Extract real neural attention from a causal LM, compute cosine similarity between centered attention signatures, and auto-generate LLM semantic explanations for retained pairs.