Why this is not ChatGPT.

When you write “here is my situation, respond as if you were Machiavelli,” you give a general assistant a costume. It still samples the next token to be useful to you. It answers at once. It speaks in one voice. It will usually join you. Nobody in that chat has a conflicting brief, because there is only one brief: help the user.

Asking the same thread to “also be Marcus Aurelius and disagree” does not create a second mind. It is still one sampler, one next-token distribution, with a softer instruction piled on top. The costume can change. The incentives of the model do not.

Each turn of a session calls a chat model (today, OpenAI's GPT-5.6 Sol). The Council is not a different intelligence. It is a constrained room: a Chair, three to five seated advisors, a JSON contract, a validator in code, one retry, and a private notepad the user never sees.

On every turn the server does this, in order:

1. 01

   Seat the room

   Only the Chair and the advisors you seated exist. Anyone else the model invents is discarded.

2. 02

   Write the contract

   The model must answer as JSON: a private notes field, a list of messages with a speaker and a text, and on a verdict a commitment with a decision, an action, and a due date.

3. 03

   Sample

   One completion, with enough reasoning budget to argue rather than slogan. GPT-5.6 Sol thinks before it writes; the retry, if needed, samples again at full reasoning effort with a note that names the exact failure.

4. 04

   Validate

   Code checks the reply. Not the model grading itself in the same breath as writing.

5. 05

   Retry once

   If the check fails, the same call runs again with that failure note, at full reasoning effort — the retry is the last gate before you see the reply. Older chat models drop temperature instead.

6. 06

   Keep the notes, show the room

   Planning notes are stored on the session and stripped before they reach the browser. You see the Chair and the advisors. You do not see the scaffolding.

A costume prompt says “be Machiavelli.” The Council injects, for each seated person, a doctrine (what they actually think), a manner, a never, and a move (what they do with a situation). When two seated people have a known fight, that fight is named in the prompt: two incompatible actions, not two tones on the same advice. Two advisors reaching the same conclusion in different vocabulary is defined as a failure.

We do not search a library of their books on each turn. There is no retrieval step that pastes The Prince into the prompt. The brief is written from their positions in advance, then locked for that session. That is still not a hat. It is a small, stubborn set of instructions the model is not free to ignore — and the validator is there when it tries.

The Chair has no philosophy and, outside the verdict, no advice. The first reply is the Chair only, with two to four fact questions (what was said word for word, who knows what, since when, what happens if nothing changes). If the answers are still thin, the Chair holds the floor again. Advisors do not speak until those facts are on the table. Most assistants answer the moment you stop typing. That is the costume prompt. This is a hearing.

Unless the turn is a verdict, the Chair's last line must be a question. If it is not, the validator fails the reply and the room is asked to write it again.

Helpful models converge. If two or more advisors speak, the reply is invalid unless one of them names another seated advisor and marks opposition — “I disagree,” “That's wrong,” or in French “Je ne suis pas d'accord,” “C'est faux” — and attacks the action, not the tone. Naming someone without pushing back does not count. Same conclusion, different words, is a miss.

Two speeches with one disagreement is also a miss. The challenged advisor must answer. Then the rest of the table speaks: some join a reason with a because of their own, some diverge. A duel of two while the others sit silent is a miss. The last advisor must say what the room shares (a fact) and what it still disagrees on (two actions). Each position needs a because tied to something you actually said — a fact plus what it implies — not an order with no reason. The Chair does not interrupt; it speaks last, restates that map, and asks which side you are actually on. At least one advisor in that exchange must ask you a precise fact question — not how you feel.

Advisors must also use something you actually wrote: a number, a duration, a quoted phrase, a name. Wallpaper advice that could fit any life is rejected. A list of banned lines (“focus on what you can control,” “trust the process,” “great question,” and the rest) is scanned in code. Flattery is not a style choice. It is a failed output.

The session has a rhythm: questions, then positions, then a verdict. The verdict is one decision, one concrete action with a date, and an instruction to come back and say what happened. The Chair does not deliver a speech. You are meant to leave having chosen, then return with news — not to collect a better paragraph and close the tab.

If you address an advisor by name, only that advisor answers, and they hold their line. They may concede a detail. They may not fold because you pushed. That is in the prompt, and it is why a costume in ChatGPT feels like it is on your side: nothing in that product is rewarded for staying against you.

ChatGPT and Claude are built to be helpful. Helpful usually means joining you. The Council is built so that joining you is a bug: the validator will send the turn back.

One voice cannot argue with itself in earnest. A good decision comes from collision — several ways of reading the same facts — so you can see which reading you are actually aligned with. The costume prompt cannot sit Machiavelli across from Aurelius. This product can, and it will fail the turn if they secretly agree.

And the output is not an essay. It is a hearing that ends in an action. That is the product. The model is the instrument.

The Council does not run a private foundation model. It does not retrieve their complete works from a database on each message. When a verdict is delivered, the Chair writes the next morning and asks what happened, so the room can continue. It does not send a phone notification.

If those distinctions did not matter, a costume prompt would be enough. They matter because a decision is not a completion. It is a room that will not flatter you, will not skip the facts, and will not let two voices pretend they disagreed.