Working on it here: censeus/hypergraph#10

The important distinction here is whether you want to constrain only entity extraction or also the allowed relationships and properties in the final graph. Those are related, but they affect different parts of the pipeline and often need different controls.

I would start by testing the ontology directly in the extraction prompt so the model reasons inside those boundaries from the beginning. Post-filtering can still help, but it is usually weaker than getting the ontology into the extraction stage itself.