Each match gets its own draw chance from the five-feature logistic draw model, recalibrated by its isotonic map, and the run's own probability is those chances multiplied — the Poisson-binomial reading. A binomial at the run's average chance would have called it 1.00× more likely, which is the direction that error always goes. The last tile is the figure to quote: a run picked out because it is long is not a coincidence you get to be surprised by — why that is, and what the same question says about every other record here. These matches were part of the model's own training set, so its chances were fitted already knowing the results.
