A regression test is accepted when it passes on the code it was written against. If that code is wrong, the test now asserts the wrong answer and will keep asserting it. Across 145 merged changes in three large Python projects, 8.4% to 16.9% of generated tests pinned the faulty behaviour in place against 2.4% to 4.8% that caught it — and between 82.8% and 90.9% of the bad ones were still passing, still guarding the wrong answer, at the end of each project’s later history.