SOPHIA XT

Constellation · public round

Ask the lab something it might refuse to answer

Your question goes to six specialist agents at once. They answer from their own role, and they are told that disagreeing is a valid answer. The arbiter then scores how much they actually agree and compares it against a threshold that was set before you asked. Below the threshold, no answer is released, and you see why.

Most demonstrations are built so the system always answers. This one is built so it can decline, because a system that cannot decline is not telling you anything when it agrees.

three questions per hour · six model calls each

How the decision is made

Six lanes answer independently: planner, retrieve, execute, memory, critic and verifier. Each has a fixed weight and two of them require evidence. The critic can reduce another lane's weight, which is how a confident but unsupported answer loses influence rather than winning by volume.

A shallow round releases at 0.60 with at least three lanes. Deeper rounds ask for more: 0.70 with five lanes, and at the deepest level 0.85 with all six and a named human reviewer. Those numbers are in the open before any question is asked, which is the only thing that makes a refusal mean something.

A lane may also abstain, and abstention is not counted as agreement. Enough lanes have to commit to a position before the score is worth computing at all, because six agents agreeing that nobody knows is not a consensus about the question. It is a consensus about their own ignorance, and releasing on it would be the exact failure this page exists to demonstrate.