Any four tracked players, any format. Uses current-month posterior values, the weakest-link penalty (γ = −0.18), and the exact race-to-target score distribution. The interval reflects real uncertainty about the players — that's the honest part.
Fine print: pair chemistry is excluded on purpose — the fitted effects are almost all within ±0.1 pts and none is statistically certifiable. Values come from a joint Bayesian fit of ~36k games; the win-probability interval is the 90% posterior range from player-value uncertainty. Probabilities are calibrated against out-of-sample games and never reach 0% or 100% — across 36k games, favorites we rate above 99% still lost about 1 time in 100, so that's the ceiling (methods). If the two sides' gender mix differs, the number rests on a modeling convention that no game data can test.