You Don't Need More Models. You Need the Argument That Fits the Question.

Everyone is still litigating whether to trust one model or a council. That fork stopped being the interesting one a while ago.

The interesting fork comes after you've decided to convene one. Because "use a council" is not a decision. It's the word "meeting" — it tells you a room will be full and nothing about whether anything useful happens inside it. The format you pick once the models are in the room is the whole game.

The same machinery, opposite results

The field spent the last month proving that structure matters. A debate step bolted onto a model and left ungoverned barely moves the number. A structured one moves it a lot. Fine. But that finding quietly hands you a harder question and walks off: structured how?

The data refuses to give one answer. DeliberationBench ran the comparison and found that simply picking the best single model beats every deliberation protocol it tested — 82.5% to 13.8%. Read that and you'd retire the council on the spot. Then look at FinCom, a financial committee that forced each agent to either name a flaw or commit with new evidence before speaking. On risk analysis, accuracy jumped from 59.3% to 90.5%. On short benchmark questions, the same machinery dropped it — 66.0% to 58.7%.

Same mechanism. Opposite outcomes. The variable wasn't the council. It was the question.

A council isn't one thing

Here's the line the verifier debate keeps skipping: more models isn't the upgrade. The right argument is.

A council isn't a switch you flip. It's a menu of arguments, and the shape of your question picks the dish. A factual lookup and a "should we ship this Friday" call are not the same problem wearing different clothes. They want different rooms, different rules, different exits. Treat them the same and you get DeliberationBench's result on one and FinCom's on the other — and conclude, wrongly, that councils either work or they don't.

So stop asking whether to use one. Ask which argument the question needs.

The decision logic, strategy by strategy

When one answer is enough, don't convene anything. A definition, a conversion, a fact with a knowable answer — that's a Quick Take. This is the case the skeptics are right about, and it's most cases. FinCom dropping on short questions is the same lesson: structure is overhead, and overhead you don't need is just cost. The discipline is admitting which questions these are. Ask a council for the capital of Australia and you've staged theater.

When the question is broad and open, put everyone in the room. "What are we missing about this market?" wants a Traditional Council — several models, full context, talking. You're not narrowing yet. You're surfacing the spread.

When you want each model to sharpen the last one's answer in turn, that's Round Robin — a sequential build instead of a free-for-all, so a good idea gets refined rather than shouted over.

When you have a field of options and need it cut down, run Survivor. Five candidate strategies, eliminate the weak ones round by round. The argument has a job: not to explore, but to prune.

When you're building one artifact instead of choosing among many, use Collaborative Editing. A contract clause, a function, a paragraph — the models pass it back and forth and improve it in place. The output is the document, not a verdict.

When the decision is expensive to get wrong, force the fight. Red Team vs Blue Team exists for the calls where a confident wrong answer costs real money. One side attacks, one defends, and the disagreement you'd otherwise average away gets dragged into the open where you can see it. This is the strategy that catches the model that would have invented an answer — because the other side is paid to find it.

When the models genuinely disagree and someone has to decide, that's Chairperson Synthesis. Not majority vote, which just launders a confident wrong answer into consensus. A chairman that reads the dissent and decides what to do with it. FinCom's "disagree or commit" is this instinct named — commit to an action without erasing the objection that survived.

Notice what the choice is actually keyed on. Not how smart the models are; they're roughly the same models in every room. It's the shape of the question — open or narrowing, one artifact or many options, cheap to miss or expensive, settled or contested. Match the format to that shape and the council earns its cost. Mismatch it and you've reproduced DeliberationBench: more voices, worse answer, and a tidy story about why councils don't work.

The part worth picking carefully

The fight over whether to trust one AI or several was always a little beside the point. The model you're going to use is about as good as everyone else's. What you control is the argument you make it have — whether you have one at all, and what its rules are when you do.

One model gives you an answer. The wrong council gives you a slower answer. The right one gives you the answer plus the reason you can trust it.

That's the part worth picking carefully.

Try it free — no signup. shingik.ai