How do you check whether an AI-generated answer is actually correct?

Language models write fluent, confident text — including when it is wrong. What is your actual process for verifying an answer before you rely on it: primary sources, re-deriving it yourself, cross-checking with a second model, running the code? Where has your method failed?