Jev vs LLMs

JEVVSLLMs

45.0 Tick Race

Model	p50	In time	Dist	Crash	Cost	$/100	Stale	Err

Ranking = decision speed. Median latency first, then the share of decisions that landed in time. Distance and crashes are context, never the tiebreak.
One tick, one decision. Every tick each model gets its lane and a forecast of the next 5 ticks. If the answer arrives after the tick ends — even by 1 ms — it is thrown away: the car holds its lane, the request is still billed, and the board counts it as Stale.
At the default 600 ms tick, a model averaging above 600 ms effectively cannot drive.

Each model drives its own copy of the same 8-lane highway: same seed, same 32 vehicles, same moves available.
Every lane flows at its own speed — trucks and vans in the slow lanes, bikes and cars in the fast ones — and four vehicles swap lanes at random. The sim guarantees there is always a survivable move.
Two ways to crash: a vehicle reaches your row, or you change into a lane where one is already beside you. The state warns about both (clear_ticks = 0 means "occupied right now").

p50 is the model's own latency, measured at the server. Your network hop to this page is not counted.
Cost is the provider's reported charge per request; $/100 projects it to 100 decisions.
Charts plot every decision against the tick budget (dashed line) and split it into three measured layers: model thinking (the provider's own timer), server↔provider network, and browser↔server — your connection to this page. Only the first two count for ranking; the third is yours.