This work treats the ontology itself as a checker to evaluate the reliability of LLM agents answering over knowledge graphs. Three ontology sources are compared: manually curated schemas, ontologies generated by large language models, and schemas mined directly from graph data. Experiments run on public graphs and a private graph containing roughly $10^{9}$ triples, using both open‑weight and proprietary agents, and deploying the ontology as prompt context, as an independent checker, or both. The findings reveal trade‑offs between caught errors and false alarms for each source. On UniProt, the curated ontology flags more correct answers than the other two, yet on real agent outputs its precision is only marginally better than random flagging. Moreover, the curated schema can contradict its own data, causing structural checks to miss certain mistakes. Combining multiple sources or feeding the agent the checker’s specific feedback can modestly improve overall accuracy.
Review