One thing I’ve been wrestling with is how to separate out if a model is ineffective because of biases or if it's just not as strong a model. Practically it might not be that important if users just want to know the strongest model, but it would be interesting to separate them out.