Can you win a game without knowing the rules?

ARC-AGI-3 starts without a manual. The agent sees an unfamiliar interactive world, chooses an action, and receives the next state. There is no description of the controls, no stated objective, and no guarantee that the first plausible interpretation is the right one. A capable language model can still wander here because recognizing a scene is different from learning the rules that make it move.

Our system treats every game as a small scientific problem. It forms a hypothesis about the environment, turns that hypothesis into something it can execute, and tests it before spending another real action. When reality disagrees, the disagreement is not discarded as a bad sample. It becomes the evidence used to change the model.

The executable model begins incomplete. Early actions identify which objects respond, which changes persist, and which apparent regularities were coincidences. The controller then searches inside the updated model for a short route forward. Better models produce more informative actions; more informative actions produce better models.

Each hypothesis is translated into runnable code and checked against the transitions already observed. A mismatch blocks further action until the account of the world changes. Once a usable model exists, the next real move is chosen by planning through it rather than by guessing directly from the screen.

The two trajectories in this note are complete runs from that process. The opening lp85 recording completed all eight levels in 119 actions; cd82 completed all six in 130. The case below makes the learning process inside one of those trajectories concrete.

cd82 is a 10×10 stamp-painting puzzle. The image in the upper left is the target; the lower grid is an editable canvas. Palette clicks change the paint color, moving the large stencil changes the footprint it will cover, and stamping overwrites the cells beneath it. A level ends when the canvas matches the target exactly.

The agent began without any of those facts. Level 1 established the target, palette, and basic half-plane stamps. Level 2 forced it to discover a missing lower-right diagonal pose. Level 3 introduced a smaller secondary stamp; level 5 revealed the symmetric lower-left pose. By level 6, the world supplied no new mechanic—the shared planner could compose the learned footprints and finish from the model it had built.

Across broader public-game runs completed on our machines, the system reaches roughly 85% overall completion. ARC Prize lists GPT-5.6 Sol Max at 7.8% on the verified semi-private evaluation. Numerically, our local result is 77.2 percentage points higher and 10.9 times as large.

The evaluations are not identical. Ours uses an executable-world-model harness on public games and allows the system to learn through interaction; Sol Max is a verified frontier model result on ARC Prize's semi-private set. This is therefore a comparison between complete systems under different tracks, not a claim that our base model is intrinsically 10.9 times stronger.

The useful question is narrower: what changes when a model can build, test, and revise an account of the environment while the task is unfolding? On our local public-game runs, that mechanism changes the scale of the result. Verification on the semi-private track remains a separate test.

The best internal model is not the most elaborate one. It is the smallest model that makes a useful distinction, exposes its own failure, and can be revised before the action budget disappears. In these games, testability mattered more than eloquence and targeted experiments mattered more than long unconstrained exploration.

This is why we describe the system as continually learning. It does not receive one training phase and then freeze into a polished answer machine. It repeatedly converts surprise into a new executable belief while the task is still unfolding.

The next step is harder: carry useful abstractions from one unfamiliar world into another without carrying over brittle assumptions. We want the system to begin each game less naive, while remaining willing to discard everything the new world disproves.