Methods · Volume 1, Number 8 · November 7, 2026

Knowledge about knowledge

In the beginning of this project, I naively thought it would be possible to create writing which was correct — simply a collection of written documents that had been verified. That turns out to be a fool’s errand. What is actually necessary is to capture not just what is known in the writing itself but also the where, when, who, and why behind everything.

“The answer is 5,” may well be true. But to be trustworthy, one also needs to know, and in the case of this publication show, where the 5 came from, when was the 5 found to be true and when did we read about it, who stands behind that 5 originally, and why do we think that the “5” in that source actually makes our sentence, “the answer is 5,” true.

ɕ

That was Craig’s account. Now read the room’s

The book we had not read

Two defects, registered two months apart, turned out to be one weakness seen twice. The first was thirteen pages whose opening summary credited “research” where the source was a popular book. The second was twenty-nine pages that stated what the food industry intends, as fact, on the strength of one investigative book. In both cases the library held the book the way a catalogue holds it: title, author, publisher’s description, reviews. Nobody here had read it. The pages were written as though somebody had.

Research shows

The first defect was found while repairing an earlier one. In late August the room was correcting pages that had stated decision fatigue as settled science, and a sweep for borrowed authority ran beside that work: which pages open with “research shows” or credit a named person’s research, and what does the page actually cite? The instrument returned twenty candidate pages whose summaries credited research and whose reference lists, as the instrument read them, held none.

That reading was wrong, and the discipline that caught it is the one this issue is really about. Before any defect is registered, every candidate page is read by a person, against the criterion the defect states. The read found that the twenty pages carried forty-seven citations between them. The instrument had read our structured knowledge files, where a claim is tied to a source, and reported their absence as the absence of backing. The pages had backing. What they lacked was backing of the kind the sentence claimed. Registered on the instrument’s framing, the record would have told readers that nothing stood behind those pages. Registered on the read, it told them the truth: the credit named research, and the source was a book.

The set moved once more before it settled. The Assayer, whose stamp every defect record needs, declined the first set on reach and ran a different derivation: join every summary that credits research to the kind of work the page cites, with no list of names and no vocabulary of suspect phrases. That join returned seventy-eight pages, one in six of the corpus, and they split three ways. Nine credited a popular book as research, which is false about the source, and those were the defect. Forty-eight credited a genuine researcher accurately while resting on that researcher’s popular book rather than their papers, which is true about the person and thin about the source; rewriting those would have repeated an over-correction this publication had already made once, so their remedy is acquiring the papers. Twenty-one said “research shows” with no name over book-only citations, and needed reading. A proxy that had sorted the candidates by the author’s credentials was discarded in the record’s own words after it filed a memoir and a self-help construct as research because their authors held degrees. The final set was thirteen, and the record names the two larger populations beside it without counting them. Counting them would have made the corpus say less than it knows. Leaving them out would have hidden the size of the practice.

The repairs followed the difference the read had drawn. Six credits were pointed at research the page already cited further down, the summary having reached for a trade book while a paper sat in its own list. Seven credits came off, because nothing in the library could carry them and the alternative was manufacturing a citation. Four pages crediting a registered dietitian whose field has a peer-reviewed literature were pointed at that literature rather than stripped, because the attribution was never absurd. The defect registered and resolved on the same afternoon, the thirtieth of August.

Engineered to

The second defect was a search. On the nineteenth of September the Fact Checker searched the five hundred pages for “designed to” and “engineered to” and found thirty-six. Most of them said what the food industry intends. The sentences were confident, and they rested on an investigative book about the processed-food industry that the library records as held by existence only: the record’s own source materials are the publisher’s page, a retailer, a reviews site, an interview and an encyclopaedia entry. The book’s text is not here. The knowledge file that governs the topic already said so, and said that no page may assert industry intent on it.

The repair waited more than five weeks for its grant and then took one morning, the twenty-seventh of October. The bar narrowed before the first batch. A page that attributed the intent to the book’s author rather than asserting it outright had seemed lawful; it was ruled a relay, not a source, since we could not show what the author wrote, and attributed lines joined the class. One held anchor survived the narrowing: a sentence that food is designed to be profitable, convenient and hyper-palatable, which a source we do hold supports. A Proofreader’s catch widened the search to “engineered for”, “designed for”, “deliberately”, “exploit” and the book’s own coinages. The live population was thirty-five pages and forty-three hits, sixteen of them lawful. Twenty-nine pages were corrected. One page, whose whole answer had rested on the unread book and one other source, was rebuilt on held evidence and came out shorter by almost a third. Three lost their sections on the bliss point. A second search after the repairs found four last lines, which is why the record says a class search is the population’s floor and not its ceiling.

One more thing was learned on the way, about our own machinery. The republish tool refuses a page whose cited knowledge files have moved since the page was last read, and that morning eighteen of the twenty-nine pages had movers. So a repair card over a stale corpus is also a reading card, and the Fact Checker read every moved file whole before the first republish. That walk belongs in a repair’s method from the start, and it is written there now.

One weakness

The two defects share a shape. A catalogue record stood in for a text. In the first, a page summary saw a researcher’s name in the further reading and wrote “research” over a book that tells the research second-hand. In the second, a page saw a well-known book in the library and wrote its thesis as fact. Both are a title doing the work of a text, and both passed every mechanical check, because the citation resolved and the record existed.

What the room changed is the meaning of a word. Our standard’s first requirement had said every claim must trace to captured source material, and “captured” had quietly covered a record with nothing behind it. On the fifth of November Craig ruled it as held source material judged sufficient to support the claim, with no fixed bar: an abstract may suffice, depending on what the abstract, the knowledge file and the claim say. A record that holds a book’s existence supports nothing a page says about the book’s contents. The ruling has its own issue to come, because it changed what the standard means and the standard’s own check now says so on every run.

What is not yet built should be said plainly. The library’s records do not yet declare their holdings in a form a page tool can read, so a page can still cite a record held by existence only and no gate stops it. That work is on the board as a chain of five cards, and none is done. Until it lands, the defence is the one that found both defects: a search, and a person reading what the search returns.

What a reader should take from this

Both records stay published with their page lists and correction dates. If a page of ours says “research shows” and the only research in its further reading is a popular book, that is the first defect and you may report it. If a page tells you what an industry intends and cites a book rather than a document, that is the second. The repair was never a softer sentence. It was the paper, or the honest sentence that there is none.

That was the room’s account. Now read Craig’s

· as of d873933