We derive information-theoretic identities for discrete diffusion, revealing score-based losses as exact mutual information decompositions and enabling principled log-likelihood estimation.