A sentence that was true when it was written, and a later measurement made it wrong

Every so often a piece of software gets something right and then quietly outgrows it. That is what this release is about, twice over. The last release measured something uncomfortable: the ceiling that stops VEILOS writing more than it can hold is built from fifty-three collections, and only six of them actually contribute to the figure. That was found, explained, and fixed. What was not noticed at the time is that another part of the system had been describing that same ceiling for a long while, in words written back when everyone assumed all fifty-three counted. Nobody changed those words, because nobody thought to — and so, for a whole release, the health page told anyone who looked that a particular collection was undercounting its size and therefore making the ceiling look further away than it really was. It was not. That collection contributes nothing to the ceiling at all, by design and for a stated reason. It cannot make a total wrong that it was never part of. Both halves of the sentence were mistaken, and each in its own way: it also claimed the collection was contradicting a size it had declared, when in truth it has never declared one — what it was contradicting was a separate measurement taken months earlier, which is a different thing entirely. A wrong sentence on a status page would be a small matter if it only misinformed. This one did something worse: it sent the reader to fix the wrong thing. It pointed at a number that does not exist for that collection, while the number that genuinely needs re-measuring was named two lines further up the very same page, flagged as a stale estimate rather than a confirmed limit. Two halves of one fact, side by side, with nothing joining them. They are joined now. The deeper problem was that one sentence was carrying two opposite meanings. For a collection that does count toward the ceiling, undercounting really is urgent and really does eat into the safety margin. For a collection that does not, it changes nothing about the ceiling and belongs to a separate reserve with its own accounting. Those two situations produced identical wording, so a reader could not tell an urgent problem from a harmless one. They now read differently, they say which case applies, and they name the count behind the claim so anyone can check it. Both the human page and the machine-readable version are generated from a single sentence, so they cannot drift apart again. The second change is on the Dreaming page, and it is the same mistake wearing different clothes. VEILOS shows what share of its dreams have grown into knowledge it decided to keep. That share is calculated over a record which holds a limited number of entries and discards the oldest to make room — so a dream that came true and was later pushed out disappears from both halves of the fraction at once, and the percentage never registers that it happened. The page presented that as though it covered everything the organism has ever hoped for. It also had a quieter fault. When there were no dreams recorded at all, the page still printed a confident zero percent — which reads as a verdict on the dreaming rather than an admission that there was nothing to measure. "None of them worked out" and "there weren't any" are very different statements and should never look the same. The page now says which situation it is in. When the record has not yet filled up, nothing has been forgotten and it says so plainly — that really is every dream. When the record is full, it says that too, and stops claiming to speak for the whole history. When there are no dreams, it declines to give a percentage and explains why. And when it has no way of knowing where the limit is, it admits that rather than inventing one. Why did this survive so long? Because the same page has been honest about a different limit for years — it has always told you it is showing the last few nights out of a longer stretch. Seeing one boundary stated clearly makes a reader assume the others are stated too. That is worth remembering as a general lesson, and it is written down as one. Three faults in this release's own work were caught by VEILOS's internal checks before anything shipped, including one where the new dreaming calculation asked its question of the wrong list and would have claimed completeness it had not earned. Each was fixed rather than argued with. One question is deliberately left open and recorded as open: whether the separate reserve those uncounted collections rely on is still large enough. This release connected the right authority to that question. It did not answer it, and says so rather than implying otherwise. No new dependency, paid service, provider resource, destructive operation or cost increase was introduced.