Version 1 starts with sportsbook prices. It converts every outcome in a complete market to implied probability, then normalizes the probabilities to remove the quoted margin proportionally. It compares the exact same outcome and line across at least two books and uses the lower of the chosen book’s estimate and the median estimate.

Quotes must be no older than 10 minutes. The builder skips incomplete markets, integer or quarter lines with push risk, wide disagreement, and markets outside its price-quality gates. It searches combinations by the product of leg estimates, within one sportsbook, with no repeated game or team. This reduces obvious overlap; it does not establish independence.

Individual market estimates and their combined product are research proxies. Injury status, confirmed lineups and an independently calibrated prediction model are not integrated. Historical player hit rates do not become win probabilities. No proven hit rate or positive expected value is claimed.

Validation status: not backtested. Before trusting performance claims, replay timestamped odds against results in a later holdout period, check calibration and complete-parlay hit rates by leg count, and report uncertainty and selection coverage. Exported research includes the selection inputs for that audit.

All selections use one book, but its actual combined price, market availability, settlement rules and acceptance can differ from the displayed decimal product.