A Markov Decision Process where each step involves two sequential decisions: selecting token values and choosing which positions to unmask next.