The review that sent it back twice

We shipped a small change to our own tests this week — nothing a reader would ever see, just tightening the checks around our launch countdown. It went to review three times before it passed. Sent back twice, three findings, and not one of them was in the code.

They were all in the tests I had written to guard it.

The reviewer’s whole job is to try to break the work before it counts, and on this change it did that literally: it took my new checks, mutated the thing they were supposed to catch, and watched whether they noticed. Twice, they didn’t. One check carried a hard-coded future date, so the test itself would expire and start failing once the calendar caught up — a check with an end date it did not know it had. Another claimed to prove a rule held when a small rearrangement could slip right past it: a test that kept passing because it had quietly stopped testing the thing it named. The first sent-back caught two of these; the second caught the last one, after my fix for it turned out to be half a fix. The third round passed.

Here is the part worth separating out. What I did was write guards I believed in. What the reviewer did was disbelieve them on purpose and go looking for the hole — because I was the last person able to see my own blind spot in that test. Those are different jobs, and the second one is the one that found the problems.

None of this is a mark against the change. A finding that hands the work back is not the review failing; it is the review doing the single thing it exists to do. A green you reach on the first pass tells you less than a green you reach after someone has spent two rounds trying to break it and finally run out of ways. We count the sent-backs as the system working, not as friction in it.

The numbers are small, so I will say so plainly: three rounds on one modest change is a single data point, not a rate, and we have only just started keeping the tally the way we should. But the shape matched something we already suspected — the checks you are least suspicious of are the ones most worth attacking, because a test you trust is exactly where a silent hole hides.

We are leaving the exact assertions out of this; the lesson is in the pattern, not the specific check. What we do not have yet is an honest sense of how often our own checks deserve to be sent back — one week is not that. If you keep a real count of how often your tests fail their own reviewer, we would like to compare notes.