Model governance
How each sport earns trust
Our goal is not to publish the most picks. It is to build a separate repeatable process for each sport, measure it honestly, and improve it without rewriting history.
Current status
Sports Rounders uses the newest implemented model introduced by the owner as the active champion unless it is explicitly marked experimental, shadow, challenger, or pending. MLB uses SR-MLB-0.5.3 with SR-MLB-LIVE-0.5. College football customer release remains fail-closed because the earlier independent projection did not add measurable 2025 holdout value beyond the closing line. Its replacement, SR-CFB-EFF-RESIDUAL-0.1, uses the market as a baseline and tests opponent-adjusted EPA, success rate and explosiveness through prospectively locked private shadows. Those shadows are emailed only to the owner, graded automatically and kept outside the Official Play record. NFL uses SR-NFL-0.2 with SR-NFL-LIVE-0.6. Every historical pick keeps its locked model and release version.
1. Sport-specific research inputs
MLB evaluates starting-pitcher quality and pitch mix, projected lineups and platoon splits, bullpen availability, park effects, weather, travel, rest, injuries, defense, and price. Football requires its own contract: quarterback and roster availability, opponent-adjusted efficiency, explosiveness, finishing drives, trenches, coverage, special teams, pace, coaching, travel, rest, weather, and the full two-sided market. Missing critical inputs lower confidence rather than being silently guessed.
2. Probability before opinion
The model produces a fair win probability. The market price is converted to implied probability and adjusted for sportsbook margin. A recommendation exists only when the difference is large enough to survive uncertainty and the available odds remain within the published limit.
3. Validation before customers
Every sport must pass its own time-aware testing that prevents future information from leaking backward. We monitor calibration, out-of-sample return, closing-line value, performance by edge bucket, and performance by game context. Results from one sport cannot validate another, and a backtest alone is not permission to sell advice.
4. Version every material change
Feature additions, weighting changes, data-source changes, and threshold changes receive a new model version. Each published recommendation stores the version that produced it so results from different approaches are never blended invisibly.
5. Publish, lock, and grade
A live recommendation stores its timestamp, selection, available price, maximum playable price, unit size, probability, estimated edge, rationale, and model version. After publication, the record is locked. The result and closing price are added later; the original research is not edited away.
6. Measure closing-line value separately
Every graded release is evaluated against a directly captured closing market when the exact paired evidence is available. CLV is reported separately from win/loss results and never changes the original grade. College-football Shadow Qualifiers and Weekly Research Picks appear in a distinct research ledger and never enter the Official Play record.
6. Responsible use
Model outputs are estimates, not guarantees. Sports Rounders uses small fixed-unit sizing, discourages chasing, and may publish no play when the evidence or price is weak. Members remain responsible for their own decisions and local laws.