The Prediction Lineup

The prediction layer is not one model with five names. It is eight engines — independent estimators built on deliberately separate information bases, so that when they agree the agreement means something, and when they disagree, we can say why.

One reads the cross-section of strikes. One reads only the clock. One reads only the volatility surface. One is a supervised learner over a pre-registered feature roster. One reads the arrival process of the tape itself. One combines the others and may never ship. One decides how much any of it deserves to be trusted. And one does not predict at all — it explains, under a firewall that forbids it from computing. Every engine states what it will not do, and an engine may abstain rather than fabricate.

Built and wired is not the same as running: every engine sits behind its own arm switch, off by default, and the page says which stages of the chain are armed and which are dark rather than letting existing code imply a live product.

Built dark and measured offline on stored sessions: no live predictive performance is claimed anywhere on the page. Derived analytics for education and research, not investment advice.