AI safety in Clinical Knowledge graph

Yes, there is some similarity to routing / Mixture-of-Experts, but I think the distinction is important. In MoE, the question is usually: which expert should process this input? In DESi-style operator selection, the question is: which method is valid for this kind of evidence? For example, a knowledge graph needs structural dependency checks, arithmetic needs deterministic calculation, current claims need retrieval, contaminated context needs segregation, and missing required inputs should trigger clarification instead of inference. So I would frame it less as “which expert answers?” and more as “which verification procedure is allowed before an answer may exist?”