We reverse-engineer the world, then build the ground where AI gets tested.
One model dreams the next version of a game. Another implements it for real: deterministic physics, verified mechanics, playable at 60 fps. What's left over is a place to test what an AI actually understands.
How the loop works
Three roles, run over and over until the build holds up.
A diffusion model dreams the next version of the game: one new mechanic, one hardware era, a bolder look. A coding model builds it for real, with deterministic physics and telemetry-verified mechanics, running at 60 fps or it doesn't count. Judges and discriminators verify the result, hunting for where it's faking it: similarity that isn't realism, realism that isn't gameplay. Then it runs again.
Measured, not assumed
Every claim above ties back to a number from the same engine.
Published work
From Diffusion Dreams to Playable Games poses target-faithful implementation as constrained program search: a dreamer, a judge, a discriminator, and a renderer, each analyzed as an optimization step, measured across three studies on one deterministic engine. Submitted to IAAI‑27, Deployed Applications track, currently under review.
Request the preprint →