Typed decision lab

LAYA × JEV

Five games, one contract. Each side is driven by a model answering typed questions over the same state — a choice and a noul, every step, no parsing. Pick an arena.

JEV API KEY

SNAKE →

Two seeded arenas, identical apples. A steering choice and a sprint noul per decision. Measures whether a model can hold a control loop together.

choice · HARD_LEFT / LEFT / STRAIGHT / RIGHT / HARD_RIGHT
noul · is sprinting safe?

KOMBAT →

One ring, two fighters, 90 seconds. Blocking is a reaction problem, so this punishes a slow decision loop far harder than the snake does.

choice · ADVANCE / RETREAT / PUNCH / KICK / BLOCK / JUMP
noul · commit to the heavy attack?

RUNNER →

Subway-Surfers-style endless run on a seeded track. It keeps speeding up, three lives each — fast-but-unsure against sure-but-slow.

noul · is there a barrier in each lane?
code picks the clearest lane

FLAPPY →

Laya's own showcase game. The model reads where the next gap is; code flies the bird. Pipes come faster every second, so reading speed decides it.

noul · is the next gap top / middle / bottom?
code flies to the likeliest one

TETRIS →

Code describes every landing spot in plain words; the model rates each one. The time per piece shrinks, so a slow answer eventually drops where it spawned.

noul · would this placement leave the stack clean?
one question per distinct spot