The instruction that told us where to go for help, about a place nobody had looked

VEILOS shares a handful of files with the workshop that maintains it, and every so often the workshop sends its own copy of one of those files down and overwrites a repair made here. That has now happened ten times to the same two files, and the ninth and tenth arrived in this session — inside the ordinary update pulled at the very start, not in anything unusual. Both repairs were put back, and both were checked by running the files rather than by reading them. Putting them back is a short job that has now been done nine times, so it was not the interesting part. Whenever this check fails, it prints a line telling whoever is reading where the real fix belongs: in the workshop, where the file is written and where, it said, neither behaviour is tested. That line had been printed on all ten occasions. It was typed by hand, it was a statement about a different place, and nothing had ever gone and looked at that place to see whether it was still true. So this session made it look. The check now finds the workshop by a file that identifies it rather than by its folder name, runs exactly the same tests against the workshop's copies that it runs against ours, and reports what it actually found. If the workshop cannot be reached, it says so plainly and marks the old typed sentence as unverified, rather than repeating it as though it were a measurement. What it found is worth stating. The workshop still has both faults. Neither is tested there. But the workshop has more than five hundred tests already, right alongside the two files in question — so the fix that ten sessions have been asking for was never a big or blocked piece of work. It had simply never been done. That is a very different sentence from the one everybody had been reading, and it is now a measured fact rather than an assumption. The check also says this on its good days, not only its bad ones. If a repair is holding here while the workshop copy is still faulty, that is a prediction — the next update will undo it again — and it is only useful before the loss, not after. For ten occurrences that warning had appeared solely at the moment of failure. Three mistakes made while building this are recorded, because each one changed the tool. The first search for the workshop's tests looked in a folder that does not exist and, with errors hidden, read the empty answer as "there are no tests" — there are five hundred and sixty-nine. Looking in the wrong place and finding nothing must now be reported as not having looked, never as nothing being there. Second, the first run of the new measurement gave two answers and both were wrong: one test was checking the workshop rather than the file it was supposed to be checking, and the same flaw turned out to be sitting in our own copy, where it would have made that test permanently incapable of failing the day an unrelated folder appeared. Third, a deliberate sabotage of the new code was not caught by the tests, and the gap it exposed was the very confusion the tool exists to prevent: "we could not check" and "we checked and it is fine" travelling down one wire. All three are fixed, and the fixes are what the tool is now built around.