Another says that they should audit the AI code using AI. Basically saying lets wrap the blackbox with another blackbox but this time with markerpen writing on top of the blackbox.

Idk how's that supposed to help, i'm not a programmer, but i'd assume if Claude misunderstood the code that they themselves wrote, the project became unfixable, and they might have to see themselves writing the whole thing from the ground up.