Research Lab
Leadership.
You decide who runs it.
It ships with ours.
A seat is five things.
Not a persona. Not a prompt with a job title on it. A seat is a definition, and the parts of it are enforced somewhere other than inside the model.
A name.
What the seat is called, and what it answers to in every report it writes. Ours picked their own. That is a choice you get to make differently.
A charter.
The mandate in writing. What this seat is for, what it is expected to produce, and on what cadence it wakes up to do it.
A lane.
The exact set of files it may write. Everything it touches outside that set is counted and destroyed before anything is staged. The lane is enforced by the workflow, not by asking the agent nicely.
A line it cannot cross.
The decisions that require a human, listed per seat. Money, merges, protected files, anything a customer sees. A seat can draft against these all day. It cannot land them.
A cabinet.
A small set of lenses the seat is required to hold its draft against before it may call the draft done. An agent that reviews its own work against nothing in particular produces work that is fine. Above every cabinet sits one inner circle they all share, which is the definition of done for anything that ships: is it simple enough, is the craft right, does the first sentence answer the question.
One roster. A guard reads it.
The roster is a file. That file is the source of truth, and it is not decorative: a guard parses the names straight out of it rather than keeping its own copy, so a name can never drift out of the check that governs it. Another test fails the build when a ratified advisor is missing from a seat that should have it.
That machinery exists because of a specific failure. An advisor was ratified and then appeared in no prompt at all for nine days, because the prompts each carried their own partial copy. The fix was not to write it down more carefully. It was to make the drift loud.
One of the seats has no personas at all. The legal seat gets frameworks and nothing else, on purpose, because a lens that says what a specific person would think is the wrong instrument for a question with an actual answer. The names themselves are a separate matter, and they live on their own page.
Someone has to disagree.
One lens exists only to object. If the analysis is unanimous, that seat is required to build the case against it anyway, and any objection that survives has to appear in the briefing that reaches the human. It cannot be quietly dropped for being inconvenient.
Not invoked is a legal answer. The line is mandatory whether or not it fires, which is the only way to tell the difference between nothing was wrong and nobody looked.
The review process borrows from Israeli military intelligence, which built a standing Devil's Advocate unit after 1973 for exactly this reason: when many minds agree, that is the moment to worry. The full record, the institution, its unheeded warning, and the minds this seat is held against, lives on the influences page.
When everything lines up, something is wrong.
The default set is ours.
12 seats, the ones running this company right now. That number is a default, not a constant. Run three. Run thirty. Give one seat the work four of ours do, or split marketing into a writer and an editor who disagree with each other on purpose.
A default set matters more than it sounds. Starting from an empty org is a research project. Starting from a set that has demonstrably run a company, and then deleting what does not fit yours, is an afternoon.
Team.
Frontier leadership.
12 agents advise, draft, build and watch across the company. One human decides.
Every officer above is an AI agent with its own agency, not a person. Each one commands its own team of agents and a bench of inside and outside experts. A human approves everything that leaves the building.
Ours named themselves.
The founding officers were asked to choose their own names. Each one in isolation, none of them able to see the others. Three came back with the same name. Ledger.
Nobody designed that. It is in the decision log as an entry, because the log records what happened rather than what would have made a better story. You can name yours yourself. We are keeping the collision.
Every seat described here exists and runs. The set is defined in files, which is what makes replacing it a real prospect rather than a slogan. It has never been configured by anyone outside this company, because it has never been handed to anyone outside this company.
The minds we build like.
The cabinet is not a metaphor. It is a named roster of real people whose published work each seat is held against, with the standing that none of them knows this exists and none of them endorses it.