#system-design

LLMs are terrible at generating correct test cases. Here's the architecture that actually works.

How I stopped asking an LLM to judge correctness and built a reliable hidden-test pipeline instead.