StatPigA reproducible research process, not a validated betting edge. Build 13 separates private evidence, immutable decision records, and public presentation. Model v1.0 is preserved in the private archive; its earlier picks are not retroactive v1.1 tests.
An official BET needs a complete, verified slate snapshot with two accessible books per spread and total; exact quote, odds and timestamp; complete personnel, injury-cluster, replacement, weather, travel, opponent and matchup audits; and complete prop-role audits where applicable.
Quotes expire after 20 minutes. Final inactives become mandatory inside 90 minutes before kickoff. Unknown information is incomplete, not checked. Source truth still requires human review.
Sides and totals combine an independent football forecast with a market-centered empirical residual distribution. The football component receives a provisional 20% weight through Week 4 and 35% afterward. Injuries adjust the independent forecast only, never the market component a second time.
Props combine leading, neutral and trailing outcome distributions with explicit weights. A higher average alone is not a win probability. Integer outcomes preserve pushes; American odds determine expected return.
All gates must clear together: estimated expected return at least 3%; conditional non-push win probability at least 53%; nonnegative return in at least two documented adverse scenarios; exact line and odds limits; and mean-based edge of at least 1.5 points for sides or 2 points for totals. These are conservative operating assumptions, not empirically optimized coefficients. A High label means the gates cleared, not that a particular hit rate has been demonstrated.
Sides/totals require at least 30 documented walk-forward residual observations; this minimum is a basic data floor, not proof of statistical adequacy. Every WATCH needs both a concrete action trigger and a kill condition. Missing required evidence forces PASS / ANALYSIS INCOMPLETE.
Official paper recommendations freeze the issue time, exact price, bookmaker, model version, input calculations and units risked. Withdrawals stop new action without deleting history. Results are calculated against that frozen contract. Corrections append a new event and preserve the old event.
ROI is settled net units divided by settled non-void units risked. Pushes remain in risk; pending recommendations do not. Closing movement is separated by market units; no spread-point/prop-yard average. Earlier unverified records never enter prospective model metrics.
Paper limits are one unit per recommendation, one unit per game, one unit per shared causal group per week, and six gross units per week. These are research controls, not personal bankroll advice. Nothing in this software places a bet.
An experimental independent football model and market collection are implemented. Football model 0.1.2 corrects team-game aggregation, enforces a pre-target-week calibration cutoff, and blocks incomplete prior-week play-by-play data. It remains separate from the decision-policy version above. Current personnel, weather, role and scenario evidence still require review; a successful research benchmark does not establish a betting edge.
The public site cannot edit the database. It reads an exported data file or a read-only API. The protected recorder and raw evidence belong on a local computer or managed backend, never in the public GitHub Pages upload. Hash chaining detects ordinary record changes but is not proof against an administrator rebuilding the database; independent backups remain important.
Open Results & Data →