Inside ASCEND-RISK: How the AI Bankroll Manager Sizes a Card
Sizing is where most records are actually won or lost. This engine weights realised edge and calibration drift above everything else, then de-rates for correlation.
It sizes from settled results, not stated confidence
Realised Edge (22%) is the settled edge of a tier over its trailing 250 plays. Calibration Drift (18%) measures the gap between stated and settled probability by bucket. Together they mean a tier that has been over-claiming gets smaller stakes automatically, without anyone editing the model.
Closing-Line Value (13%) tracks the average beat against the closing number, which is the leading indicator that survives even when a short run of results does not.
Correlation and variance shrink the ticket
Ticket Correlation (16%) reads shared game, player and market exposure across the card, because six plays on one game is one play with extra vig. Payout Variance (15%) accounts for the dispersion of returns at the prices actually available, not the idealised price.
Limit Availability (9%) asks how much stake clears at the quoted price without moving it. A stake you cannot get down is not a stake.
Drawdown state is an input, not an afterthought
Drawdown State (7%) places the account inside a modelled drawdown envelope and reduces exposure as it deepens. Fractional Kelly with a hard cap does the rest: the engine will not put a card's outcome on one read, however strong that read looks.
Nothing here removes variance. A genuine, thin edge still produces losing weeks, and the engine's job is to keep those weeks survivable rather than to promise they will not happen.