I think it shows "100% test coverage" isn't testing what matters otherwise AI would get signals back telling it things are bad so keep working until they are good. I don't have a solution for this but more and more tests isn't the answer. I had AI create a small-ish CLI the other day and it added >2k tests and things still felt garbage. Now imagine that at GH.com scale. AI won't solve it with the current paradigms. I would love if someone could point me at good research in this area (not prompt evals).