Controlled ML training experiment on antiquity corpora proving **block-based historical curricula** prevent representational collapse and embedding anisotropy in small language models (**SLMs**).