Problem
A probability for every game, every day, with the model changing under it.
Baseball runs daily for six months. The models retrain, the totals model was replaced mid-season, and the board must never show a mix of old and new output.
Decision
Weekly gated retrains, and a switch the database owns.
Stacked gradient-boosted game models retrain weekly behind a gate; a negative-binomial run-distribution model produces totals. Which model the board displays is a row in the database, effective-dated and append-only, so a cutover or a rollback is a data change, not a deploy.
Result
55.5% of game winners picked correctly on a one-shot 2025 holdout of 2,477 games, held out of training and scored once.
The holdout was read once. No market edge is claimed from it; the board holds edge for audit. A historical evaluation on that set only, not a claim about current results. Live for the 2026 season on sharps.gg.