In supervised machine learning, evaluating a model is usually straightforward. Given a labeled data set, we can compute standard metrics such as accuracy, mean squared error, or cross-entropy loss by comparing predictions to true labels. In causal inference[1], however, evaluation becomes much more challenging. This is because for any individual, we can observe only one of the two potential outcomes—the outcome under the treatment they received. The other outcome, the counterfactual, is inherently unobservable.