Two architectures, one board.

Two ways to make a decision.

Autoregression commits one move per pass, and each move constrains the next. Diffusion evaluates the whole board in a single pass, which is the difference the hardware feels.

Battleship.

Agentic decision under uncertainty
Autoregression · one move per pass
Shots 0 Hits 0 Sunk 0/5
Diffusion · whole board per pass
Shots 0 Hits 0 Sunk 0/5

Sudoku.

Constraint satisfaction
Autoregression · one move per pass
Filled 0 Errors 0
Diffusion · whole board per pass
Filled 0 Errors 0 Passes 0

Both sides run the same puzzle. The heatmap is the posterior over the whole board.