The old shape was backwards: pick one interesting market on structure — liquidity, timing, momentum — then form a view on whether it's priced right. Too often the deep analysis landed on a well-priced market and ended in a pass. The screening funnel flips the order.
Two layers
Every run, a light model prices the ~40 most liquid non-sport markets in one cheap call — a rough fair probability per market, no web research, deliberately crude. The selector then adds a divergence term: where the screen disagrees hardest with the market price, *that's* where the one expensive deep analysis goes — web research, heavy reasoning, an actual betting decision.
A screening fair is a pointer, never a bet signal: the deep pipeline still builds its own number from scratch. But pointing the costly look at probable mispricing instead of structural interestingness raises the hit rate of every run — and the screens themselves feed the platform-wide mispricing feed in the Intelligence API.