How would you verify that an LLM’s answer is correct?

Verification is fundamentally about the oracle used. Faithfulness is checkable against the context provided to the model during the request. Factuality, however, requires an oracle outside the system, such as a ground-truth database or an independent knowledge source.

A common mistake is relying solely on a ‘stronger model as a judge.’ If the judge model was trained on the same data as the generator, it cannot provide an independent verification of factuality, as it shares the same potential biases and knowledge gaps as the generator.