The Council Protocol

One answer is testimony, not a verdict. This appendix is the copyable version of the room. Everything in quotes below is meant to be pasted straight into a chat, word for word if you want. Change the words once you understand why they're there, not before.

One rule sits over all of it: no costumes. Never tell a model to pretend it's a senior engineer or a skeptical CFO. You'll get the impression of an expert and less of the goods. Same real context to everyone, no acting, and let the native differences between the models do the work.

The one-model version

Most days it's just you and one witness. That's fine, as long as you remember what a witness is. You can trust it on the what: restating your problem, laying out options, doing the mechanical work fast. You cannot trust it on three things: its report of its own success, any confident claim with nothing under it, and its agreement with you. It wants to please you. That's the tell to watch.

The cross-examination still works with one witness on the stand. These are the solo moves:

• "What did I miss?"
• "What evidence would make us reject this?"
• "For every claim you just made, tell me where it came from. Label each one: documented fact, supported inference, hypothesis, or unverified."
• "Restate what I'm asking for and why I want it, in your own words." Read the restatement. The part that comes back wrong is the part you never actually said.

And whatever it tells you it verified, go look with your own eyes. One model checking its own work is not verification. Ever.

The three-model version

Bring in three when money or time is riding on the answer. Anything that changes what you build, what you sign, or what you spend deserves at least one fork in the road, and one model can't fork with itself. Same question, same context, to three different models, separately. Then read for real disagreement, not reworded agreement. Two answers saying the same thing in different clothes tell you nothing. Two answers standing on different ground tell you where the problem actually lives.

The five-model version

The full room is for foundational decisions. Architecture. Legal and compliance weight. The thing you'll be living with for years. Here you add roles, and a role is not a costume. A costume is "pretend to be a person." A role is "here is the job you're responsible for in this room." The Architect proposes something concrete, because critics swinging at fog produce nothing. The Red Team breaks it. The Designer asks if a real human can stand to use it. The Librarian demands sources. The Operator, the boring one everybody wants to skip, asks whether you can actually run this thing on a tired Tuesday. Keep the Operator. The unglamorous question is usually the one that saves the project.

Don't run five models on everyday questions. A bad council is worse than a single good model. You'll just crowd the room with the same confusion wearing different clothes and feel more confident about it.

The instructions, in order

Independence first. Each member answers blind, before anyone sees anyone. One confident wrong answer at the top will pull the whole panel into its gravity. Paste this to each model separately:

• "Answer this on your own. You have not seen anyone else's answer and you won't until yours is done. Give me your honest independent read, and say plainly where you are not sure."

Surface the disagreement. Collect the answers, then hand the stack to each model with this:

• "Here are independent answers to the same question. Do not blend them. List every place they genuinely disagree, what each side is claiming, and what each side would have to be right about. Ignore wording differences. I only want real forks."

Provenance before consensus. Agreement is cheap. Four models can inherit the same bad assumption and nod along. Every claim earns its place first:

• "For every claim still standing, name where it came from and label it: documented fact means you can point at the source right now. Supported inference means you can't point directly, but the reasoning runs from things you can point at. Hypothesis means it's a guess worth chasing. Unverified means it needs primary research. No claim enters the final answer without a label and a source. And for any claim we're getting comfortable with: what evidence would make us reject it?"

Synthesis. A synthesis is a verdict, not an average. The smooth gray paste that offends no model and commits to nothing is worthless:

• "Write the verdict using only claims that survived challenge and named their sources. Do not average. Separate what every member independently established from what only some did. Say which disagreements the evidence resolved, which are still open, and what got excluded and why."

The minority report. When three converge and one won't fold, the holdout is usually seeing something the others aren't:

• "If any member held out and would not be talked down, write its objection into the verdict as a live, named disagreement. Do not resolve on paper what was not resolved in the room."

Human signoff. The last instruction is for you, not the machines. Before you commit, stop reading the words, because reading well is the trap, and ask yourself the only question that matters:

• "Would I stake my name on where this came from?"

If there's any drag in your gut, any hesitation you can't name, don't sign. The room makes the answer survivable. Only you can make it yours.