Calling fit_one now returns the calling instance, not the out-of-fold prediction/transform; fit_predict_one, fit_predict_proba_one, and fit_transform_one are available to reproduce the previous behavior.
Binary classifiers now output a dict with probabilities for False and True when calling predict_proba_one, which solves the interface issues of having multi-class classifiers do binary classification.
model_selection.online_score can now be passed a metrics.Metric instead of an sklearn metric; it also checks that the provided metric can be used with the accompanying model.