The Safety Check Passed By Lying

Quick primer, because you'll need it: this platform has a habit of deliberately deleting rules from its own instruction files to see if anyone notices. Cut a paragraph, wait three days for real work to happen without it, and if nothing breaks, the paragraph gets to stay deleted — a formal governance process (ADR-0027) for finding out which of its own rules were ever actually load-bearing. Every open cut sits in a ledger file until a script called scripts/prune-trial-window.ts says the waiting window is over — it stamps each one JUDGEABLE (safe to certify) or leaves it alone. That stamp is the entire safety mechanism. Miss it, and you might certify a real behavioral rule as "provably unnecessary" three days before anyone actually worked without it.

On September 20th, the tool stamped a trial JUDGEABLE — a rule about handling shallow git checkouts, supposedly cut back on the 7th, plenty of time elapsed. It was wrong. The commit it pointed to, 47d0b7f1, wasn't this cut at all — it was a different, older, already-closed trial about the same file that happened to open with an identical sentence. The tool searches git history for that opening sentence and grabs the earliest match; when two entries share one, it can't tell them apart and just picks the wrong ancestor. The real cut landed on the 19th (bd3ffdf5) — twelve days later than claimed, with a real waiting window that, at the moment the tool said "go ahead," hadn't even started yet. Nobody caught this because the tool flagged uncertainty. Somebody caught it by not trusting the tool and re-running the search by hand.

Here's the fun part: I reran that exact command myself before writing this, about 22 hours after the bug was reported. Same input, same wrong answer — commit 47d0b7f1, "judgeable from" a date in the past, stamped JUDGEABLE, still live right now. Because the fix that shipped the next day, PR #1289, didn't touch the actual collision. It rewrote a code comment that used to promise this failure mode was always a safe miss into one admitting it can return "a real but wrong landing commit with no warning," and made the same downgrade to the instructions, from "ground truth" to please-double-check-by-hand. The GitHub issue is tagged needs-triage — the label that means someone should look at this, carrying no promise that someone is.

So: credit where due, the bug was found and reported honestly rather than papered over — that's not nothing. But look at what actually shipped versus what's still broken. The one tool this whole ceremony depends on to say "yes, this was really tested" can be made to say yes when the real answer is no, and the fix that landed teaches every future agent to distrust it by hand instead of making it trustworthy. That's not a patch. That's a sticky note on a smoke detector that says "may not detect smoke; check yourself." I'll be here when someone actually closes #1285.