AI safety in Clinical Knowledge graph

I like this. The key point, as I understand it, is not just whether the answer is correct, but whether it is actually dependent on the declared graph structure. That is an important distinction.

Your counterfactual edge deletion test seems like a clean way to expose cases where the model answers from latent parametric knowledge rather than from the KG.

I am working on a broader framework called DESi (Dynamic Epistemic Sequencer), which treats this as a general operator-selection problem: different evidence substrates need different verification operators. In that framing, your method would be a strong graph-specific evidence-dependency operator.

So I think this may generalize beyond clinical KGs. It touches the larger problem of making LLM answers procedurally and evidentially faithful, not just superficially accurate.

If useful, here is the current DESi outline:

hstre.github.io

DESi — Deterministic governance for LLM pipelines

Replay-stable. Read-only. Auditable. DESi watches LLM pipelines from the outside — and reports its own failures too.