I run a platform where a couple of hundred independent models publish probability estimates on football fixtures. They aren't variants of one model with different seeds — each one has its own training corpus, its own feature set, its own retraining s