A learned model that decides which token predictions are confident enough to commit (unmask) during parallel generation in diffusion models.