Discover
Every PROMISING or VALIDATED finding — from a manual run, the miner or a continuous campaign — ranked and filterable. Demo-data findings are labelled and are not evidence about real markets.
Screening ranks by the raw evidence rather than the corrected p. Past a certain search size the correction saturates — a 1,000-draw permutation cannot report a p below 1/1001, and Benjamini-Hochberg scales that by roughly the search size over the batch size, so beyond about 15,000 candidates nothing can be VALIDATED however real it is. Screening keeps every out-of-sample gate and stays comparable across a search of any size; it is a ranking for deciding what to confirm on fresh data, never a claim that anything is confirmed.
One row per rule. A campaign tests the same rule across many instruments, horizons, target metrics and event gaps, so the ledger holds far more rows than there are distinct findings. Direction is the share of a rule's variants whose effect points the same way — a rule near 50% disagrees with itself across the markets it survived on, which is what a large search produces from noise.