AI safety in Clinical Knowledge graph

hello, Yes you got it correct. The main idea is that model should not hallucinate, rather it should generate based on true knowledge. And as you also guessed correctly, this is not only for clinical AI, it is applicable widely. In my research I used medical dataset. One good point you said about “different evidence substrates need different verification operators”, this sounds like Mixture of Experts that most LLMs are using for chain of thoughts.