They don’t though. LLMs are massive cheaters when it comes to unit tests. Claude will just conveniently not write tests for things it knows will fail.