We did not build a bigger model. We put six small, different ones in front of the same question.
A chat like any other: a white box, a blinking cursor, the question written the way you would write it to a friend. Nothing to learn.
The moment you press enter, the question leaves for six labs in six countries. Each answers on its own, without seeing the others' work. They don't talk, don't copy, don't agree in advance.
A program receives the six answers, lines them up and measures how much they resemble each other. It does not understand the question. It has no opinion. It counts.
At the bottom of every answer you find the consensus. Ninety-four percent means almost all of them say the same thing. Below sixty we tell you to check for yourself, and we say it in large type.
A single model is confidently wrong. It has nobody beside it that is wrong in a different way. Six models trained in six places, on different data, by different people, have errors that don't overlap: where one invents, the other five usually don't.
That is why we don't use one large general model. After the Global Binding Protocol the large models became cautious to the point of uselessness: they answer everything with a care that helps nobody. Six small labs answer. And the moderator keeps count.
Below sixty percent we tell you to check. We never round up.
Every question is a new question. We don't build a profile of who is asking.
None of the six models speaks to you directly. You only see what they agree on.
It helps you ask the right question to the right person. With a percentage in hand.