// Why Council

You double-check with the same model.

Ask one AI a hard question and you get one confident answer. Ask again and you get the same answer in different words. Confidence is not the same thing as being right, and a single model cannot tell you which one you are looking at.

// What is missing right now

Four blind spots.

One voice, no dissent

A single model hands you its best guess in a settled tone. You never see the reasonable case against it, because nothing in the system is trying to make that case.

It agrees with itself

Asking a model to check its own work mostly produces agreement. It shares its own blind spots, so it cannot audit them.

Confidence is flat

The same certain tone covers a settled fact and a coin flip. You cannot tell them apart from the writing.

Facts and judgments get one treatment

“What year did that happen” and “should I leave my job” are different kinds of question. Most tools answer both in the same voice.

Why disagreement is the useful part

Most multi-model tools combine AIs to give you more answers. Council combines them to show you where they clash. When three models trained differently land on the same conclusion, that agreement means something.

We do not smooth that into a tidy consensus. The verdict arrives with the disagreement visible underneath it.

“When they split, the split is the most valuable thing on the page.”

Verification, and its limits

On Plus, factual claims inside the verdict will be checked against an outside source; this is not switched on in the free preview yet. Three AIs agreeing is still three guesses that happen to match. Only an external check makes “is this actually right” an answerable question.

One model is a guess. Three models and a chairman is a better reasoned guess. Outside verification is the first point where correctness becomes checkable at all, and only for the parts that can be checked.

The research behind it

Comparing several answers to catch a wrong one is an established technique, not something we invented. It was tested properly at the CONVERSATIONS 2024 workshop on human-centred chatbots.

What that work found, and we would rather say it ourselves: it helps, but it is not magic. Accuracy swung widely depending on the subject, and it worked far better on a narrow domain than as a general truth detector. That is why Council sticks to decisions.

We do not build models. We call them through their providers’ APIs and put them into a structured argument. What we built is the argument.

Try it on a decision you are stuck on.