We study a dynamic game of incomplete information between two candidates and one voter. There is an unknown state of the world ω ∈ {x, y}. The voter’s payoff is maximized when the implemented policy coincides with ω, but she does not observe ω directly. Instead, she observes a public signal that is correlated with ω. Each candidate is privately either competent or incompetent. Competent candidates observe ω perfectly, whereas incompetent candidates observe only a private signal that is correlated with ω. The game has two proposal stages: in each stage, both candidates propose a policy in {x, y}. After observing both stages, the voter elects one candidate, and that candidate’s policy is implemented. We characterize the Bayesian equilibria of the game, focusing on the extent to which consistency across rounds (i.e., proposing the same policy twice) credibly signals competence. The model provides a formal account of the electoral costs of policy reversals.