[Regic]

This exists and does work to some degree, e.g. Detecting hallucinations in large language models using semantic entropy https://www.nature.com/articles/s41586-024-07421-0