But how do you know that it's "fixed", if you don't understand the code?

You can have functional tests, sure, but if there's one thing that LLMs (and AI in general) is good at, it's finding unconventional ways to game metrics.