Claims of "Zero-hallucination" usually fall apart when the engine has to derive a new insight from the graph. If the LLM isn't reasoning, how are you enforcing

Claims of "Zero-hallucination" usually fall apart when the engine has to derive a new insight from the graph. If the LLM isn't reasoning, how are you enforcing Fidelity between the graph source and the final output?