TokenShooter

LOCKED

Model not loaded

Pilot: language model

Health100

Ammo24

Monsters9

Last decision–

What the model perceives

Waiting for the first decision…

Tactical map what the robot knows

explored ray unexplored monster: ● watcher, ◆ seeker fireball fog of war

Decision one forward pass, no text generated

Latency–

Pace–

On format–

Mirror averaging (cancels left/right bias)

Timeline latest decisions

No decision yet.

Strategy part of the system prompt, cached on the GPU

Edit it live: the prompt cache is rebuilt in a fraction of a second.

Exact prompt sent to the model

–

Monsters

Watcher · floating eye. It lines up with the robot, its mouth glows, then it spits a fireball. Takes three hits.

Seeker · burning skull. Fast: it charges in and bites. One hit is enough.

They sleep until they see the robot or hear a gunshot, then hunt it down. Shots and fireballs fly straight along the rows and columns of the maze: to hit, you have to line up.

How it works

Sensors to text. The game describes what the robot perceives: free cells in each direction, the way to the nearest unexplored area, the crosshair, and for a monster off the line of fire, the move that lines it up.
One forward pass. Qwen3.5 runs on your GPU through WebGPU. The system prompt and worked examples are computed once and cached.
Read, don't generate. Jev-like, we read the probability of four words: Forward, Left, Right, Shoot. The situation and its mirror image are averaged.
Act. The most likely action wins; physically impossible moves are masked. Keyboard: ↑←→ Space.