Eight signal sources vote on live market data. A Bull and a Bear argue the evidence; a Judge weighs it with weights earned from each source's own track record. A separate Risk Sizer sets the stake — direction never sizes its own position. Every call lands in a Brier-scored calibration log.
Direction, sizing, and approval live on different agents on purpose — the agent that wants the trade never decides how big it is, and the sizer never overrides the gate.
Each source's rolling hit-rate is tracked in the calibration log below. Sources that earn their seat get amplified; sources that don't get muted. The weights you see here adapt from resolved calls in your browser's log.
Not "BUY 73%" from a black box — a structured argument built from the same evidence the sources voted on, with the losing case preserved in full.
Every run files a probability call. One hour later it resolves against the actual price and is Brier-scored — 0.0 is prophecy, 0.25 is a coin flip. Per-source hits and misses feed straight back into the fusion weights. Real money stays gated behind a written Brier improvement, not vibes.
The terminal is how the swarm sees markets. The intelligence layer is how it reads the world. The strategy builder is how it encodes a thesis. The prediction terminal is how it checks its beliefs against people betting real money.