For months, puzzle 42 of the Abstraction and Reasoning Corpus sat in the reject pile — close enough to tantalise, far enough to humble. This week the iterative solver landed it cleanly, and the trick was not a bigger model. It was better eyes.

Delta-based cell-level features ask a small, almost impertinent question of every cell on the grid: what changed, and by how much? That local delta, folded back into the feature pipeline, was apparently the missing hinge. A seq-to-seq hint stream carried over from the previous checkpoint did the rest.

One puzzle is one puzzle. The more interesting news is the mechanism.