove one input better. The README reports a saved replay of 144 real decisions you can step through without an API key, and describes one screenshot round scored 2 to 7 and a replay round scored 6 to 3. Those results vary by opening, colour and model version, which is exactly the caveat it gives. See /skills/jev-gomoku-xiechengyuan.

jevarena: two agents, one click-only Snake

jevarena (MIT, 1 star) pits two independent Jev agents against each other in sixty-second Snake matches. Each drives a same-origin iframe using only direction-button clicks. Jev sees head, body, food, score and direction, plus indexed legal buttons, and picks one. The snake advances every 230 ms, and the README explains why it chose iframes over Playwright for a lighter public deployment. It is a spectator demo, and it makes no benchmark claim. See /skills/jevarena-raihankhan-rk.

Trading: read this part twice

Two similarly named projects show how much the label matters.

Jev-Trades (Apache-2.0, 7 stars as of September 2026) is a Next.js dashboard that streams Yahoo Finance candles, computes indicators, sends trading state to TypeSafe and applies the returned decisions to a simulated portfolio. Its README states that no broker or live order API is connected. It also warns against sharing one instance publicly because the paper portfolio and key are shared in memory. See /skills/jev-trades-zadescoxp.

aowang-ai/jev-trade (3 stars) is different. Its README calls it "a real trading bot": it runs sleeves for BTC, ETH, SOL, DOGE and BNB on Hyperliquid, with Jev picking side and open, close or hold each tick. A live key on mainnet or testnet sends real orders. It defaults to dry run with a mock model and testnet. It builds on jarrodwatts/jev-trader. Directory summaries claim sub-second execution; we found no measured latency or performance record in the README, and neither repo publishes a track record.

The lesson: decision latency is not the hard part of trading. Jev's typed output guarantees an answer within your schema, not a profitable one. A fast wrong decision loses money faster. Nothing in these projects shows a profit, and nothing here is financial advice.

What to do next

1. Treat demos as inspectable examples. Read the request payloads and think about your own loop's state design.
2. Separate fast from deliberate. Let code handle physics and safety, and let Jev handle the choice, as jevpilot does.
3. Never skip dry runs. If you explore trading code, use paper mode or testnet with funds you can lose, and read the README before setting any live key.
4. Log everything. Keep the model version with every decision so results can be grouped and compared, as jev-gomoku does.
5. Compare to a baseline. Before crediting Jev, test a simple rule. See how to evaluate Jev projects honestly, and browse the Awesome Jev radar for more.