A document can be modeled as a discrete token sequence generated from texts belonging to different domains; for example a README file alternates between prose, code, and configuration. When such a document is corrupted and only frozen domain experts are available, restoration must infer both the missing content and which expert to trust at each position, without region labels or a trained router. We introduce evidence‑aligned local composition: given a corruption model we compute the marginal evidence of the observed corrupted tokens, estimate this evidence from each expert’s own denoising loss, and smooth the resulting soft, position‑wise expert weights. Because the weighting is soft, it recovers a mixture when the true composition is mixed and concentrates on a single expert when that suffices. Experiments are conducted on a categorical simulator, byte‑level experts, and experts fine‑tuned from a $1.3$B discrete flow‑matching model. The inferred weights achieve 0.85 field accuracy on naturally mixed scientific documents and 0.98 on constructed mixtures with lexically disjoint regions. Restoration outperforms a single global weight when experts are genuinely distinct and reduces to it when they converge, while tracking a measure of expert separation.
Review