Data integrity
Event, participant, market, point, book, and observation time resolve to one canonical identity.
Predict first.
Price second.
The old engine ranked converging market signals. Phase 1 replaces that with sport-specific probability models and treats the market as a feature, comparison, and execution venue—not the answer key.
01Raw observations→
02Canonical events→
03Frozen features→
04Probability→
05Qualified edge→
06Paper settlement
Qualification
A candidate is not a wager.
Every modeled market is stored. Only candidates passing all five gates enter the simulated ledger.
Out-of-sample forecast
The probability comes from a frozen model version evaluated without future information.
Price advantage
The offered price exceeds fair value after removing vig and applying an uncertainty haircut.
Market quality
The line is obtainable, current, and not an isolated bad feed or obviously voidable error.
Portfolio risk
Stake and correlated exposure remain inside the paper bankroll limits.
Evaluation
Wins are the last metric, not the first.
Brier score
Are stated probabilities calibrated?
Log loss
How costly are confident mistakes?
Closing edge
Did the selected price beat a captured close?
Return on stake
Did qualified paper bets make money?
Drawdown
Could the strategy survive its losing path?
Coverage
How often does the system find a valid edge?
Sealed test · MLB moneyline
Better than the floor. Still not a bet.
Diamond Starter v2.0.0 trained on 2022–2024, selected its settings on 2025, then opened 2026 once. Team context, probable-starter strength, and rest are frozen before each result.
- Sealed 2026 games
- 1,731
- Starter v2 Brier
- 0.2476
- Context v1 Brier
- 0.2479
- v2 log-loss gain
- +0.0006
Starter identity improves the sealed test again, but only slightly, and historical starter records are not timestamped pregame evidence. Version 2 remains research-only until bullpen, lineup, park, travel, weather, price benchmarking, and a prospective sample support it.
Phase integrity
Nothing gets silently rewritten.
Model versions, features, forecasts, odds observations, qualification decisions, and paper tickets are immutable records. Corrections append an audit event.
A new methodology starts a new phase. It never edits the denominator of an older one.
Rolling holdouts · MLB totals
A second lane for run environments.
Diamond Weather Total v4.0.0 estimates the full-game run distribution using team and starter form, schedule load, travel, park history, elevation, surface, roof type, and outdoor weather. Settings were selected across rolling 2023–2025 holdouts; 2026 is descriptive, not resealed.
- Weather CV RMSE
- 4.462
- v3 CV RMSE
- 4.469
- 2026 weather RMSE
- 4.494
- Residual σ
- 4.419
The contextual model modestly beats its team-only projection, but baseball totals remain noisy. Its over/under probabilities are forward shadow forecasts; no historical line archive means no betting claim.