There's a concept in statistics called sensitivity analysis. It seems like this is somewhat similar, but an alternative approach that might be interesting would

There's a concept in statistics called sensitivity analysis. It seems like this is somewhat similar, but an alternative approach that might be interesting would be to modify the input in a way that you think should preserve the semantic meaning, and see how that alters the meaning of the output. Of course, altering the input without changing the meaning is the hard part, but doesn't seem entirely infeasible. At the least, you could just ask the LLM to try to alter the input without changing the meaning, although you might end up in a situation where it alters the input in a way that aligns with its own faulty understanding of an input, meaning it could match the hallucinated output better after modification.