Released after
the meeting.
the meeting.
SAPIENSQ · MIRROR REAL ONES · PRODUCT 1
We don't forecast rates.
We simulate the committees that decide them.
Built for trading desks and risk teams that position ahead of the Fed. Each meeting is reconstructed from the public record: named members, rotating seats, prior debate. The model returns the room's next decision. FOMC is our first case.
The April 29 call
Locked before the meeting.
The FOMC statement releases on April 29, 2026 at 14:00 ET. Until then, this card does not change. The actual outcome appears beside it once the Fed publishes.
What's inside
Three parts. One room.
Each part is visible in the deliberation transcripts. Request access to see it turn by turn.
01 · Persona
Who's in the room, correctly.
Each member is represented from their own documented posture: statements, votes, speeches. The seat list is refreshed per meeting, so rotations and new appointments show up where they should.
Where to look. Each member's call in the transcript traces back to the speeches and votes it was built from.
02 · Memory
No look-ahead, ever.
Every document carries a release date. The briefing for a given meeting loads only what was public before it: Beige Books, Minutes, prior Statements, Transcripts. Later meetings do not leak backward.
Where to look. Citation tags in every deliberation turn link back to the source document and its release date.
03 · Deliberation
A structured deliberation, end to end.
The room runs in four phases: briefing, debate, a draft proposal, a formal vote. The draft anchors on the prevailing policy regime, so the starting point matches the period rather than a generic baseline.
Where to look. In every deliberation, the draft's reasoning chain shows turn by turn. Request access to read it.
Recent calls
What the Committee decided, and what we called.
Recent meetings where the call matched on direction and magnitude.
N=3 reproducibility · Gemini 3 Flash
Run it three times. Read what stays.
Three independent reruns over the same 17 meetings. Means below; per-run breakdowns on request.
94.1%
Direction AR
vote-direction accuracy (CUT / HOLD / HIKE)
4.07
Statement LLM judge
3-judge ensemble (Claude · Gemini · GPT), 1–5 rubric
3.37
Minutes LLM judge
3-judge ensemble (Claude · Gemini · GPT), 1–5 rubric
83.6%
Vote AR
per-member vote alignment with the real FOMC
0.820
Statement embedding cos
semantic similarity to real FOMC Statement
0.727
Minutes embedding cos
semantic similarity to real FOMC Minutes