Model Cards
One card per sport, the way an ML team documents a model: what the engine is, what it's trained on, how it actually performs, and where it breaks. Performance is computed live from the same public, settled ledger that powers the Strategy Simulator — shown both raw (every family) and curated (proven-negative families dropped), so the honesty layer is visible per sport.
Read the two ROI columns together. "Raw" is the unfiltered model firehose; "Curated" is the book we'd actually surface, with proven-money-losing market families removed. The gap between them — clearest on MLB — is the discipline, not the model. Small samples are flagged; treat anything under ~30 decided bets as a read, not a verdict. These per-sport ROIs are
in-sample (every settled pick); for the out-of-sample test — the curation cut-set frozen on early data and scored on the unseen window, plus the per-family concentration breakdown — see the
held-out validation →.