We’re a research studio building environments to test AI. We take existing systems apart, rebuild them in code, and change how they work. We want to know whether an agent understands the world it’s in, or has just learned its way around one version.
Games are where we started. They have rules to recover, decisions to make, and plenty of room to experiment. These are recordings of the software running.
We reconstructed Re-Volt and added a car made from photographs of a real toy. Once it’s in the game, it has to drive, collide, and make it around the track.
What else could we bring into this world?
Codex, Claude, Gemini, and Grok reconstruct the same game. Put the running versions side by side and the differences become visible, from the interface to what happens on the farm.
Excerpt from the four-model evaluation, recorded from the running reconstructions.
Fields, livestock, machinery, and the occasional disaster. A reconstructed SimFarm with a new set of graphics, right down to the crops and tools.
Our Tom & Jerry reconstruction running on an RG Rotate. The original-style and claymation versions share a selection screen and physical controls. This edit comes from a phone recording of the device.
Watch the claymation gameplayIndependent research projects. Original games and characters belong to their respective owners.
Once a world runs in code, we can alter it. Add an object. Change a rule. Make a familiar task harder. Each version gives us another way to test what an AI has learned.
Recover a working world.
Give it something new.
Put an agent inside it.
Find out what held up.
Study the original. Work out what each action does, how state changes, and where the constraints are. Rebuild it in code, then run it against the reference to find what we missed.
ORIGINAL → RUNNING RECONSTRUCTIONA new object, a different visual style, or a rule the original never had. Generative models help describe the next version; coding models make it run. We check the result before using it as the starting point for another change.
RUNNING WORLD → VERIFIED VARIATIONThe research question is whether an agent can handle a world that has changed. Give it a task in both versions. Compare its decisions, its mistakes, and how it recovers. Use those results to choose what to change next.
AGENT BEHAVIOR → NEXT EXPERIMENTTHE IDEA BEHIND THE PAPER
A diffusion model imagines the next version of a world. A coding model builds it. We run the result, compare it with the target, and work through the gaps. Once it passes its checks, a fresh capture becomes the starting point for the next dream.
That’s the Dream Loop. Our paper studies where it works, where visual similarity is misleading, and why a convincing image is only part of the job.
Request the paper
The dream proposes a change. The implementation has to make it work. The next cycle starts from what actually ran.
Conceptual illustration, generated with AI. Gameplay recordings above show the actual projects.
Two patent projects sit behind this research: turning generated targets into executable worlds, and keeping later versions on course as those worlds change.
Turning a generated visual target into executable code and assets, then building the next version from a checked, running result.
US provisional application filed, September 2026.
Using both a long-term visual goal and the current implementation to guide the next change, with checks on local progress and the larger direction.
Separate provisional filing package prepared.