Regional league win probability model

Trained on 1,400 maps of parsed demos, updating on utility spent rather than on kills. Calibrated within four points on held-out playoff data, which is where most public models fall apart.