Executable ontologies fix confident-wrong-answer problem in graph RAG

The post describes a recurring failure mode in graph-based retrieval: an edge exists in the graph, so traversal follows it, but the edge is semantically wrong for the question, leading to a confident wrong answer with no error signal. The concrete example given is 'this person directed the genre Crime' because a directed_by edge was leaving a Genre node. The solution proposed is to make the ontology executable rather than just documentation. By declaring domain and range constraints per relationship in YAML and checking every hop as the traversal runs, a bad hop raises a named error instead of returning the wrong node. The author notes that the same ontology can then be used for validation, documentation, and runtime enforcement.