Playing ARC-AGI-3
Pure symbolic prediction against the hardest interactive reasoning benchmark in AI. No LLM. No GPU. No fine-tuning. Observe → predict → act. Against the live engine it completed 24 of 25 games and 176 levels — one game is still standing. The gym numbers below are the offline harness, kept separate on purpose.
Gym Benchmark
1,000 procedurally generated games · 5 core-knowledge families · 3 difficulty levels each · 0.6 seconds total runtime
Measured on internal gym (arc3_gym.py) — procedural puzzles matching ARC-3 protocol. Not the official ARC-AGI-3 leaderboard. Official eval pending benchmark release.
Family Breakdown
Agent→goal shortest path. Walls avoided. Computed in microseconds.
BFS on (agent, block) state space. Finds minimum-move push sequence.
Read target half → copy to work region. 16 clicks. Zero guessing.
Lights Out → linear algebra. Gaussian elimination over GF(2). 2–3 clicks.
Detect filled half → mirror column-by-column. 8 clicks.
The Paradigm
GRADIENT CHASING (before)
Click a colour. Did the score go up? No? Try another. Blind exploration exhausts the action budget. Navigate scraped to 52.5%. Everything else: 0%.
THALAMIC PREDICTION (after)
Read the grid. Detect the puzzle type. Compute the exact solution. Execute in minimum actions. Zero guessing. Every move is predetermined.
The shift: stop treating the environment as a black box. The answer is in the grid. Read it, compute it, paint it. Prediction over exploration. Measurement over iteration.
Solve Pipeline
Read the raw grid. No preprocessing, no embeddings. Just the matrix of cell values.
Universal family detection from grid structure alone. No labels, no hints from the environment.
Compute the exact solution using the matched algorithm: BFS, GF(2), pattern-copy, or state-space search.
Emit the minimum action sequence. Every click and move is predetermined before the first action fires.
Universal Family Detection
No labels from the environment. No hints. The classifier reads grid structure alone and routes to the correct solver.
Architecture Components
Thalamic prediction policy. Observe → detect → solve → execute. Zero gradient chasing. Plans the full solution before the first action.
Procedural game generator. Speaks exact ARC-3 FrameData/GameAction protocol. 5 families, graded 0–254 scoring, solvability-verified.
Universal classifier. Reads grid structure—separator columns, cell value distributions, entity counts, half-density. No family labels used.
BFS pathfinding, composite (agent,block) BFS, target-copy, GF(2) Gaussian elimination, mirror-copy. Each solver is exact—zero trial-and-error.