Abstracting fairness: oracles, metrics, and interpretability