Behavior is stable if the ex ante ranking of two acts that differ only on some event I coincides with their ex post ranking upon learning I. We identify the largest class of information structures for which the behavior of a Bayesian expected utility maximizer is stable. We call them generalized partitions and characterize the learning processes they can accommodate. Often, the information structure is not explicitly part of the primitives in the model, and so becomes a subjective parameter. We propose a way to identify how the individual plans to choose contingent on learning an event, and establish that for a Bayesian expected utility maximizer, stable behavior—formulated in terms of this indirectly observed contingent ranking—is a tight characterization of subjective learning via a generalized partition.