In the repeated Prisoner's Dilemma, predictions are notoriously difficult. Recently, however, Blonski, Ockenfels, and Spagnolo (2011, BOS) showed that experimental subjects predictably cooperate when the discount factor exceeds a particular threshold. I show that this threshold implies existence of an equilibrium robust to two standard refinement assumptions (utility perturbations and imperfect monitoring). The equilibrium is "Semi-Grim": Cooperate after mutual cooperation, defect after mutual defection, randomize otherwise. Testing six resulting predictions on existing data, comprising 37.000 observations, I then find that subjects indeed play Semi-Grim strategies, and switch to cooperation in round 1, when the former turn into equilibria -- at the BOS-threshold.