Model quality comes first.
We are reviewing the calibration and ranking power of our early football models. Editorial research may be published before a market is qualified as a Certified CORE recommendation.
Probability is not a guarantee.
For Premier League Under 3.5, among 25 later-season test games where our historical Model A forecast was at least 70%, its average estimate was 75.2% and the observed occurrence rate was 64.0%. This sample is too small to establish a precise long-run calibration error, but it does not justify high-confidence publication claims.
AUC = ranking discrimination (0.50 approximately random). Distinct seasons produce different results; these are archived research evaluations and not independently certified sport-market model performance.
New evidence-control layer
AIONORA 4.10 introduces separately tested research checks for coherent football-market probabilities, source and kickoff timestamps, market/line consistency, calibrated-versus-uncalibrated forecast provenance and operator-price verification. All 88 tests in the extended offline engine suite passed.
The research quality checks do not prove that future tips will win more often. No weak calibration mapping was promoted. Formal betting certification still requires separately approved models, positive conservative value and legitimate executable odds.
What we are improving next
Richer as-of historical team-strength and opponent-adjusted xG data; robust promoted-team priors; meaningful player-status effects; separately calibrated totals, sides, basketball and tennis forecasts; and an untouched chronological evaluation period.
Current production has no independently approved model version for formal Certified CORE release. Price-implied probabilities in the Daily editorial feed are identified as market references, not independent AIONORA win estimates.
Explore the historical data audit ↗