This manuscript examines how candidate tokens activated in earlier transformer layers influence subsequent token predictions and explores decoding informed by those dependencies.