First, the plumbing. The Federal Open Market Committee - the FOMC - is the body inside the US Federal Reserve that sets the policy interest rate. It meets on a published calendar, releases a short statement the same afternoon, and its decision reprices the cost of money worldwide. Nineteen people take part; twelve of them vote. Section 02 explains why those two numbers differ.
For two decades the Committee told the market a great deal about its own thinking. It published forward guidance - a sentence in the statement describing what it expected to do next. Four times a year it published the Summary of Economic Projections, or SEP, whose best-known page is the dot plot: one anonymous dot for each participant showing where they think the policy rate should end each year. The Chair then spent an hour at a press conference being asked to interpret both.
In 2026 all four of those channels narrowed at once, under a new Chairman. This is documented in the Fed's own published statements, not inferred. hard data
None of this is a scandal, and we do not present it as one. A central bank is entitled to talk less, and there is a serious argument - the Chairman has made it in public - that guidance which the Committee cannot honour is worse than silence. But it changes the economics of research. When the institution hands you its reaction function, modelling it is redundant. When it does not, the reaction function becomes the thing worth estimating. judgement
The Committee has two kinds of member. The seven governors sit on the Federal Reserve Board in Washington and vote at every meeting by law. The twelve Reserve Bank presidents run the regional banks; the New York president votes permanently, and four of the other eleven vote on an annual rotation. All nineteen attend, all nineteen speak, and all nineteen are simulated here - because the seven who cannot vote in a given year still argue, still publish, and still move the people who can.
A dissent is a recorded vote against the Committee's decision. The statement names the dissenter and says what they preferred. It is a costly, public act: most meetings are unanimous or carry a single dissent, so a run of three in one direction is a genuine event rather than noise. hard data
Kevin Warsh was sworn in as 17th Chairman of the Board of Governors of the Federal Reserve System on 2026-05-22. The corpus behind his persona holds 54 dated documents and 168,085 words as at 2026-08-28 - every one of them his own writing or speech, with a date and a source. The frozen persona used for the September prediction was distilled on 1 August from 53 of them; the 54th is the Jackson Hole keynote of 28 August, which is what section 08 checks the persona against. hard data
He has not once described financial conditions as restrictive - not in June, not in August. Any model that has him waiting for the bond market to deliver tightening on his behalf has taken a sentence about a six-week change and promoted it into a judgement about the level. judgement
We know this failure mode intimately, because it is the one our own model committed. Section 08 is that story.
Before every meeting the Fed's staff circulate a briefing book. We cannot see it, so we build a shadow of it: a fixed list of 26 public series covering the two halves of the Fed's legal mandate - stable prices and maximum employment - plus activity, plus the financial conditions the Committee argues about. Each simulated member reads the same packet. Differences in their conclusions therefore come from the people, not from the data.
A word that will keep coming back: a vintage is the value of a statistic as it stood on a particular day, before later revisions. Statistical agencies revise heavily. A model that replays June 2026 using today's corrected numbers is being handed information nobody had. hard data
The four stages of a simulated meeting mirror the real sequence described in section 02. The one place we impose a hard rule rather than imitate is the memo stage.
The second design choice is where the probability comes from. Language models can be asked the same question many times and will answer slightly differently; averaging those answers produces a number that looks like a probability and is really a measurement of the model's own jitter. We do not do that. Instead the uncertainty is put where it belongs - on the state of the world.
We can tell you, with a transcript to back it, how this committee behaves if inflation re-accelerates. We cannot tell you the chance that inflation re-accelerates - nobody can, and a model that pretends to is selling you its own noise.
So the two are kept apart: the behaviour is simulated and fixed; the weight is yours to set. judgement
The first time this machine was checked against a real document it had not seen. The Warsh persona card was distilled on 2026-08-01 from 53 historical documents; the Jackson Hole keynote did not exist until 2026-08-28, so the model could not have read it. That makes the check half-clean. It is not a rate forecast score - the September meeting is what produces one of those. What is compared here is the model's baseline chair proposal and memos against the delivered text. hard data
The engineering consequence has already been made. The data pack carried only the first kind of gauge, which is the same one he reads - so the model had no way to tell the two apart. A 10-year real yield series and a 10-year term premium series were added, taking the pack from 24 series to 26. That is the change behind Figure 13. What has deliberately not changed is the prediction: The record for the September meeting is already closed. Editing it after new information arrives means it was never a prediction. hard data
Everything the committee does inside a scenario is fixed - it was simulated once, end to end, and stored. Drag a slider and the headline distribution recomputes; the rooms do not move. Click any member to read the memo they wrote before the meeting, what they wanted, what they voted, and what would have made them dissent.
Open the Room - the archive reads instantly and costs nothing; the two things that call a model are opt-in and are described below.
| You can | What it does | What it costs |
|---|---|---|
| Look | Open any run - the official batch or one somebody else ran - drag the scenario weights, read all nineteen memos, browse the archive and the ledger. | No model call at all. Instant, and it works with the service switched off. |
| Ask | Put a data print, a headline or a question in front of one member and get their reaction in character, from the same persona card the memo came from. The answer is not added to the run. | One model call, about thirty seconds, well under a tenth of a yuan. |
| Run | Describe a scenario in plain sentences, optionally override one number in the data packet, and run the whole committee: nineteen memos, the Chair's proposal, twelve votes. The result becomes a room with a link you can send. | 32 model calls, about 6.5 minutes queued, roughly ¥1.5 on deepseek-v4-pro - the calibrated figure, which is about twice the rate card. |
When the ledger resolves we will publish the outcome whether or not it flatters the model, together with both benchmark dates, the two checks in section 08, and the three residual leaks in Figure 25. A score reported without its contamination boundary is not a score. judgement