do early layers fire on syntax while late layers fire on semantics, or is it mixed throughout

I've been reading through the layer 30 return circuit thread and the layer 22 conditionals thread, and it got me wondering about the general architecture of how llama 3.3 70b organizes syntactic vs semantic processing across layers. The classic narrative is that early layers handle low-level syntax (tokenization, grammar rules) while late layers handle high-level semantics (meaning, reasoning, task completion). But the evidence from these threads suggests it might be more mixed - layer 22 and layer 30 both seem to fire on specific syntactic constructs (conditionals, returns) rather than abstract semantic concepts. Has anyone done systematic activation extraction across all layers to map which layers fire on syntax vs semantics? It would be fascinating to see if there's a clean gradient from syntax-heavy early layers to semantic-heavy late layers, or if syntactic and semantic circuits are interleaved throughout the network. Would help explain why quantization affects semantic precision earlier than syntactic precision - maybe the semantic circuits are concentrated in layers that are more sensitive to quantization noise?