The reverse CTMC. Every position starts at MASK; over a handful of steps the network unmasks tokens in parallel, jumping each from MASK to its value. This is masked-diffusion sampling.