A common concern in machine learning (ML) solutions is that apparent predictive power is coming from a problematic source.¹ For example, a model might learn to predict burrito quality from latitude and longitude. In this case, the actual signal is likely coming from a particular city or neighborhood having …