Continuously ingest fixtures, team identities, match states, results, and targeted market data for the next 36 hours across the 300-event allowlist, then validate timestamps, completeness, and duplicates.
Builds stable mappings from unique fixture, league, and team identifiers; recognizes home and away roles, competition tier, season phase, and cross-league relationships; and prevents name variations from causing mismatches.
AI maintains team attacking and defensive strength, home and away performance, recent form, and cross-league ability in true chronological order. Every match is analysed before its state is updated, strictly isolating future information.
The online goal-intensity model generates expected goals for both teams and builds a correlation-adjusted joint score matrix, deriving analyses for every validated market from one probability space.
The system increases scan frequency as kickoff approaches and tracks only valid quotes for the same match, market, selection, and line, continuously capturing price changes.
Global priors, league context, team state, and the maturity of historical evidence enter the calibration layer together, converting raw model probabilities into publishable probabilities closer to observed frequencies.
AI converts different markets into verifiable match-state sets and automatically detects hard cross-market conflicts. Member analyses take priority; equal-tier conflicts are resolved by calibration strength, and all are suppressed when they cannot be separated.
Eligible fixtures first receive baseline analyses. Only selections that pass data quality, historical maturity, probability threshold, market validation, and consistency checks enter the member candidate pool.
The system calculates a model reference price from the calibrated probability. It keeps monitoring while the market is below the price requirement; the first qualifying observation freezes the record, and later observations are appended without overwriting it.
After the match ends, the result is filled in automatically. Standard analysis results and real market records are settled separately; only research records that actually meet the recording conditions enter backtest metrics, volatility, and drawdown statistics.
Every analysis is permanently frozen before kickoff, and team states are updated sequentially only after results arrive. New data continues into later calibration and validation, while historical analyses can never be revised to fit outcomes.
Fixtures, analyses, member research records, price observations, later prices, and post-match archive results are stored separately. Each output uses an atomic snapshot and SHA-256 verification to form a complete, traceable AI decision chain.
