Model Calibration

We publish our calibration. For every market family, this is what the model predicted versus what actually happened across every settled bet — straight from the results ledger, no synthetic data. A positive gap means the model was overconfident. Families proven to lose money are curated out of the boards entirely; well-sampled families have their displayed probabilities calibrated toward the realized rate. This is the engine behind the High Confidence Board.

Curated-out families split two ways: overconfident — the model's probability itself is broken (realized far below predicted), so the edge is fake at any price; and priced-short — the probability is roughly right but the bet loses at fair odds (it could still be +EV at a soft book line). The distinction is why we curate by ROI, not hit rate.

Curation runs on every pipeline cycle, so a family that turns into a proven loser is dropped automatically — the book can't quietly accumulate dead weight. The converse holds too: no family that survives the cut is a persistent loser (one that bleeds across both halves of the settled record). What's left is the clean subset — the losers are removed, not hidden.