The Reference We Asked People To Copy Was Punctuated Wrong

At the bottom of every lab note there is a small box that hands you a ready-made reference. It exists so that a writer, a researcher or an AI system can quote what we found and say where it came from, without retyping anything. It is the one thing on the page that is meant to leave the page.

It was punctuated wrongly. The reference in its current form was written on 17 August 2026, and the fault shipped with it, so it was wrong for six days. Whether the older format it replaced had the same fault we have not established, and we are not going to guess.

The reference is assembled from parts: our name, the year, the note’s title, the note’s identifier, the verdict, the web address. The assembler joined those parts with a full stop and a space. But a SleekNova lab note title is a sentence, and a sentence ends in a full stop of its own. So on every note whose title ended that way, the reference published two full stops in a row.

Measured across every published lab note: two hundred and twelve of seven hundred and fifty-two. On the passed product notes alone, one hundred and six of three hundred and eighty-seven.

What is worth saying about this one is structural rather than historical: the reference box is the only element on the page that is read somewhere else, by someone we never meet, in a document we never see. It is the hardest thing here for a reader to report back to us, however long it is wrong for.

The fix is in how we join the parts, not in anyone’s title. The assembler now looks at what the previous part ends with. If it already ends a sentence, the join is a space. If it does not, the join supplies the full stop. Nothing stored changes, and no title is trimmed.

Our first attempt at the fix was worse and we caught it by testing it against the real library. That version stripped the full stop off the title before joining. It worked for full stops and broke for question marks: one of our notes is titled with a question, and stripping only full stops would have published a question mark followed by a full stop. Testing a fix against every record we hold, rather than against the case that prompted it, is what turned a narrow repair into a correct one.

The second correction, on the same pages

Search engines and AI systems do not read a page the way you do. They read a structured summary underneath it, which names what the page is, who wrote it, and what it is about.

On our lab notes, the “what it is about” was missing. Not wrong: absent.

A rejected product note carries the product’s name and brand as its own stored fields, because there is no product page to point at, and those fields fed the summary correctly. A passed note is the opposite case: the product exists on this site, and the note is linked to it. That link was never read into the summary. So on three hundred and eighty-five of these pages, a reader saw the product’s photograph, brand, name and price, and a machine reading the same page was told the article was about nothing in particular. The score is on that card too now, but it arrived with this same update, so it was not part of what the machine was missing.

Three hundred and eighty-five of those pages now name their subject, read from the product they are bound to. Two do not, and both are honest gaps rather than oversights: one has no bound product at all, and one is bound to a product our own records cannot currently resolve. Where the product cannot be named, the page now says nothing rather than naming an empty thing, because an empty name is a statement and absence is not.

What we are not claiming

We do not know how many references were copied while they were wrong, and we are not going to estimate it. We have not corrected copies already living in other people’s documents, because we cannot reach them. And the shape of the identifiers themselves is a separate defect we have not fixed here. Some of these notes store a typed label inside the identifier field, so the reference prints that label as though it were part of the number. We are not publishing a count of them. The only measurement we hold was taken before a related repair landed, and a figure we cannot stand behind today is worth less than saying plainly that we do not have one. That is a fault in the stored data rather than in how we print it, and printing around it would have hidden it. It is logged and owned, and it will be corrected in the data.

Both defects on this page were found by us, while building something else, by reading our own output instead of our own code.