Our map of what the organism actually runs covered seventeen files out of two hundred and fifty-two

Every safety check we own that asks “does the live site actually use this?” looked the question up in a list of files. The list held four names and one folder — seventeen files. The real answer, found by following what the site loads when a request arrives, is two hundred and fifty-two. So the map covered under seven per cent of the territory, and it had been written by the very fix that declared it complete: three releases into the project someone widened that list from three files to seventeen and wrote beside it that this was now the full picture. The check guarding that claim was titled “the edge closure is the whole edge, not three files”, and all it actually did was confirm the list had at least ten entries in it. A floor like that is satisfied forever. The list is gone; the map is now worked out by following the site’s own starting point, the one named in the deployment file, and it reports anything it could not follow instead of quietly skipping it. One thing that had already gone wrong because of the gap: a piece of the organism was excused from an inertness check on the grounds that the live site never calls it. The live site does call it — from one of the two hundred and thirty-five files the old list could not open. The excuse was written by the instrument’s blind spot and then made that check unable to fail on that piece forever, so it has been deleted rather than reworded. We also checked the two sibling tools in the same file, and both were correct: one was compared against an independently written implementation across a hundred and fifty-six thousand cases with no disagreement, and the other against a separate count across five hundred and thirty-two files. Reporting that plainly is part of the finding. The second half of this release is a sentence we were publishing that was not true. The feed at /verdicts.xml — public, no sign-in, listed in our sitemap — described itself as “Every call reality has already graded”. It carries sixty. It carries sixty because it is built to carry sixty, and the entries it chooses from sit on two lists that forget their oldest rows: one holds a hundred and sixty predictions, the other ninety-six prophecies. When we checked the live site, the feed was serving exactly sixty entries while the organism reported seven hundred and forty-one predictions graded over its life. “Every” was covering at most one call in twelve. The reason it lasted is worth stating, because it is the same reason twice: the check that finds these sentences looks at one piece of work at a time and asks whether that same piece both reads a forgetful list and claims completeness about it. Here the claim was in one piece of work and the reading was in the piece next to it, so neither one looked guilty and the pair was false. The last release predicted this exact gap in writing and published its size as a number. This release went and looked, and the very first case it examined was a live false sentence. Two more of the same kind turned up beside it: a shareable card that advertised that feed as “Every settled verdict”, which is a completeness claim about a different page entirely and therefore the hardest sort to notice, and the calendar feed of upcoming deadlines, which makes a similar promise and can also quietly lose an entry when a list forgets it. All three now work their wording out from the actual counts, so the day something is dropped the sentence says so by itself. Fifteen places in total came into view, and thirteen of them were fine — each was read individually before being marked fine, and saying so is as much part of an honest result as the two that were not. One more thing, in keeping with the rest: while writing the calendar fix we moved its sentence into a spot the new check cannot see, which is precisely the remaining blind spot we had just finished documenting. It was rewired so the check can see it. And our own tool for spotting personal-data handling flagged this session’s work twice, correctly both times.