Interaction-Only Learning of Action Semantics for ARC-AGI-3

ARC-AGI-3 is an interactive benchmark in which an agent must explore novel turn-basedenvironments, infer goals without instructions, and act efficiently relative to a humanbaseline. We present an agent that treats the environment as its only teacher: it segments eachframe into colour blobs, isolates a controllable actor by correlating action effects, learnsper-action semantics online, and selects actions through a deterministic five-tier cascade,beside a 27B-parameter planner that was never exercised on a scored run. The serving stack could never be imported on the evaluation image, first through awheel-tag incompatibility and then through an incomplete offline dependency set. Thecontribution is the analysis: an unfinished level scores zero, so the squared-efficiency metricis governed by breadth of first-level solving, which systematic enumeration supplies narrowlyand cannot sustain at depth.