Ilmarinen selects a neural architecture -- its computational contract (arena), its primitive operations, and its width and depth -- for a dataset by minimizing a single description-length objective, with every complexity charge derived as a code length rather than a hand-set constant.