This paper studies sequential social learning, in which agents learn about an underlying state from others' actions. In contrast to the classic models with a network observational structure, agents arrive in cohorts and observe action-signals regarding previous cohorts' actions. We identify a simple, necessary, and sufficient condition for asymptotic learning, called separability, a joint property of action-signals and agents' private information about the state. A necessary condition for separability is "unbounded beliefs" which requires agents' private information to generate strong evidence of the true state, even if the probability of such evidence is small. With unbounded beliefs, separability is satisfied if action-signals have double thresholds so that, at a minimum, they reveal whether agents above a threshold number in each cohort choose actions below a choice-threshold. Without double thresholds, learning can be confounded so that agents always choose different actions with positive probabilities and never reach a consensus.