Audit an AI football prediction
A seven-part test for publication proof, model identity, evaluation quality and a complete result history.
Evidence-first learning centre
Start with the probability, then follow its evidence. These guides explain what a forecast says, how a genuine model should be tested and which public records are needed before an accuracy claim is worth attention.
Published by Football Proof AI · Updated · Educational material, not betting adviceStart here
A model output is a distribution of possible outcomes. The first task is to understand that distribution; the second is to verify how it was produced and recorded.
A seven-part test for publication proof, model identity, evaluation quality and a complete result history.
Interpret a complete home-draw-away split, calibration, fair odds and Brier score without treating uncertainty as certainty.
Rebuild a three-outcome Brier score and see how one result changes the error assigned to a probability split.
Inspect the system
A polished explanation cannot validate a probability. Model inputs, time-ordered evaluation, artifact identity and data limitations must remain inspectable independently of the interface that describes them.
Paste or upload a complete 1X2 history and audit timing, exclusions, hit rate, uncertainty, Brier, log loss and calibration locally.
Reproduce hit rate, multiclass Brier score, calibration, ECE, log loss and Wilson intervals from one worked 1X2 forecast.
Build rolling-origin folds, point-in-time features, separate calibration windows and fixed gates without future leakage.
Follow the feature window, leakage controls, LightGBM estimate, isotonic calibration and walk-forward acceptance gates.
Check registered model versions, evaluation windows, sample sizes, metrics and immutable artifact fingerprints.
See which fixture, historical, model and settlement data enters the product—and which inputs remain deliberately disabled.
Verify the record
Accuracy is credible only when misses remain visible, corrections preserve the original call and responsible-use limits sit beside the forecast rather than behind it.
Audit every eligible settled forecast, league and model split, calibration bucket and row-level result.
Understand how later score corrections are appended without silently rewriting the original probability record.
Compare complete probability shapes using only historical rows that had settled before the target was published.
Keep uncertain model output separate from staking, profit promises and harmful gambling behaviour.