02Case study
Jev Arena
Write a fighter in plain English and let a model pilot it in real time.
- 230 msmedian decision latency
- ≈3.7 / sdecisions per fighter
- 2–3¢model cost of a 60-second fight
- $0.49model cost of the whole 30-fight ladder
The problem
LLMs think well and react slowly. By the time a chat model has written "I should raise my shield", the heavy shot has already landed.
Jev is a different kind of model: it reads a state and returns typed decisions with calibrated probabilities in roughly 70 to 500 ms by its vendor's numbers, fast enough to be a fighter's reflexes. Jev Arena asks: if a person writes the strategy and the model supplies the reflexes, how strong a fighter can one paragraph make?
How one decision works
The engine does the geometry and turns it into facts the model can read, such as enemy charging heavy shot: YES, fires in 0.4 s. Then one request asks three kinds of question in parallel: how to move, what to do with the weapons, and whether each reflex condition is true right now.
Every probability shows live in each fighter's "mind panel", so you can see why it did what it did.
Key design decisions
Two fight modes
Real-time never waits, so it measures speed and judgement; lockstep freezes every 0.2 s until both brains answer, so it measures judgement only.
Exact replays
One seeded RNG and only exact floating-point operations mean a fight is fully determined by its seed and the tick each decision landed on; verify re-simulates a replay and checks it.
A ladder you join with a pull request
CI smoke-tests the fighter with an offline brain and posts a results table on the PR; on merge a workflow re-runs the ladder with Jev, updates the leaderboard and republishes the site. Round robin up to 16 fighters, Swiss above.
Fair-play limits
Strategy up to 700 characters and at most 4 reflexes; every fighter gets the same ladder brain and prompt budget.
Swappable brains
Any System One model, chat LLMs through a JSON adapter, or an offline mock brain for CI and for people without a key.
Built to be streamed
A 1920×1080 broadcast view generates live commentary from the model's probabilities and drops straight into OBS; the UI also speaks Traditional Chinese.
Results
Season one's 30 real Jev fights all ended in a KO:
| # | Fighter | Elo | W-D-L | Style |
|---|---|---|---|---|
| 1 | Trickster | 1063 | 8-0-2 | circles, baits, dashes away |
| 2 | Zen | 1062 | 8-0-2 | one sentence, zero reflexes |
| 3 | Tortoise | 1039 | 7-0-3 | shoots from behind pillars |
| 4 | Berserker | 994 | 5-0-5 | rushes in, punches |
| 5 | Glass Cannon | 938 | 2-0-8 | heavy shots, paper armour |
| 6 | Sniper | 905 | 0-0-10 | long range |
The most interesting result: Zen is a single line, "Win the fight. Do whatever a wise, calm fighter would do in this exact moment.", with no reflexes at all. It finished 2nd, one Elo point behind the leader, while the carefully tuned Sniper went 0-10.
A decision is about 1,400 input tokens, roughly $0.00006; both fighters decide about 3.7 times a second at a 230 ms median latency.