Many minds · one answer
One AI will answer you with total confidence and no way to show you how sure it was. AI Council puts up to six models from different makers in the same room, makes them answer independently, argue, and vote — and only replies once they agree. Joshua Sorell, operating as Lunar Ley Lines built it.
The idea
You choose two to six seats from eighteen models — GPT-5, Claude, Gemini, Grok, DeepSeek, Qwen, Llama, Kimi, Mistral and their smaller siblings. Every seat answers your question on its own first, without seeing anyone else's reply.
Then the rounds start. Each seat reads the others' answers, can change its mind, and votes for the answer it now thinks is best. If every seat picks the same one, the debate stops immediately. If a majority holds across two rounds, that's your verdict. If nothing settles, you still get the best available answer, clearly labelled as unsettled. The model in seat one acts as chair and writes the final answer from what survived.
Open Deliberation under any answer and you can read every round: who said what, who switched, who voted for whom, and how many credits the question used.
Choosing your seats
Seat count is the one dial you control. Every extra seat brings another independent first answer, another argument into the debate and another vote on the tally — and also another model to wait for and another charge on your balance. Two seats is a quick check; six is a full room.
Two models that answer separately and then challenge each other is already a long way from asking one. Consensus tends to arrive in the first round, the cost is small, and it's the right setting for a sanity check before you spend anything serious.
Enough spread that a weak answer has to survive several different backgrounds at once. This is where most of the useful disagreement shows up, and the debate usually runs a round or two before it settles.
The most collaboration available: six independent drafts, the widest mix of makers, and the best odds that someone objects before the chair writes the final answer. Slowest and most expensive per question, and worth it when the answer matters.
A practical pattern is to open with three seats, then re-run the questions that actually mattered with the full six. Since the conversation is saved, the follow-up keeps everything the council already said.
Accuracy
A hallucination is an invention delivered in a confident voice: a ruling that was never decided, a statistic nobody published, a command the software doesn't have. It is the most expensive thing a single model can do to you, and a council is built to resist it structurally rather than by hoping the model behaves.
Each seat writes its own answer before it sees anyone else's, so a fabricated citation or invented number shows up in one reply while the others produced something different. That mismatch is the cheapest, fastest detector there is.
Seats have to defend their own answer, then choose the best of what's left. Adopting a neighbour's made-up detail doesn't earn a vote, so the pressure in the room runs toward the claim that actually holds up.
Smaller models are poor at inventing detail they were never trained on and good at spotting a claim that doesn't match their own. One cheaper seat beside two heavyweights is a real check, not a spare body.
A Majority or No consensus verdict is the useful signal, not a nuisance. Don't take the summary at face value — ask a follow-up that names the split, ask the seats to show their working, or narrow the question until the council can be specific.
Treat a contested verdict as a research prompt. Something as plain as "three of you said X and two said Y — what would settle it?" usually comes back with the honest version of what is actually known, and tells you quickly whether the question was too broad to begin with. Where the stakes are real, verify the surviving claim against a source you can point to: the council narrows the field, it doesn't replace the check.
Why it's better
A lone model writes one fluent paragraph and you have nothing to weigh it against. In a council, that claim has to survive six sceptics who each produced a different answer. Fabricated citations, invented statistics and invented API methods are the first things a rival seat notices.
Each lab trains on different data with different guardrails, so each has its own blind spots. Seating models across vendors means one lab's habitual error no longer gets the last word — it has to beat five answers that came from elsewhere.
Models are good at telling you what you seem to want to hear, especially when you hint at a preferred answer. A peer model has no reason to flatter you. A contrary seat keeps the reply honest instead of agreeable.
A unanimous verdict and a three-versus-three split read the same when the answer is one paragraph. Here the verdict is stated out loud — Unanimous, Majority, or No consensus — so you know how much to lean on it.
Seats are told to vote for another seat when they genuinely agree, and to keep their own answer only when it is truly different. The tally reflects real agreement, not politeness or a model voting for itself.
Every round is saved with the conversation. Weeks later you can go back and see which seat first raised the objection that changed the answer — useful when a decision is questioned.
Where it helps
Under the hood
Seats run on live models served through OpenRouter. One question becomes one request per seat per round, and the whole answer is assembled only after the vote settles — the final text is not streamed word by word, because consensus has to be finished before it's worth showing you anything.
Every model carries a power tier from S down to E so you can mix a couple of heavyweights with cheaper seats instead of paying top rate six times, and public parameter counts are shown where the maker discloses them. Diagrams in an answer are drawn only when one actually appears, so nothing loads that you don't need.
Sign in and conversations are saved to your account, with follow-up questions keeping the earlier context. Guests can ask three questions before signing in; new accounts get three free ones too. Beyond that, questions are billed against a prepaid credit balance — bought as one-time packs or a monthly plan — and a 30% service fee applies to model usage, covering provider costs, hosting and payment processing.
Being straight with you
Agreement is not proof. Models share a lot of the same internet and often the same misconceptions. When six seats confidently repeat one popular error, the vote comes back unanimous and looks perfect. Consensus raises the odds; it does not retire your judgement.
Early answers can anchor the room. The first confident reply sets the frame, and later rounds sometimes drift toward it rather than testing it. Reading the deliberation is the antidote.
It is slower and costs more than asking one model, by design — you're paying for six opinions and a debate.
No live browsing, and no professional standing. Model knowledge goes only as far as its training, and a consensus answer is still not legal, medical or financial advice.
Pick your seats, ask something you'd rather not get wrong.
Open the council chamber