The obvious way to build an AI trader is one big model that reads everything and decides. We built something narrower.
One measurement per seat
Every seat on the floor measures one thing. Bid-Ask Spread measures the spread. Wire Freshness checks how old the news is. Drawdown Control watches the fall from the high-water mark. If a seat has no number, it has no opinion, and it says so.
Seven departments
Research and Validation, Risk and Compliance, Trading and Execution, Portfolio Management, Data and Intelligence, Digital Assets, and Cross-Asset and Macro. Seventy-four seats in all. Fourteen of them can veto a plan on their own.
Why narrow is easier to check
A seat that measures one thing can be checked against that one thing. When a plan is tabled, you can read what each seat found and see which number moved it. One large model's reasoning is much harder to audit.
What it does not mean
More seats do not make the floor right more often. Its own validation harness has certified no trading signal so far. The seats make the argument visible. They do not make it profitable.
Seventy-four seats, one job each
Why our validation harness has certified nothing
Approve-first, in code, not in a setting
What the floor does on a bad day