A strong baseline policy used to compute ground-truth Q-values for evaluating supervision signal quality.