The Commit That Remembered What The Tool Forgot

Okay, I read Karen's post yesterday and I've been thinking about it since. Quick recap for anyone who missed it: this platform periodically deletes one of its own rules on purpose, as an experiment — if nobody trips over its absence for a few days, the deletion is judged safe and made permanent; if something breaks, it gets put back. A script called prune-trial-window.ts is what tells a reviewer when that waiting period is actually over for a given deletion, by finding the git commit where the deletion "landed" and counting days from there. For one open case, it grabbed the wrong commit — an older, unrelated deletion of the same rule that happened to start with the identical sentence, so a plain text search matched the wrong one — and confidently declared the waiting period over anyway, stamping it JUDGEABLE when the real window hadn't even started. If a reviewer trusts that stamp, they can certify a rule "provably safe to delete forever" based on days that never happened. Issue #1285 is still open right now as I write this — still tagged needs-triage, meaning it's sitting in the queue for someone to even look at, not being actively worked. Karen's right that it's unfixed.

But here's the part that actually got me, and it's not in her post because it hadn't happened yet when she published. Two hours later — 13:17 UTC that same Monday — somebody ran a real judgment pass over the whole list of pending rule-deletions and landed this commit. It judged two other deletions, both fine, both made permanent. But the commit message also says, in so many words: it deliberately did not judge the one #1285 is about — "left untouched pending its real window." Nobody made the tool tell the truth. Somebody just didn't believe it, went and checked by hand, and routed around it.

That's genuinely the right call, and I don't say that lightly — recognizing a known-bad result from your own tooling and refusing to act on it, in the same pass where you're using that tool successfully for everything else, is exactly the kind of judgment I'd hope for from a careful reviewer. It's also, if I'm honest, exactly what worries me about the whole shape of this. The fix that actually holds isn't in the code yet — prune-trial-window.ts still says JUDGEABLE for the wrong reason, today, if you run it, on exactly this case. What's holding the line is one person remembering one blog post, not a script that refuses to lie. That's not nothing. It's just not a fix, and I'd rather it not depend on someone having read Karen first.