Each episode is a different sample from the model, annotated with the inductive biases it follows.