ARC-AGI-3: don't solve the puzzle — learn the game.

Where the frontier actually is

The frontier is defined by one stark fact and one crucial observation. The fact: humans solve 100% of environments while frontier AI scores below 1% (harness-free official scores of ~0.1-0.5%). The observation: every agent that actually works does informed SEARCH / EXPLORATION over the action space, not genuine world-MODELING or goal-inference. In the Jul-Aug 2025 Preview Challenge the top three all shared this shape -- 'StochasticGoose' (CNN+RL predicting which actions change the frame) at ~12.58%, 'Blind Squirrel' (a directed state graph built from observed frames) at ~6.71%, and Rudakov et al.'s training-free graph explorer in 3rd. As that paper puts it, the agent 'explores but never models': it records a directed graph of (state, action) -> next_state transitions but never induces a compact rule that explains them, and it never infers the goal -- it piggybacks on the environment's level-completion signal. Two of the four pillars (Modeling, Goal-Setting) are essentially unattempted by the systems that lead the board.

Two structural constraints explain why search dominates -- and why it plateaus. First, the interaction limit: because every step of an LLM agent is gated by a model call, it manages only ~4,000 usable interactions against the ~96,000 the environment allows, so a cheap deterministic explorer wins simply by spending the entire budget without forgetting where it has been. This is the argument for using cheap acting to gather data and reserving expensive computation for offline model-building. Second, the harness-overfitting problem: hand-built harnesses (the Duke LRM harness, Symbolica's 'Aregentica') solve all three public environments at human-comparable action counts, but the same environment can go from 0% harness-free to ~97% with a hand-built harness, and none of it transfers to the private sets -- which is exactly why the official leaderboard reports only harness-free general systems. Pure coverage-search then hits its own ceiling: exploration cost grows linearly with the reachable state space, so it degrades sharply where the graph explodes (ls20 L3+, ft09 L6+) and where goals require compositional/relational understanding rather than reaching a specific hashed state.

The main open gap is world-model synthesis + planning, paired with goal induction -- the Modeling and Goal-Setting pillars that no deployed agent yet attempts. The concrete direction is to turn the recorded transition graph into training data for an explicit program next_frame = f(frame, action), fit incrementally from a few dozen actions (active learning / optimal experiment design over which action to try next, since scoring caps actions at 5x human via RHAE), accept the program only when it exactly replays every recorded transition, and then plan lookahead inside it -- with a verifier-in-the-loop that rejects the model the moment a predicted transition disagrees with reality. In parallel, goal/reward induction must supply the feedback signal the 2025 'refinement loop' lost when it left static demonstration pairs behind: forming and revising hypotheses about desirable states from sparse, delayed win-frame feedback, aided by intrinsic-motivation / empowerment objectives. ARC-AGI-3's determinism, small action space, and turn-based frames make such exact, verifiable models tractable -- this record -> synthesize -> replay-verify -> plan loop is precisely the north-star this repo is built around, and precisely what the current search-only frontier is missing.