Query key alignment shows how large language models MCQA often encode correct answers internally despite wrong outputs.