September 2026

What was wrong. The Quillery’s What this is, and what it is not note, live since 2026-08-15, read that Kaya Oldaker’s artwork “is not ours to redraw stroke for stroke.” The Easter Eggs card called the sheet “the first about somebody else’s character,” twice. A CSS comment in the page’s own source said the palette was “not ours to fold into our own.” And the FACTCHECK.md row asserted “Esmx the Porkypine is Kaya Oldaker’s” — in the same sentence that recorded the piece as commissioned by Stimpunks.

What is right. Stimpunks commissioned Esmx and holds the rights. Kaya Oldaker is the artist. The CC BY-SA 4.0 licence on the mascot is one we granted rather than one we are subject to, and the invitation attached to it — feel free to customise! Make variants of Esmx to represent your community — is ours as much as anybody’s. The eight drawings were rebuilt from scratch rather than traced, which was always a good decision and is now stated on its actual grounds: a pasted PNG would be the only drawing in the collection not in its own hand, and a variant is the thing the licence goes out of its way to ask for rather than merely allow.

The credit does not move, and that is the half that was doing the work. Kaya’s name stays under every drawing, in the colophon, on the collection card and in the structured data. Holding the rights to a commission is not a reason to stop naming the hand that drew it. The one thing that changed in the structured data is an addition: isBasedOn named a creator and never a copyrightHolder, so the machine-readable version of the page had the same gap as the prose.

Why nothing caught it, which is the part worth keeping. Every rule in FACTCHECK.md is tuned to catch a claim that overreaches — an attribution taken from a summary, a metaphor promoted to a proof, a number gone stale. This claim under-reached. Told that a drawing came from a named artist, the sentence reached for somebody else’s, handle carefully and never asked who had paid for it; and the result reads as scrupulousness, so no gate and no reader flagged it for five weeks. It is the mirror image of the failure Too Good to Check catalogues: not a story too good to verify, but a caution too flattering to question. Being careful in the wrong direction is still being wrong, and it is harder to see, because the tell that usually gives an error away — that it makes us look better — is present in exactly the same way.

Caught by Ryan Boren on 2026-09-20, while stimpunks.love was building an arcade room around Esmx and copying this page’s framing into it. A wrong sentence had already travelled one site before anybody read it twice, which is the argument for correcting it here rather than only there. The Quillery now records the correction in its own Notes, beside the citation it nearly got wrong the first time.

Not affected, and checked rather than assumed: The Autism Infinity Dragon, also Kaya’s and also on this site, is her own work under her own public-domain release — not a commission, and the open item about stimpunks.org over-restricting its licence stands unchanged. The ledger drew that distinction correctly for the dragon on the same day it got Esmx wrong.

The hold that earns the section is a hundred-year correction whose evidence is nearly two years old. Textbooks may have misdrawn this basic brain structure for 100 years rests on two papers from one laboratory: the headline finding, that unmyelinated axons are strung with 200-nanometre swellings rather than being smooth cylinders, was published in Nature Neuroscience on 2 December 2024 — 652 days before the release that carried it — and its companion in Neuron was registered 295 days ago. We read the 2024 paper in full and held the story anyway. A digest whose entire axis is how long did that take to reach you cannot file a two-year-old finding under a present-tense headline.

The second wanted hold is a physicists’ survey, and it was chased from this page’s own unchecked leads. Last week’s edition listed a survey of researchers’ beliefs as “an epistemics item rather than a discovery, and possibly the most useful of the four” and did not open it. A sibling arrived this week and was traced — the release names no journal at all, and the source turns out to be an arXiv preprint found by reading the author’s own listing. Its finding is that several positions described publicly as field-wide consensus are held by narrow majorities or pluralities. Held, because a preprint read at its abstract is not a version of record, and that rule was written here last week; bending it in the first week it costs something would make it decorative.

Four more held, each for a nameable reason: four separate semaglutide items on one feed in seven days, the loudest of which drops both words from its paper’s own title, female mice; a Venus moon that the paper’s title calls hypothetical and the headline calls swallowed — the same “parse the title’s subject” trap this page recorded against itself a fortnight ago; an exploratory crossover in 23 men whose mechanism, the fasted state, does not survive into the headline; and a genuinely extraordinary Cretaceous feather preserved in dinosaur dung, whose closing claim about which lineages survived the extinction rests on one specimen and would lose its hedge in a single retelling.

Three of our own sentences were cut before publishing rather than after, and are logged here because the ledger is where they belong. A draft of the inductive-effect entry asserted a student count that nothing supports — an invented number, in a frame arguing against numbers fitted to expectations. A draft of the muon entry gave a flux figure without marking it as an approximate textbook value not taken from the paper. And a refusals line described the semaglutide result as “twenty female mice-months,” which is not a unit and not what the study did. None reached the site.

Where the coverage was right, said plainly: ScienceDaily’s summary of the Earth-accretion paper — “less than two percent of Earth’s total mass, and possibly none at all” — checks out exactly against the paper’s own bound. Naming that is part of naming the misses fairly.

The first was found by reading the spread, not by any of our gates. Ryan asked what “flat” meant on spread eight — a fair question, since the spread never said — and the answer turned up something worse. A draft paragraph read “positive means the geometry closes on itself…so space is finite,” while the figure directly beneath it labelled the negative side closed, and finite. The page was arguing with itself, and not one of the eight checks that guard this site can see that.

The figure was right. The quantity is written Ω_K and it runs opposite to the curvature: Planck’s own paper says the pull is “towards negative values of Ω_K” and that “closed models predict substantially higher lensing amplitudes.” So negative is closed and finite. The trap underneath is that two standard quantities carry opposite signs for the same universe — a paper reporting “positive curvature” and a table reporting a negative Ω_K describe the same thing. The spread now defines flatness in plain language first, then states the convention outright.

The second was found while checking the first. The draft cited −0.0106 as the figure from Planck’s spectra alone. It is not: that equation already includes the lensing map, and the spectra-alone figure is −0.044. That mattered more than a mislabel, because adding the lensing map is precisely the step the paper says pulls the answer back toward flat — so the wrong label hid the mechanism. Correcting it replaced a two-number standoff with the three-step walk the paper actually describes: the spectra alone give −0.044 and put about one sample in ten thousand at flat or beyond, the lensing map moves it to −0.0106, and the oscillation data lands it on 0.0007. Each dataset added pulls it toward flat, and the figure now plots all three.

Neither error reached the public site, so by this house’s own rule — a mistake corrected before shipping is a fact about the build, not disclosure — they are logged here and in the ledger rather than confessed on the zine itself. An earlier draft did confess them on the page, and those sentences were removed. What stays on the page is the part a reader can use: the sign convention is a real trap in the literature, stated as such.

Three more defects were caught by looking at the figures at the size they actually render, which is the only instrument for this and is not optional: an arrow pair that drew contraction while the prose beside it said expansion, a panel showing five points against a caption claiming the same people where the facing panel had nine, and a label detached from the dot it named. All eight gates passed all three, and each was right to — a technically perfect picture of the wrong thing is still valid markup, legible, correctly coloured and in the right place.

And the source is a lead, not a citation, which this piece states on its own face. The Brian Cox compilation that prompted it is an automatic transcript of clips of several vintages: it gives the age as 13.75 billion years in the older segments and 13.8 in the newer, renders branes as “brains” throughout, and turns “ten with 120 noughts after it” into “knots.” Nothing is quoted from it and no figure is sourced to it. Two limits are also printed: the Nature Astronomy paper arguing the other side of the curvature question is paywalled, so only its abstract phrases are attributed; and Friedmann and Lemaître were confirmed at their bibliographic records rather than read in the original languages.

The Dolly rack told search engines and agents it was called “One Body at a Time.” That is the David Byrne rack’s title, copied across when the page was built. Its visible title, its tab title and its card were all correct; only the machine-readable headline was wrong, which is precisely why it survived — there is no rendering for it to get wrong. And the Sound collection’s own structured data listed seven of its eight racks, omitting the newest. Both fixed.

Our markup gate passes both of these and is right to. It asks whether a structured-data block parses and whether it declares a type; it cannot ask whether the title inside it is this page’s title. A check that compared the two would have caught both in a second, and is worth writing.

One sentence elsewhere stopped being true and was corrected on two pages: It Take a Joyful Sound was described as “the one that argues first,” which read as the L★S rack does not argue. It now says what is actually the case — that it was the first rack here to carry an essay.

The design system opened by saying every artifact here is one self-contained file “with no dependencies beyond Google web fonts”. On 9 September the four typefaces were moved onto this domain, precisely because that dependency was handing every reader’s address to Google before a word appeared — and the privacy page written that same day says so at length. The design page was not updated with it. For four days the site told readers two different things about itself, and the page still claiming the dependency is the page whose whole subject is how the thing is built.

It was swept rather than patched, because fixing only the instance somebody noticed is how a correction inherits the original defect. Four more copies of the claim were live: the repository README and the working notes both still specified what “the Google Fonts request must be”, down to the weight range — instructions for reintroducing the exact third-party request the privacy page says is gone — a decision-log entry still listing Google fonts as the one permitted dependency, and a comment in one of the checking tools naming the Google stylesheet as an expected cross-origin fetch. All five corrected. The lesson each carried was kept: the old font request asked for a weight range the typeface does not publish, so it returned an error and no font at all, and nobody noticed for weeks because the font was installed on the machine doing the checking. That is about verification, not about Google, and it outlives the mechanism.

The replacement claim was checked against Google rather than against our own notes. Saying the self-hosted files are the same set Google used to serve is a provenance claim, and our working notes asserting it are not evidence for it. The site’s own font request URLs were recovered from the commit before the move, fetched from Google as they stood, and compared face by face against what this site now ships: 20 against 20, identical in family, style, weight range and character subset, with nothing added and nothing lost. The four families are OFL 1.1 and unmodified, each licence shipped beside the files.

What was deliberately left alone is the other half of this. The privacy page’s “until 9 September 2026, every page loaded its typefaces from…”, the migration account further down the design page, and every earlier entry on this page are all correctly in the past tense. They are dated records of what was true when written, and editing them backwards would be tidying away the evidence — which is the opposite of what this page is for. Only claims still written in the present tense were touched.

The first attempt moved 59 rules and changed 2,878 elements across 53 pages. The reasoning had been: a page that writes its own version of a rule overrides the shared one, so sharing is safe. That is true per property, not per rule. A page whose version set the colour but said nothing about letter-spacing kept its colour and silently inherited a letter-spacing it had never had. Forty-seven rules fail that test and stayed where they were.

The second attempt moved 27 and changed 52 elements, all on the two scroll zines. One rule — the one that shows the current spread — is more specific than a scroll zine’s own layout rule, and would win. No reader would ever see it, because nothing marks a spread “current” on a page that scrolls. But four of our own measuring tools add that mark in order to see every spread at once, so sharing the rule would have changed what those tools look at. Three rules dropped for touching state the tools inject.

What shipped: 24 rules, verified at zero change. Every computed property of every element on all 207 pages, before against after — 140,667 elements, not one difference — with the “before” copies pinned to a frozen duplicate of the shared sheet, so the comparison measured the move and not the sheet swap. Both wrong theories sounded right and neither survived being rendered.

A shared pager that loads is not a pager that pages. So every one of the 107 pages was driven in a real browser: start on spread one with the back button disabled, click forward and check the spread, the counter and both buttons, press the right-arrow key, walk to the last spread and check the forward button disables, then follow a #spread-4 link and land on spread four. 107 pages driven, none failing.

Then the same test again with the actual security policy served, because this site has learnt that a policy tested only against attacks is untested: a policy that breaks a hundred page-turners is worse than none. Pages paged, zero violations reported.

One test failure was the test’s fault and is worth naming. The page that keeps its own pager was reported as failing the deep-link check — because it has never had deep links, that feature living in the shared script it does not load. The assertion was demanding a feature the page was never offered. A test that does not know what a page is for will confidently report it broken.

First, it did not know about the shared script. Pages get some of their classes at load time, from the one JavaScript file the whole site shares — the pager controls on every zine, among others. Those classes are not written in the page, so the first version of the audit called their styling dead: 87 live rules, on the pager of every paged zine. Acting on that run would have broken the thing readers touch most.

Second, comments were being read as part of the rule that followed them. That hid any block introduced by a comment, and — far worse — meant a comment that happened to mention a class could condemn the perfectly live rule underneath it. On a site whose stylesheets are heavily commented, that is a loaded gun. Comments are now parsed as their own thing.

And the verification was wrong before it was right, in the same direction as yesterday’s correction. The first before-and-after comparison reported 3,611 changed elements, which looked like a disaster and was an artefact: the browser lists an element’s custom properties in a different order once rules are removed, and one animated panel was sampled at two different moments of the same fade. Sorting the properties and freezing animation gave the true answer of zero. Three times in two days a measurement has been wrong in a way that looked like a finding.

What we said, yesterday, on this page and in two working files. That the cover artwork lay across cover type on “every one of the 98 covers that carry any: over the title on 95, the issue line on 94, the subtitle on 93.”

Those came from the artwork’s bounding box, not the artwork. A cover motif is an SVG, and the drawing rarely fills its own frame — the median motif here inks about 30% of the box it sits in. So a box can sit squarely on a paragraph while every visible mark clears it by a wide margin. Re-measured by ink — the union of the shapes and labels actually painted — at the same phone width, on the same pre-fix state: 98 of 99 covers, with the issue line crossed on 90, the title on 84, the subtitle on 23 and a byline on one. The subtitle figure was wrong by seventy pages.

The error has a precise shape and we had it written down already. Our collision checker measures where the letters are, and its own notes explain why at length: an early throwaway version compared bounding boxes and, in its author’s words, drowned in false positives. That tool did not make this mistake yesterday and does not make it now. A one-off probe written beside it did — and the subtitle claim happened to be true on the one cover we checked by eye, so a real observation on a single page got generalised across ninety-eight by a method that could not support it.

What still stands, stated plainly so the correction is not read as wider than it is. The count of covers affected is unchanged at 105 of 108. The text-on-text figures — 313 collisions across 103 pages — came from the checker’s own ink measurement, not the bad probe, and are unaffected. The fix shipped yesterday is correct and is verified at zero findings on both measured widths. Only the artwork figures were wrong, and only because of how they were counted.

The zine will not end on the standard as a happy conclusion, and the source for the reason is one we could not open. Once contributions became recordable they became countable, and Larivière, Pontille and Sugimoto used contributorship statements to measure how scientific labour actually divides. The citation is confirmed at Crossref; the publisher returned 403 and we did not read the paper. A search summary offered us specific percentages, and a search summary is not a citation — so the spread states the direction of the finding and quotes no figure at all, and the colophon asks anyone who can reach it to send the numbers. A vocabulary makes work visible; it does not make it valued.

And the headline figure has two versions. Nature’s report gives 5,154 authors; another source gives 5,153. We did not count the list ourselves. Rather than average them or pick one quietly, the zine prints Nature’s figure and states the discrepancy. The two-column arrangement of the fourteen roles is also ours rather than the standard’s, which publishes one flat alphabetical list and ranks nothing — and the spread says so on its face, because presenting an editorial reading as a published norm would be self-refuting in a zine about attribution.

One link came out before shipping. A draft linked the repository’s fact-check ledger as a reader-facing document. Nothing on the site links it, it is not in the sitemap, and it is raw markdown with no navigation — making a zine the first public door to a maintenance file is an editorial decision rather than a citation, so the sentence now describes the ledger and links the changelog instead.

The zine shipped with one source unread and said so; hours later Ryan put the PDF in front of us and spread seven was rewritten from it. The hedged version is gone and the figures are on the page: 30,054 PLOS papers across 2017 and 2018; 57 per cent of women contributed to investigation against 49 per cent of men, with a similar gap for data curation; men 30 per cent likelier on funding acquisition and supervision.

And the paper contains a result that argues this zine’s case better than the zine did. The older, coarser scheme PLOS used had a single category for writing, and under it men appeared to dominate. CRediT splits writing into original draft and review & editing — and the picture inverts: women are 6 per cent likelier to have written the original draft, men 8 per cent likelier to review and edit it. The authors call the inversion striking and attribute the earlier finding to how ubiquitous the reviewing half is. So a coarse vocabulary did not merely fail to record something; it produced a false picture that a finer one corrected — which is the zine’s whole thesis arriving as a measured result rather than a claim.

The method’s limit went on the page with the finding, because the finding rests on it: gender was assigned by algorithm rather than self-report, reaching 82.2 per cent of authorships and reported as 98.3 per cent precise for men and 86.7 per cent for women. A measurement less accurate for the group a finding is about is worth stating out loud. It does not overturn a gap that size; it sets the resolution.

Jocelyn Bell is on the paper. Second of five, Nature 217, 709–713. The claim that she was left off it circulates constantly, it is false, and it is checkable in ten seconds — so anyone hostile to the subject can knock it over in one line and take the genuine complaint down with it. The card puts that in its first row rather than quietly omitting it. Sharpening a true grievance with a false detail is the same offence as the grievance.

And Hedy Lamarr did not invent Wi-Fi. This is the most repeated story in the genre and it does not survive the patent record. Frequency hopping has prior art in Tesla’s patent applied for in 1901 and in Zenneck’s 1908 book, which notes Telefunken had already tried it; Ellison Purington filed a year before she did. BLADES, the first frequency-hopping system actually deployed, cites neither her nor Antheil — nor does the field’s first textbook across nearly two thousand references. And 802.11 Wi-Fi does not generally use frequency hopping at all. What is true is quieter and still worth saying: the patent is real, it is US 2,292,387, and it is filed under Hedy Kiesler Markey — a name she is not known by, which makes the document itself a missing name. A props culture that overclaims is not a props culture.

Three cards carry no reframe table, following the rule FG 10 set and FG 18 followed. Henrietta Lacks is one: tissue taken without consent in 1951 is not an attribution dispute, and describing it as one proposes a remedy — add her name to the papers — that addresses none of the injury. Her card’s diagram is the only one in the guide built out of no name-slots at all. Ada Lovelace is another, because the historiography is genuinely contested in both directions and the flattering reading is the tempting one. And the third is the survivorship problem: every case here was recovered by somebody doing the work of recovering it, so eleven is not a sample.

Same run, opposite error, and worth publishing because it is the more dangerous of the two. Alongside the real collisions the checker reported 256 pieces of text as cut off — sliced by the edge of the page with no way to reach them — across twelve pages of working notes. Every one was a wide table on a narrow screen, and every one of those tables sits in a wrapper you can swipe sideways. Nothing was cut off. The reader reaches all of it.

The tool was looking for a container that hides overflow and walking straight past containers that scroll it, then blaming a rule much higher up the page that exists to contain the starfield. It now stops at whichever container takes responsibility for the overflow first, and answers differently depending on which kind it is. It also prints how many pieces of text it declined to report for that reason — because that is precisely the number it used to get wrong, and a silence nobody counts is a silence nobody questions.

Both halves were tested by breaking them. With the cover fix removed the checker reports the collisions again on the phone width and stays clean on the desktop one; with it restored, both are clean. A page that genuinely cuts text off still fails; the same page with a scrollable wrapper passes. A green gate is only evidence if you have watched it go red.

We had the famous image and it is not in the source. The line everybody repeats — a quarter of an inch of scruff in 400 feet of paper — was to be the spine of the whole zine. Her own 1977 account gives no such figure; what it gives is 96 feet a day and “the 3½ miles of chart recordings from the radio telescope that I analyzed.” Those two agree with each other, and they are hers. The quarter-inch is gone, and the zine says on its own face that it was cut and why. A secondary source had also given us 98 feet a day; the primary says 96.

And the argument itself was wrong. This piece was proposed as she was robbed, and it was unfair. In the same speech she declines that reading in print, with four reasons, ending “Finally, I am not myself upset about it—after all, I am in good company, am I not!” A zine asserting her grievance would have contradicted its own cited source on the page. So the credit argument was rerouted through what she does record as absurd: the press, who on discovering a woman was involved “descended even faster”, photographed her “examining bogus records”, and asked “how many boyfriends did I have at a time?” The child gets the harder and truer distinction instead — being kind about something is not the same as the something being fair — and is told she may reach her own verdict. So, arguably, did Bell Burnell: in 2018 she gave away the entire £2.3 million of a later prize to fund students who get left out, disabled students named among them.

One more we are keeping on the page rather than quietly fixing. The circulating story in which she was left off the paper is false — she is second author — and the zine insists on her by-line rather than dropping it to sharpen the grievance. Sharpening a true complaint with a false detail is the same offence as the complaint.

This piece started somewhere else entirely. Last week’s Glimmer Wire graded a September 2026 result in which an ordinary bacterial enzyme transcribed a synthetic eight-letter alphabet as readily as the natural four. Ryan read it and asked for a Young Readers primer on the genetic alphabet.

The wire’s own filing note said not to build a zine on it — that the argument needs the biology to stay in the test tube, and a zine would want it out. That judgement was followed rather than argued with. The synthetic alphabet appears nowhere in the finished piece. A read-aloud ending on scientists made new letters sits one small mishearing away from scientists made new creatures, and the 1977 pond virus carries the same four is a habit, not a law claim without ever leaving the water.

One source is declared unread on the page itself. Kirnos, Khudyakov, Alexandrushkina and Vanyushin (Nature 270, 1977) is behind an authentication redirect; the substitution and its completeness are taken from the 2021 literature citing it, read at full text, and marked as such on the spread and in the colophon rather than carried as the 1977 paper’s. And one number is deliberately not quoted at all: the 2021 papers disagree about how many Z-genome phages are now known, and an unreconciled figure is worse than none.

The Franklin acknowledgement is quoted in full and left uncommented. Watson and Crick recorded being “stimulated by a knowledge of the general nature of the unpublished experimental results and ideas of Dr. M. H. F. Wilkins, Dr. R. E. Franklin and their co-workers at King’s College, London.” How thin that sentence is, is the point, so the grown-up layer reproduces it exactly rather than characterising it. The child layer names nobody.

Two claims the zine refuses to make, both named on its own refusals spread. The pond virus is an exception in chemistry and is not an argument about people — reading it as one is the naturalistic fallacy with a nice feeling attached. And the piece declines the obvious ending, nobody else has your order: it is not quite true for identical twins, and worth that rests on being rare is worth that somebody similar can take away.

The last three links go where this house lives, and the fairness in them is structural rather than decorative. In 2018 the retired metaphor was the title of Robert Plomin’s Blueprint. The objection the chain quotes is Kathryn Paige Harden’s, a working behaviour geneticist — “Insisting that DNA matters is scientifically accurate; insisting that it is the only thing that matters is scientifically outlandish” — chosen precisely so the chain cannot be read as an argument against heritability, which it is not. The spread also records that others defend the book as misread, and that we do not adjudicate.

Link 11 is the one we are inside, and it is deliberately not a villain story. Spectrum 10K opened in August 2021 to collect DNA from ten thousand Autistic people, was paused within weeks after our objections, consulted on for years, and closed in 2025 with Wellcome’s support. Proposed, objected to, consulted on, ended — that sequence is the thing working, not failing. The chain asks only why the objection had to be that loud.

And the last link is a leap we refuse to cross. The tempting sentence is the blueprint metaphor caused the gene-hunts. We cannot show that. Did the picture make the search thinkable, or did people who wanted to search reach for a picture that suited them? Both fit every fact in the chain. What the zine will say without hedging is narrower: the picture is wrong — on the gene count, on the onion, on the 2010 review, and on the testimony of the man who coined it.

Two things declared rather than hidden. Link 5 is graded contested because we could not source a first use of “genetic blueprint” — a grade given for our own sourcing failure rather than a live dispute, which the spread says outright while asking readers to correct it. And Plomin’s book was not read; its thesis and the responses to it come from reviews and reference summaries, which is the weakest sourcing in the zine and is labelled as such on the page.

No. 106 shipped naming Pigliucci’s 2010 paper as a lead it had not read. It has now been read at full text, and this zine is what came of it. Its verdict is blunter than ours: the blueprint and programme metaphors are “not only woefully inadequate but positively misleading.” The zine tests one of the four failures he names and lists the other three for the adult.

Three of four sources read directly; one figure marked second-hand on the page. The onion assembly, the worm genome and Schrödinger were read; the Allium genome-size range of roughly 7 to 31.5 picograms was not traced back to the measurement papers and says so where it is used.

What it refuses. No position on whether non-coding DNA has function — that argument is live, the onion comparison comes out of it, and the zine needs only that raw length is the wrong ruler. No swap to recipe, which keeps a dish somebody is working towards. And no argument about people anywhere in the child’s layer: every measurement is about a plant, a worm and a gene count, because a fact about an onion licenses nothing about anybody.

The half that is missing is deliberate, and it is the better piece. How a qualifier falls off between a 1944 founding text and an ordinary sentence about a child is a chain with graded joints, and it belongs in How We Got Here rather than in a read-aloud. It is planned. One refusal points at it: a plan is the kind of thing that can have a mistake in it, and that is what made the word useful to people who wanted to search a human being for the error.

What is true: PubMed Central’s web reader does serve that article as page images, and our fetcher could not read them. What we then said, and should not have: that the paper was therefore unreadable.

The PDF has a text layer. Fetched from Europe PMC, the fifteen pages come out as searchable text — imperfect OCR, but plainly legible — and the result the zine would have used is two lines of it: “the addition of 10 µg of polyuridylic acid per ml of reaction mixture resulted in a remarkable stimulation of C¹⁴-L-phenylalanine incorporation… no other polynucleotide tested could replace polyuridylic acid.” A tube of nothing but one letter, producing a chain of nothing but one bead.

The failure is not the blocked route, it is stopping at it. Our own working notes name Europe PMC as one of the metadata-and-full-text routes to try when a publisher or a repository refuses, and that note was written after a previous paper hunt. It was not applied. “We could not read it” is a claim about the world, and it needs the same standard as any other claim on these pages — which is that you try the routes you already know about before writing it down.

The spread stays cut, and now for an honest reason. Twelve spreads is the right length for that zine and its argument does not need the history — so it is cut by choice rather than by a wall that was never there. The colophon now says so, quotes the result, and records that we got this wrong.

The page said seven of the ten for weeks. Adding Zine No. 105 on 12 September, we could not reconstruct that seven from the page’s own named sets, re-derived it as six of the eleven, and shipped that. Adding No. 106 hours later we carried it to seven of the twelve. Both of those were ours and both were wrong.

The cause is not carelessness, and it is worth naming because it will happen again elsewhere. Two different sets were being described by one number. Something to do is wider than a demonstration: No. 102 has no demonstration — you cannot do a whole autumn on a table — and it still hands a child a method to carry out, which is to mark one leaf and look at it every day. It belongs in one set and not the other. The two lists differ by exactly that piece, and a count that does not say which set it means has nowhere to stand.

The original seven does reconstruct, and only on a reading the page argues against. It works if No. 95 is counted — and that zine’s “push on the bone, the bone answers” is the one case the page keeps deliberately on the far side of the line, because the same zine says the current is “not enough to feel.” So the figure and the boundary paragraph beneath it could never both have been right. That inconsistency predates this week; our two corrections inherited the ambiguity rather than the error.

What changed, beyond the number. The eight are now written out by name rather than counted, the paragraph says out loud why the two sets differ, and the list below it now reads “no demonstration” rather than implying the same four pieces hand the child nothing. No gate can see a number written out in a sentence — this site’s oldest standing lesson, and the reason the fix is to name members rather than to correct a digit.

Every codon on the page was checked against NCBI’s own translation tables rather than recalled, because this zine is the first here whose demonstration fails if a single row is wrong. A child following spread seven with one bad row gets the wrong answer and learns that the book lies.

Two sources are declared unread on the page itself. Freeland and Hurst’s 1998 result that the standard code is extraordinarily good at limiting the damage of errors is carried at second hand — PubMed refused us — and Crick’s 1968 paper is quoted from the same open-access review. A third, Nirenberg and Matthaei 1961, the experiment that read the very first codon, had a spread outlined and then cut because we believed PubMed Central held it only as scanned page images. That reason was wrong and is corrected below.

The famous result is not allowed to travel without the half that usually falls off. The same review that gave us the figure puts the code “halfway from an average random code to the summit,” notes that billions of alternative codes would do better, and finds Crick’s rival account — that the code is a frozen accident, arbitrary and stuck because any change now would be lethal — essentially unfalsified. So the zine does not say the code was arranged kindly, or arranged at all. It says the checkable thing: there is more than one right spelling, so a great many changes change nothing. Why that is so is an open argument, and the page does not borrow its warmth.

One departure from its companion, made deliberately. No. 105 stayed safe by never saying what DNA does. No. 106 cannot, because the code is what it does — so spread ten corrects the popular meaning of “genetic code” in the child’s own layer, in plain type: it says which bead comes next, and holds no entry for what anybody will be good at, or love, or be like. And one overclaim was caught during the build: synonymous does not always mean no difference, since codon usage can affect how a chain is built and folds. The child layer claims only what the table says, and the grown-up box states the limit.

Anyone who opened What’s New handed their IP address to Google before a word rendered — exactly the thing the self-hosting pass existed to stop, and exactly the thing the privacy page says no longer happens. The page is otherwise unremarkable; the reason it survived is the interesting part.

A generated page does not conflict. It reverts. whats-new.html is written from a template inside tools/build-derived.mjs, so the 9 September sweep edited the file, and the next routine rebuild quietly wrote the Google links back — with the tool reporting a clean, successful build. This site already had that failure on record. The same tool did the same thing to this page’s skip link on 9 September, and the lesson written down then was that a repo-wide sweep owes the template the same edit. The sweep that followed did not pay it.

Found while reading a diff for something else, which is the honest account: no gate looks for a third-party font request, and none of the nine can. The two other pages that match a search for those hostnames — this changelog and the privacy page — only mention them in prose, which is the difference between a thing and a sentence about the thing, and is why a raw-text search would have called all three guilty.

The template no longer carries them, and the lines have been replaced with a comment saying why, so a future copy does not restore them. Verified by loading the rebuilt page and watching the four self-hosted woff2 files arrive from this site — not by reading the stylesheet, since both families are installed on the machine that builds this and a silent fallback would have looked identical.

Ryan proposed this piece with a second claim attached: that sitting near the inner edge is part of why we are sensitive to carbon dioxide now. Checking it against Goldblatt & Watson (2012), read at full text, produced a clean refutation — “adding carbon dioxide does not increase the outgoing longwave flux, so it cannot cause a runaway greenhouse” — and on that basis the climate thread was cut from the outline entirely.

That was an over-correction, and it is worth naming plainly. It answered runaway greenhouse, which was not the claim being made. Ryan came back with astrophysicist David Kipping’s version: the zone migrates outward as the Sun brightens, so the headroom shrinks. That holds. Scaling the 2013 limits by the square root of luminosity puts Earth about 13 per cent across the band two billion years ago against 1.4 per cent now — the same planet, a ninefold smaller margin. The thread went back in, and the zine is better for it: spread seven carries the mechanism, spread eight carries the distinction and names the misapplication out loud.

Kipping is credited here as the lead that prompted the re-check and is not cited on the page. The figures there come from the papers. A video is a lead, not a citation — the same rule this site applies to aggregators, applied to a source that happened to be right.

Everything else, graded on its own face. The cloud-free hedge is printed verbatim on spread five, because Kopparapu’s own abstract says the boundaries “may extend further in both directions” and cloudy models since have argued the inner edge moves substantially inward. The band positions for two and four billion years ago are our arithmetic and are labelled as ours on the spread itself. Gough (1981) and Rushby et al. (2013) are marked cited-not-read and read-at-abstract respectively. The tulip figure is extension-service consensus rather than a primary paper, and says so. And caves, campfires and watering holes are credited to David Thornburg, with the count stated as four — the fourth, life, named in the grown-up box rather than quietly dropped.

Five held, and the first one is about us. On 4 September this page filed the bat immunoglobulin paper, read it in full, and corrected one figure: the widely repeated more than 500 vesper bat species. The paper measures 26 species with duplicated antibody gene loci and confirms both functional in one; it gives the family no number at all. A release carrying the old figure surfaced on this week’s scan, unchanged. A correction that lives only on our own page is a correction that has not happened — so it is held and recorded rather than passed over. In fairness to that release, we went back to the paper and its other numbers check out.

Held second: a mouse study from February 2022, carried on 2 September 2026 under a headline in the present tense, crediting the discovery to Stanford when the author list is led from Shanghai Jiao Tong University. Also held: a magazine feature about a decade-long bird-cognition programme whose headline result is a 2024 paper, and which reads as a merit argument if lifted; two papers on cortical development folded into one headline, with both publishers refusing us so we could not tell which finding belonged to which; and three large observational studies reported as percentages a reader might act on, which turns a question about prescribing and access into a personal decision. That last decline is the second time this collection has made it, and declining twice is the point of keeping the list.

What we got right by checking our own sentence rather than the coverage’s. A correction had been drafted saying the cosmology summary’s “1,700+ supernovae” confused light curves with supernovae. The paper itself says 1,701 SNe Ia. The draft correction was deleted before publication, which is the discipline the second edition’s amendment pass bought us.

We checked the names against each vendor rather than copying a list, which the reference we were working from explicitly tells you to do — and it was itself stale. Anthropic now publishes two crawlers that list does not mention, and OpenAI has added a third. All three are named in our file. Anthropic's own documentation has moved host, too.

Separately, and worse: Young Readers had shipped with broken structured data on 6 September and it was live for four days. The page's description quotes a caregiver asking “is that really true?” — and the straight quotation marks around it ended the machine-readable block four words in. Search engines and agents got nothing typed from that page at all. Nine automated checks passed it, and each was right to: a broken block of that kind has no appearance, so nothing about the page looks wrong.

It is the same fault as one we fixed last week — an unescaped quote cutting a value short — in a different container, which is why the check written for that one could not see this one. There is now a tenth check that parses every such block on every page, tested by restoring the real breakage rather than a synthetic one. 198 blocks, 198 parsing. A fix applied to one container is not a fix to the mechanism, which is the second time this week that sentence has been the lesson.

Yesterday's audit note claimed that the 401 collapsible descriptions across the front page and the collection pages were invisible to find-in-page, and listed fixing them as the next job. They were already fine. Those use the browser's own disclosure element, which has handled this for years — and the specification we were auditing against says so in as many words: prefer the native element, because it gives you find-in-page reachability with no code at all.

Checked before touching anything, which is the only reason this is a paragraph rather than 401 needless edits: the text is present in the page, three accessibility-tree nodes carry it while the card is shut, and the browser's own find call locates it. The lesson is the same one as yesterday morning — an inference about what a reader can reach is worth nothing next to a measurement of whether they can reach it.

Every page here carries an HTTP header pointing at the things a machine might want — the index for language models, the feed, the licence. Our working notes said “every relation is IANA-registered; inventing one is a bad signal and crawlers ignore it.” We fetched the registry to add a new one and read all 236 entries. sitemap is not in it. security is not in it. Both had been going out on every response.

Why that is an error and not a nitpick. The standard for these headers permits a made-up relation only as a full web address, never as a bare word. A bare unregistered word is not a loose choice; it is invalid, and a strict reader is entitled to discard the whole header. We had written the rule down correctly and then broken it twice, in the same file, in the same line.

Nothing is lost by removing them. robots.txt has always carried the sitemap, which is where every crawler actually looks, and a security policy is found at its standard path — that is what a standard path is for. The lesson is narrower than “check your work”: sitemap reads exactly like a standard word. Plausibility is not registration, and the registry is one fetch away.

The same inconsistency is in the specification we were auditing against, which is worth saying rather than quietly working around: its example for this file keys a link as sitemap and lists it as useful, while the same page says to use only registered names. Both cannot be true. We followed the registry.

What we measured, and why it sounded right. A heading painted with a clipped gradient takes its colour from a fill property that beats color outright. A forced palette throws the gradient away. The fill, read back from the browser, was still fully transparent. That is precisely the mechanism by which 90 headings across 45 pages printed blank on 12 August, a fault we found and fixed — so 229 elements on 130 pages, the ★ stuff wordmark among them, read as the same thing in a medium we had not thought to emulate.

It was wrong. A browser under a forced palette paints those glyphs in the reader's own colour anyway, whatever the fill says. The proof is two screenshots of the same masthead — one with the transparent fill still winning, one without — and they are identical. Paper genuinely honours the transparent fill; this mode does not. The two are indistinguishable if you ask the browser what the fill computes to, and obvious the moment you look at the pixels.

The cost, and the reason this is a correction and not a footnote. A rule was written for it and applied to all 198 pages, where it did exactly nothing. It has been removed. Reading a property and inferring what a reader sees, instead of looking at what a reader sees, is the error this site is organised against — we have a page about it, a gate that exists because of it, and five entries below this one from a single day last week cataloguing tools of ours that mistook a description for the thing described. This is a new instrument making that exact mistake in its own first draft, and finding it took a screenshot rather than a better argument.

What the checker does now. It counts gradient-clipped headings and offers no verdict on them, with the reasoning written where a future version would be tempted to add the test back. A real check would sample the rendered pixels inside each glyph; that is a bigger instrument than this, and until somebody builds it the gap is recorded rather than papered over. A tool reporting “I cannot check this” is reporting a gap, not clearing it.

What we said, in the entry below and in our working notes: that their generator has no --check, that the check-markdown.mjs its own header names was never built, and therefore “a page edited without re-running it ships a stale mirror and nothing notices.”

The first two clauses are true and the conclusion does not follow. The check was built somewhere else and under another name: check-metadata.mjs, one of that site's seven ship gates, regenerates every derived file into memory — each Markdown sibling, the two index files, the feed — and compares. Its own error message for the case reads “a stale .md is the quiet one.” It separately asserts that every page still advertises its sibling. We tested both rather than reading them: corrupt a sibling and it exits 1; delete a discovery link and it exits 1.

How we got it wrong. We read the comment at the top of their generator, which names a file that does not exist, went looking for that file, did not find it, and concluded the coverage was missing. We did not read their list of gates — which is in their working notes, seven lines long, with the answer in it. A comment naming a file that was never written is a claim, and we repeated it. Their coverage was never the problem; their comment was, and it has been fixed.

And it is the same fault we spent the day cataloguing, turned outward. Five times today a tool of ours mistook prose for the thing the prose describes — a broken tag quoted inside a script, a comment about a deleted attribute, a changelog entry quoting a template. This is the sixth, and the only one where the prose belonged to somebody else. Reading a description instead of measuring the thing is the error this site is organised against, and we made it about a neighbour.

Their arrangement is arguably better than ours. One gate asking is every derived file current? is a cleaner shape than a --check flag on each of three generators, which is what we have.

This changelog was excluded from the Markdown set, and the reason was this changelog. The tool decides whether a page is a client-rendered field guide by looking for a template string; an entry published this morning quotes that template while explaining a fix to it. So the page describing the mechanism was mistaken for the mechanism, and the largest and most useful document on the site silently had no Markdown sibling. The check we had just added caught the consequence within the minute: 56 of 57 advertise their sibling.

Five tools, one day, one fault. The markup checker flagged a broken tag written inside a JavaScript string. The policy generator counted a comment about a removed attribute as a live attribute, and would have kept a weak keyword in the security policy indefinitely. Two of our own verification commands reported problems that were sentences. And now this. A regex over a whole file cannot tell a thing from a sentence about the thing — and this site writes about its own markup more than most sites ever do, which is why it keeps stepping on the same rake. The rule is now written down where the next person will find it: ask the structure, never the raw text.

One smaller one, in the same hour. The sentence in llms.txt announcing how many pages have a sibling was written with the number typed into it. It was wrong within the hour, when the misclassified page above joined the set. That is the fault the longest section of our working notes is about, committed inside a sentence whose whole job is to be accurate for a machine. It is counted from the filesystem now.

After the last attribute was removed, the policy generator still reported one inline handler and still emitted 'unsafe-hashes' for it. The handler was a sentence. The note we had just written into search.html explaining why the onsubmit attribute had been removed contains the words of the attribute, and the tool was scanning the raw file rather than the tags — so a comment about deleting a handler was read as a handler, and would have kept the weak keyword in the policy indefinitely on the strength of a description.

This is the third time today the same mistake has appeared in a different tool, which is the part worth publishing rather than the instance. check-markup.mjs flagged a broken tag written inside a JavaScript string; the search index once indexed a page's own listing of other pages; and now this. The fix is the same each time: walk the tags, not the text. The generator now uses the same quote-aware walk the markup checker does, and skips scripts, styles and comments.

The first run reported Proportioned to the Groove as a failure: four spread ids, all undefined. Nothing was wrong with it. It is a scroll zine — every section laid out at once, no paging, so no pager buttons and no active spread, exactly as it was built. The page had not been modified at all; the test had simply clicked a button that does not exist and reported the absence as a break.

The suite now checks that a page has a pager before it tries to use one, and skips the two scroll zines by name in its output. A test that cannot distinguish a broken page from an inapplicable one files the wrong bug, and the wrong bug on a page that is working is worse than no test: it sends somebody to fix something that is correct.

What we said, in the entry below this one and in _headers. That a full policy was deliberately absent because every page carries an inline <style> and many an inline <script>, so it would need either nonces — which a static site with no build step cannot produce, since a nonce must be unique per response and appear in both header and HTML — or 'unsafe-inline', which is a policy in name only. We concluded that frame-ancestors was the only half worth shipping.

Both premises hold. The conclusion does not, because there is a third option, and the specification we were reading names it in the same sentence: hashes. This site has 49 distinct inline script bodies across 157 elements and six distinct inline event handlers across 233 attributes. Forty-nine and six are small numbers. Every one of them is now hashed into script-src.

What made it look impossible was counting elements instead of distinct bodies. 157 inline scripts sounds like a hash list nobody could maintain; 49 unique ones, most of them the same pager repeated across a hundred zines, is a 3.3 KB header. The measurement that mattered took one command and we had not run it.

What was wrong. Stars We Grew Up On described itself as “The culture end of Star Stuff, where ” — and stopped. Not a truncated sentence: the description reads “…where ‘star’ means icon rather than nucleosynthesis”, and the quotation marks around star were plain " characters inside an attribute already delimited by ". So the parser closed the attribute at the first inner quote, kept nine words, and treated the rest of the sentence as stray attributes on the same <meta> element.

It affected all three of that page's descriptions — description, og:description, twitter:description — so a search result, a link unfurl and a shared card all showed the same nine-word fragment. Two more pages had the same fault in one tag each: A Field Guide to Being a Shark lost everything after “a ‘living fossil’”, and Two Hundred Years after Alfie Kohn's “Better Get Used To It”. Live since 11 August in the first case, which is the day that page shipped.

Why nothing caught it. The JSON-LD on the same page carries the identical sentence correctly escaped, which is how we know the wording was never in doubt — only the attribute. And the source reads perfectly: the sentence is all there, in order, in the file. The defect exists only in the parsed document, which no gate here was looking at, and a meta description has no rendering to look wrong. Found by reading seventeen descriptions in a row while building the index below, which is the only reason it turned up at all.

The result: clean. 5,591 interactive targets across 198 pages. 3,005 are inline links inside running prose, which the criterion exempts outright; 1,379 are smaller than 24×24 px but sit far enough from their neighbours to satisfy the spacing exception; none fails both.

Getting there took three attempts and the first two were wrong in opposite directions, which is the part worth keeping. Testing only a link's immediate parent under-detected the prose exemption — a link wrapped in <em>, or two links joined by the word “and”, both read as isolated controls, and it reported 11 failures that were not failures. Testing whether the nearest block ancestor is mostly prose over-detected it: the front page's pills sit in a block full of prose, so a genuine standalone control was exempted and a real finding disappeared. The criterion is neither. A link is inline in prose when real text sits immediately beside it in its own inline flow, and only that test gives the same answer as reading the page.

A number that comes out clean on the first run has not been checked; it has been asserted. These three runs disagreed with each other by eleven findings in one direction and one in the other, and the site never changed.

What we said this morning, in the entry directly below this one: that moving 66 embeds to youtube-nocookie.com was “the cheapest privacy win on the list.”

What Google actually promises. Privacy Enhanced Mode “prevents the use of views of embedded YouTube content from influencing the viewer's browsing experience on YouTube”; ads on such a video “will likewise be non-personalized”; and the view “will not be used to personalize advertising shown to the viewer outside of your site or app.” Every one of those is about personalisation, not about collection. Google does not say the request is not made, that nothing is logged, or that no cookie is ever set. The domain name says nocookie; the documentation says non-personalised.

Why it matters more than the wording. An embedded player is a frame from somebody else's server, so 292 of them across 13 pages reach Google the moment those pages load, whether or not a reader presses play. We had read the domain name and not the small print — which is this site's own too-good-to-check reflex, aimed at a piece of infrastructure instead of a quotation. The privacy page now states the trade on its own face and marks it as unfinished rather than settled, and names the 13 pages so a reader who would rather not be seen by Google knows which they are.

What we said, this morning, on this page and in the ledger. That “282 of 299 <th> carry no scope” and “28 of 83 tables have no <caption>”.

The first was a broken count, and the bug is worth naming because it is one keystroke. The figure came from grep -o '<th' — which also matches <thead>. There are 81 <thead> elements here, and 218 + 81 = 299 exactly. The real figure was 201 of 218. The correct pattern is grep -oE '<th[ >]'. An audit that inflates its own findings is not being cautious; it is being wrong in the direction that feels safe.

The second was literally true and materially misleading. 28 tables did lack a <caption> — but 24 of them carry an aria-label on the <table>, which gives them an accessible name just as a caption does. Only four tables on the whole site had no name by any means. Presenting 28 as an accessibility failure counted a naming convention we had deliberately chosen as an absence.

Caught by us, while fixing it — the inventory taken before the first edit disagreed with the audit that ordered the edit. That is the good case, and it is the second time in one day that the error was in our own check rather than in the thing being checked. The lesson is already written down here in another form: a derivation that returns a number nobody reads back is not a derivation.

What we were asked to check. A Popsci video essay claims Texas horned lizards are recovering in “protected limestone corridors” and reducing fire ant numbers through “grazing pressure” that keeps harvester ant colonies compact and defensible. Ten papers were gathered in the hope of finding support for it.

None of them contains it. The chain needs three links: lizards eat harvester ants (documented, half a century over); moderate predation makes a harvester colony better at resisting fire ants (no published evidence found); that lowers fire ant numbers across a region (no published evidence found — Texas Parks and Wildlife’s volunteer data suggested a negative relationship in some years, and the same report says “pooled data do not reveal a clear relationship”).

The causal arrow is reversed, and that is the part with a cost. The literature runs one way: suppress fire ants and small vertebrates recover. Lee Ann Johnson Linam’s conclusion is an ask — “recovery efforts will have to address their management in order to be successful.” Reverse it and a request for work becomes a reassurance that no work is needed.

And the geography is exactly backwards. The claim puts the recovery in limestone. The eastern Edwards Plateau — the limestone country — is precisely where that report records volunteers “confirm some pessimism.” The encouraging reports came from the Post Oak Savannah, the coastal barrier islands and the Dallas–Fort Worth Metroplex, and are described as “small and isolated.” Mason Mountain, where releases actually happen, is roughly two-thirds granite.

The true half of the headline is the bad half. Fire ants genuinely are being eliminated across parts of Texas — by the tawny crazy ant, Nylanderia fulva. One invasive replacing another, which is worse news, not better. The invented half is the comforting one.

The 1992 paper, and the numbers we did not print. Middendorf & Sherbrooke’s canid-elicitation study in Copeia is paywalled and every route we tried refused us. A search-engine summary offered specific percentages for how often lizards squirt at a dog versus at a human imitating one. Those figures are deliberately absent from the spread, and the spread says why — a synthesised summary is not a source, least of all for a number. The qualitative finding is carried instead by Henry and colleagues, read in full: a defence “employed primarily against canids.”

The 1989 venom paper. The 0.119 and 162 µg/g values are quoted as they appear in the 2021 paper we did read. The 28 µg/g figure for the spiny lizard and the plasma-transfer detail come from secondary reporting, and are logged as an open item.

The video. YouTube returned the title and nothing else; a site republishing the script returned 403. So the zine grades the claim as it was described to us, and says so on the spread. The claim has meanwhile propagated into synthesised search answers phrased as though it were research, which is the part that should worry anybody.

A 2019 study of introduced populations in South Carolina reports Solenopsis at 1.20%, 8.53% and 1.50% of the diet by number across three sites. That could be written up as “actually, horned lizards do eat fire ants” — and it is not: the identification is to genus, and Solenopsis contains many native species. The spread says so rather than spending the number in either direction.

Two other things were left unasserted for the same reason. That fire ant baits kill native harvester ants is contested — pesticides are repeatedly implicated in the decline literature, and at least one field trial found broadcast baits did not reduce native ant populations. And no arithmetic was done on whether one lizard could dent a colony, because inventing that calculation would be the same error as the video, pointed the other way. Grading the chain is the honest move; a counter-calculation would not be.

A release figure was also corrected before shipping: a press summary gave “260 released, 205 from Fort Worth Zoo”, where Texas Parks & Wildlife Magazine gives 263 released in autumn 2021, 159 of them hatched at Fort Worth Zoo. A widely repeated “more than 3,000 juveniles since 2017” total was dropped rather than printed unverified.

What we got wrong. Spread ten read: “The species was listed as threatened in Texas on 18 July 1977 — within a few years of that schoolyard. Which means the thing those children did is now illegal.” It then quoted Henke & Fair, “Handling horned lizards is illegal even if your intentions are good.”

What is right. Those children did not handle it. They took turns looking into a tree hollow and left the animal where it was. The statute Henke & Fair are describing names three things — it is illegal to “pick up, touch, or possess” one — and looking is on none of that list. It was lawful then and it is lawful now. Ryan caught it, within the hour of shipping, from his own account of the afternoon.

Why the error mattered. The spread was building to a real point — that the law is a measure of how scarce the animal became, not a verdict on a child — and it reached for the sharpest available version and overshot into a claim the source does not support. The refusals two spreads later already said “looking is not extraction”, so the zine contradicted itself in its own closing argument and no gate could see it: this is a claim about a Texas statute, which is not a colour, a tag tree, a position, a class or a card order.

And the correct version is the better one. The law draws its line at the next step because that is where the harm was, and collection sits in the decline literature beside the roads and the poison. Those children were on the right side of that line — and looking leaves the animal where it is, which is precisely why a schoolyard memory can end up as evidence. Spread seven gained the sharper half: nobody touched it, so the animal was on the first rung of its own defence ladder. Henry and colleagues set the order out — cryptic coloration first, then “flatten itself against the substrate… run a short distance and then ‘freeze’”, with hissing, lunging, horn-jabbing and blood-squirting all further up. The stillness those children remember was not tameness. It was the defence, working.

The cover began as a top-down horned lizard, positioned so its fringe ran through the issue numeral and its label. Moving and scaling it cleared the numeral but left it reading as a spiky ball rather than an animal — the fault recorded here before as four of five whale attempts drawing a fish. It was replaced rather than iterated: an eye, and the drop leaving it made of ants, which is the argument and cannot be mistaken for a different animal. Clearances against the numeral block and the title’s ink were measured, not eyeballed.

And a print fault no gate on this site can see. The nine diagrams sit on a dark ground rectangle for screen. starstuff.css blanket-inverts SVG text and strokes for print but deliberately does not blanket-invert shape fills — so every one of those grounds would have printed as a slab of ink with #111111 labels on top of it. A page-level print rule now sets them to fill: none. The contrast gate reports 0 screen and 0 print failures either way, because its print pass composites against a bare sheet; confirming the fix needed a separate probe reading computed fills back under print emulation. Same lesson as the August stylesheet case: run the gate for regressions and something else for confirmation.

What we planned. The zine was proposed as The Yellow Was Already There — the standard account, in every popular explanation of autumn colour: the green goes and reveals the yellow that was hiding underneath.

What the one source that actually separated the yellow pigments found. Keskitalo and colleagues, following a single aspen through an autumn: “Lutein and beta-carotene were degraded in parallel with chlorophyll, whereas neoxanthin and the xanthophyll cycle pigments were retained longer.” The two main yellow carotenoids went down alongside the green, not after it. And the paper that states the standard version hedges it — Mattila et al. write, in their introduction rather than their results, that autumn colours are “partly caused by exposure of carotenoids due to faster degradation of chlorophyll.” The popular retelling drops the partly.

What the zine does instead. The uncontested half stays: the pigment really is in the summer leaf, and the spread that says so is the claim the piece is most confident of. The contested half gets its own spread, in the child’s own layer, ending “nobody has finished this one either.” A read-aloud could have printed the tidy sentence and nobody would have complained, which is exactly why it did not: for kids is the shelf where a verify-everything standard quietly lapses, and this was the sentence on the page most likely to have lapsed.

Adding this piece, the Young Readers line about how many pieces hand the child something to do was bumped by arithmetic — from “six of the eight” to “six of the nine”, on the assumption that the new piece added none. Deriving the set from the read-aloud layers of all nine pieces gives seven, not six. The page now says seven, and the derivation agrees with the page’s own separately-worded sentence naming exactly the pieces that have nothing. An inherited count incremented is not a count derived — and no gate on this site can read a number written out in a sentence, which is why this file keeps saying so.

Two more figures on that page had gone stale in the same way and were re-derived: the list of what the pieces measure, which still stopped at “spadefoot toads, salamanders, seeds or bees” four pieces after it stopped being true, and the note on which pieces have no demonstration.

Third cover in a row where the motif had to be measured rather than placed. Here the binding constraint was horizontal — this motif is a wide shape and the numeral block’s ink starts at 864 px — so it had to be scaled down to 0.70 rather than merely moved. check-overlap.mjs passes it and is right to: art over text is the half of the No. 48 fault that gate was explicitly built not to cover, so the clearances are measured against each element’s ink box every time.

The print probe was clean on the first run for once, because this zine’s figures were written with the lesson from No. 100 and No. 101 already applied: every pigment bar is labelled in text as well as coloured, and the page carries its own print rules pushing the shapes to two distinguishable inks, so the figures still read on a black-and-white sheet where the green/gold distinction cannot survive.

How many times ear asymmetry evolved independently among owls is given as at least five by Norberg’s own 1977 abstract, as at least four by a 2018 paper summarising that very same paper, and as up to seven by a 2019 one. The count depends on which owl family tree you use and on how different two asymmetries must be to count as separate inventions. The read-aloud uses the primary’s own figure and the grown-up box states the disagreement, because picking the tidiest number would have been the easiest possible place for this collection’s standard to lapse. The independence is not what is in dispute — the counting is, and the reason nobody doubts the independence is the fact the child gets: the anatomy is different every time, soft tissue over a symmetrical skull in one owl and skull bone in another.

An open question from the proposal, now closed. Which of the barn owl’s ears sits higher was reported inconsistently across the sources we had at proposal stage. Two independent primaries settle it and agree: Knudsen and Konishi on Tyto alba, and Krings, Rosskamp and Wagner on Tyto furcata pratincola — “the left ear opening in the skin is located higher than its counterpart on the right.” The zine’s figure is labelled from the owl’s point of view and says so on its face, because an owl facing you has its left on your right, and an unlabelled drawing tells half its readers the mirror image.

The cover art collided with all three cover elements at once — the issue number by 5 px, the masthead line by 10 px, and the title by 57 px vertically. The cause is worth keeping: this title wraps to two lines where No. 100’s took one, so its ink box is 112 px tall instead of 62 and starts 42 px higher up the cover. A position inherited from a sibling is a position measured against a different title. Rescaled to 0.85 and repositioned by measuring the painted group against each element’s ink box. check-overlap.mjs passes all of it and is right to — that gate compares text to text, and art over text is the half of the No. 48 fault it was explicitly built not to cover.

And the print probe earned its keep a second time. Two arrowheads were drawn as filled triangles in a light rgba(), which is matched by neither the shared print sheet’s list of light fill literals nor this page’s own print rule — so they would have printed near-white and vanished beside a bracket printing at #444444. Converted to strokes, which the shared sheet handles for free. Re-probed clean: every shape #333333, all seventeen labels #111111.

What we planned. Cut a paper circle the width of the seed head, drop both, and watch the solid one fall faster. It is the obvious version, it sounds right, and it is wrong.

What the paper says. Their own arithmetic runs against the intuition: for a solid disc to supply the same drag as the pappus at the same falling speed, its diameter “is given by D = 8.6 mm, which is 38% smaller than” the real 13.8 mm plume. Per unit of width, a solid disc is the better brake. A same-size paper circle has more drag — and, being some twenty times heavier, lands first anyway. A child following that instruction would have got the opposite of what the page promised, which is the exact failure the collection names on its own face: an imperative a child follows and gets nothing from teaches them that the book was wrong.

What the demonstration is now. The observable, correct difference is steadiness, not speed. A paper circle flutters, tips and slides all the way down; a dandelion seed comes straight down upright. That is the finding: the paper measured the critical Reynolds number at which the ring breaks up into vortex shedding at 149 for an impervious disc against 429 for real seeds, with a porous disc built to the dandelion’s own porosity giving 457 — and “all of the dandelion samples tested flew at a Re below Re_c.” Spread ten states the correction to the reader rather than quietly benefiting from it, and the figure was redrawn to show orientation, which is what a child actually watches. No gate could have caught any of this. A demonstration that quietly proves something other than what it claims is worse than no demonstration, and the only thing that catches it is reading the primary source.

The zine’s second figure is a hundred dots — eight filled for the hair, ninety-two hollow for the air. The shared print stylesheet turns every stroked SVG shape #444444 on paper and inverts a fixed list of light fill literals; this page’s one-off accent is not on that list and could not be. So on paper the ninety-two hollow dots would have come out dark and the eight filled ones pale: the emphasis inverted, and the picture saying the opposite of its own caption. check-contrast.mjs cannot see it — it measures text against its background, and a shape fill carries no text. Found with a separate print-emulation probe reading computed fill back off the elements, and fixed with a small page-level print rule whose selectors name circle, ellipse and rect explicitly so they cannot reach a label. Re-probed after: shapes #333333, every label #111111.

The cover motif’s descending flow arrows crossed the last word of .cover-title — art over text, measured at 1280 px as arrows at x 742–806 against a title ink box running to x 843. check-overlap.mjs passes it and is right to: that gate compares text to text, and text over artwork is the half of the No. 48 fault it was explicitly built not to cover. Caught by looking, measured with a Range rather than by eye, and the motif moved up 56 units to leave 14 px of clearance.

The paper is Nature 562(7727), 414–418 (2018) — not 559, which appears in secondary accounts and which we would have carried had we not checked it against Crossref and Europe PMC before building on it. Publisher access is blocked, so the text was read from the peer-reviewed accepted author manuscript in the University of Strathclyde’s repository; the zine says so in its colophon, and says that it is the accepted version rather than the version of record.

While this was being built, Ryan supplied the source we had flagged as untraced for a different, rejected proposal about dog body-size variation. Reading it settled the matter in an unexpected direction. Shearin and Ostrander (PLOS Biology 8(3), e1000310) is an Essay rather than a research report; its headline claim is hedged in its own text — “arguably the most morphologically variable land mammal” — and it does not contain the hundred-fold Chihuahua-to-mastiff range that circulates widely and gets attributed to the literature. What it actually states, in a figure caption, is that the difference between a Cane Corso and a Yorkshire terrier is over 30-fold. Nothing in No. 99 rests on any of it, and the discrepancy is recorded on the zine’s own colophon. A source you are handed can be a lead that dead-ends, and saying so is the finding.

Read-aloud is exactly the shelf where a verify-everything standard quietly lapses, so the two figures most likely to be quoted back are hedged where the adult will see them. The four-to-seven-sniffs band is real, but the paper also reports sniff frequency scaling with body mass — so “the big dog and the little dog did it at the same speed” is true at the resolution a five-year-old is being offered, and the grown-up box says exactly that. The 12–13 per cent split is one modelled case of a “relaxed, undilated nostril”, and the authors hedge it upward themselves. A whole spread is given to a distinction between two verbs: that each nostril samples a separate patch of air is a measurement; that the dog uses this to work out direction is the authors’ proposal, written “may be used”. The child is told the difference.

Ryan handed over six items on dogs. The chain’s ending came from none of them: Genomic history of early dogs in Europe (Nature 651, 25 March 2026) turned up while we were chasing a radiocarbon date for the Goyet skull. It reads the oldest dog genome yet recovered — 14,200 years old, from Kesslerloch in Switzerland — and rules out one long-standing hypothesis before declining to settle the rest. Without it the chain would have stopped at 2020 and been six years stale on its central question. A chain is only as current as its last search, and the source drop is a starting point rather than a boundary.

Darwin is quoted at the 1868 first edition, not the one everybody transcribes. The public-domain text of The Variation of Animals and Plants under Domestication that circulates online is the 1875 second edition, and the zine dates its quotation 1868 — so the sentence was checked against the first edition as well. It appears in both. A sentence quoted at the wrong date is a sentence nobody checked. And the retelling of Darwin is wrong in a way worth printing: the secondary accounts have him wondering whether dogs came from one species or a wolf-jackal cross. He considered it and then concluded the opposite of a wondering — “the balance of evidence is strongly in favour of the multiple origin of our dogs.” Separately: the 1997 paper that founded modern dog genetics is universally reported as dating dogs to 135,000 years ago. Its own abstract says “more than 100,000 years before the present.” The zine prints the abstract’s figure and names the other as the number that entered circulation.

How We Got Here said that No. 52 “carries the most contested joints of any chain here.” It did, with four. No. 98 has five, so the sentence was corrected in the same pass rather than left to rot, and the collection tally was re-derived from the pages rather than incremented — fifteen chains and 135–33–14 becomes sixteen and 141–38–16. A new piece can falsify a sentence on a page it never touches, and nothing in the eight gates can read a count written out in prose.

Nagasawa et al.’s oxytocin gaze loop (Science 348, 2015) is the obvious thing to build a zine like this on, and we did not. Published commentaries flag confounds between its dog and wolf conditions and note the effect was not uniform across the dogs. Contested is not refuted — and it is also not a floor. Albuquerque et al. (Biology Letters 12, 2016) went the same way for a different reason: we could find it neither replicated nor overturned, which is an absence of evidence about the evidence. Both are named in the colophon rather than quietly omitted, because a reader is entitled to know which famous thing is missing and why.

The outline carried a sweep about what dogs have done for us across millennia. The survey behind it — Fallahi, Masoudi, Maghsoudi and Vaez Torshizi, Veterinary Medicine and Science 10, e70012 — is titled “A 40,000 years history” while its own text gives the range as “between 10,000 and 40,000 years ago”, a factor of four, cites secondary sources for it, and carries the typo “∼40,0000–10,000 BP.” The confidently-assigned archaeological floor is far more recent: about 15,000 years in Europe and over 12,500 in Far East Asia, read at Frantz et al. (Science 352, 2016). The sweep is gone rather than restated at a number we could not stand behind, and the deletion is printed in the zine’s own colophon.

The zine closes on Carol Millman — an Autistic professional dog trainer — who answered “is ABA dog training for children?” by pointing out that the comparison is unfair to the dog trainers: meet the animal’s fundamental needs before any conditioning, relieve the fear rather than train away its display, and know that a dog taught not to growl is a time bomb dog, because you removed the warning and not the fear. We found her piece through a saved highlight file in our own library, which credited it to “neuroclastic.com” rather than to her. Both quotations were verified at the article instead, where her name and credentials are. An aggregator can erase an author as easily as it can alter a quotation — and this one is in a store we built ourselves.

The paper we cite for the match is good also reports where it fails: “most reflectance spectra of cuttlefish had a peak around 800 nm in the infrared range, whereas natural substrates tested did not have this spectral characteristic.” In infrared the animal does not match at all. That is a whole spread rather than a footnote, because printing only the flattering half of a paper is the tidier version of it. And the grown-up box says exactly what we did not establish: vertebrate visual pigments are not known to reach 800 nm, but we found no study asking whether any cuttlefish predator can detect it, so “it does not matter” is an inference from an absent sense rather than a tested result. Two other caveats sit on the page: the 492 nm pigment measurement traces to a single study from 1958 that nobody has redone with modern methods, and “cephalopods are colour-blind” is false as stated — the firefly squid carries three photopigments.

Too Good to Check is the page that lists what we got wrong, and for two days it did not list the compression-versus-shear error — the one we swept across seven pages. It is now the eighteenth row: “Press a bone and it makes a voltage — negative where compressed, positive where stretched. Ours, on seven pages, for seven weeks.” It is filed under Mode 03, the story told backwards, which is exactly its shape: the facts survived the retelling and the mechanism did not. It also carries the detail that makes it worth printing — a 2021 review in Bioelectricity states the mechanism correctly on one page and repeats the compression gloss four pages earlier, which is how a wrong sentence stays alive in a literature that also holds the right one.

The changelog entry for the Tsing matsutake case ends: “That page’s worked-correction count goes from seventeen to eighteen.” It did not. That case was written into the page’s prose, as a worked example under Mode 03 — the cleanest instance we have of a hedge dying in quotation with the author at no fault — and not into the receipts table. So the table stayed at seventeen rows, its caption kept saying Seventeen receipts, and the four other places that cite “seventeen worked corrections” were right all along. The entry asserted the consequence of an edit rather than the edit, which is the same species of error as everything else on that page. Today’s addition is what actually takes it to eighteen, and the caption and all six downstream references now say so.

Owed, Not Earned is the sheet we print for a door, and section vi read: “if you have flesh and blood, we accept you. If you are made of stars — and you are — you are already one of us.” That is membership derived from what you are made of, laid out as a formal step. The colophon then named the step as the piece’s own contribution: “that belonging is constitutive because we are made of the same star stuff — is ours.” The sheet never needed it. Its actual warrant is Timothy R. Clark’s — “not earned but owed … a nonnegotiable right” — plus Douglass, and both are moral arguments that stand without any physics at all. Section vi now reads “You are made of stars, and the welcome goes all the way down to them,” and the colophon says the belonging rests on Clark and Douglass, “not on the physics: it would be owed in a universe made of one element.” Section v was deliberately left as it stands — “Kinship is not a courtesy we extend to you. It is the material you are made of” is a true claim about what kinship is, since kinship means shared origin and we do share it. The fallacy was only ever in the step from kinship to membership.

This week’s shear correction belongs in Six Ways’ list of the ones we got wrong — it is precisely that sheet’s Mode 03, the paper’s own hedge is missing. It does not fit. One more entry costs about 48px and Side B has less slack than that; a full entry overflowed by 64px and a two-line one by 48. We did not cut an existing correction to make room — trading one published error for another is not a fix. Its real home is Too Good to Check, which the sheet already points at for the full list and which is also missing it. That page holds seventeen worked corrections and states no count of its own; “seventeen” is asserted in four other places instead. So the eighteenth correction is owed, and so are eight count updates across four files. Logged here first, because the gap is the finding.

What We Refuse ranks our refusals by how often they recur, and the one at the top — on nine pieces — is a rhyme, not a proof: “That a pattern recurs in nature is not why difference deserves welcome. It would deserve welcome in a universe that made exactly one shape.” Side B then read: “Difference is variation, not deficit. The universe doesn’t pathologize its own variation. Neither do we.” That is the derivation, printed on the reverse of the sheet that counts it nine times — and Side B’s own commentary says why it matters: take that refusal out of the natural-history pieces “and what’s left is an argument that biology licenses a politics.” The panel now reads “True even in a universe that made exactly one shape. We are the ones saying it, and we are saying it all the way down.” The sheet also now says so on its own face, in the section it already keeps for this — And: we notice we are doing it too, whose whole point is “not a claim to have escaped the trap, but a note, in the piece itself, saying where we are still standing in it.”

The Manifesto said bone generates charge “under mechanical stress — compression, tension, bending” in one paragraph, and correctly described collagen fibres slipping past one another under shear two paragraphs later. Both were already on the page, and nobody had noticed they disagree. It also stated the contested part as settled — “It is how bone knows to grow, where to strengthen, how to heal. The skeleton is a living signal processor” — and carried “Every step you take, every breath, every impact — your bones are generating voltage,” which is verbatim the sentence Bone Song’s refusals name. All three fixed. The same framing was also in the read-aloud’s grown-up box, its colophon, its structured metadata, its card on two pages, and the broadside’s card on the Print collection. Seven pages, one grep.

Rewriting Bone Song’s poem removed the universe doesn’t pathologize / storms are not disorders stanzas. Bone Song’s own refusals then still said “the poem says the universe does not pathologise its own variation” — pointing at a line the poem no longer had. And No. 95 described that removed argument as No. 1’s in four separate places, including its refusals and the note explaining how read-aloud translations are numbered. Editing one piece can falsify another piece’s account of it, and every gate passes cleanly: the sentence stays grammatical, the link still resolves, the markup is fine — the claim is just no longer true. All five now describe what the zine actually argues: a nothing found is a fact about the instrument at least as much as about the thing it was pointed at, and the belonging is a promise being made rather than a finding being reported.

Carter and colleagues (Bioelectricity, 2021) write that a charge appears “when a shearing force is applied to collagen fibers and they slip past each other” — the 1957 finding, stated correctly. Four pages earlier the same review says “compression on bone is shown to produce a negative electric charge … whereas traction produces a positive charge” — which is, almost word for word, the sentence this zine carried for seven weeks and corrected yesterday. A peer-reviewed review holding both at once is the measure of how completely the compression gloss has settled into the secondary literature, and of why nothing short of the primary could have told us which one was Fukada and Yasuda’s. It is now said plainly in the colophon rather than left as our own bad luck.

The same review reports that the piezoelectric charge opens voltage-gated calcium channels in bone-repair models. That would tie spread two’s electricity directly to spread four’s star-forged calcium — a genuinely satisfying join, and cited. We have left it out of the argument. It sits inside the very mechanism spread three spends a page saying is unsettled, so spending it here would spend the hedge for a rhyme. It is named in the colophon, so the choice is visible rather than a gap. And both papers carry the same limit, flagged twice: their subject is overwhelmingly engineered piezoelectric materials — barium titanate, PLLA, ceramics, scaffolds mending defects in rats. That applied stimulation helps a bone heal is well supported and is the lineage Bassett’s clinical work belongs to. It is not the claim that native bone uses its own voltage to decide where to grow, and this zine keeps the two apart.

Spread two said the voltage goes “negative where the tissue is compressed, positive where it is stretched.” That is the later stress-generated-potential literature, not Fukada and Yasuda. Their paper says the effect “appears only when the shearing force is applied to the collagen fibres to make them slip past each other,” that its magnitude depends on the angle between the pressure and the bone’s axis, and that the sign and amount of polarisation “vary considerably with the direction of pressure.” Push straight down a bone’s axis and there is nothing to read. It answers to being twisted, not merely pressed. The diagram caption was wrong for the same reason. Not one of the abstract-level sources or review summaries we had used said any of this — which is the whole argument against treating “confirmed at the abstract” as a small discount on reading the thing.

Friday’s pass added “Not: physics as a warrant for a moral claim” while the paradigm spread was still making exactly that argument — “a star doesn’t diagnose its own magnetosphere … but the bone still made its electricity.” That closing clause is the fact vindicating the person, which is the move A Promise, Not a Finding exists to refuse. The spread also restated the poem in prose, and asked “what is the universe trying to do through this particular configuration of matter?” — the universe is not trying to do anything.

The entry said it was — in its summary line, in a tag, and in a whole fact row reading “Piezoelectric · pressure becomes voltage · bone sings under load.” The effect belongs to bone as a composite, and is attributed to the collagen rather than the mineral, which we had already established elsewhere on this site in August. An element guide is exactly the wrong place to hang a composite-material property on a single element. All of it is gone.

“Every step you take, every load your skeleton bears, gets quietly translated into electricity that tells the bone where to grow.” That asserts the remodelling mechanism outright — the precise part the literature does not settle, and a step further than the sentence removed from No. 1 the same day.

Bone piezoelectricity is Fukada and Yasuda’s. This page named Yasuda alone in four places — an eyebrow, the body prose, the colophon, and the poem. We had already fixed this on the Manifesto and the L★S Broadside in August and never swept it back to the founding zine. The poem line is now “Fukada and Yasuda knew this / They pressed on the crystal and it answered.” Editing verse is a bigger act than editing a citation, and it is recorded as one.

That sentence was a step past the evidence, and nothing on this site had ever mentioned why. New spread three: Fukada and Yasuda measured dried bone. Piezoelectric response falls as tissue is hydrated — Fukada came back to measure exactly that in 1982. In a living body there is a second mechanism, streaming potentials from fluid squeezed through the tiny channels in bone, and a voltage measured in wet tissue is not by itself evidence of piezoelectric origin. It is not refuted either — recent work finds collagen’s response measurable at the humidity a real body runs at. Unsettled in both directions, for fifty years. The claim that holds is narrower: squeeze a bone and it makes a little electricity, and this was measured. The effect is also now attributed to the collagen rather than to collagen and mineral jointly.

The first draft of this revision said the page had carried the fault “for three years” — in four places. Ryan asked: three years or three months? Seven weeks. 17 July to 6 September 2026. The figure had been invented to make the correction sound weightier, which is precisely the failure this site exists to catch, committed inside a correction to our founding fact. All four fixed; a grep confirmed it had not spread to any other page.

Bone piezoelectricity is Fukada and Yasuda’s — Eiichi Fukada is first author on the 1957 paper. This site credited Yasuda alone for a long time. We fixed that on the Manifesto and the L★S Broadside in August, logged it, and never swept it back to No. 1 itself, where it still stands in four places. Fixing the instances you noticed is not the same as fixing the fault, and this is that lesson landing on the oldest page here. No. 1 has not been edited yet: one of the four is inside its prose poem, and rewriting the founding artifact is a decision, not a correction to make in passing.

This is the bigger one. Fukada and Yasuda measured dried bone. Piezoelectric response falls as tissue is hydrated — Fukada himself returned to that in 1982 — and in living wet bone there is a second mechanism, streaming potentials from fluid squeezed through tiny channels, which some researchers think matters more for cell signalling. Voltages measured in hydrated tissue are not by themselves enough to show a piezoelectric origin. It is not refuted either: recent work finds collagen’s piezoresponse still measurable at physiological humidity, and a 2023 study reports both mechanisms coexisting and possibly coupled. The question is live in both directions. Until now, nothing on this site mentioned any of it, while No. 1 and the masthead broadside both say “every step you take is a small current” without a hedge. What No. 1 says is not false — the effect is real and was measured. It is stated more strongly than the evidence about living bone supports, which is a different fault and still worth fixing.

Found in passing. check-markup.mjs verifies that a piece’s collection badge is labelled with the name the collection page gives itself, and it takes that name from the page’s <title>. If the title does not match the canonical form, the tool deliberately stands down rather than failing every member — an unverifiable claim is not a failing one. Triggers was the one page of sixteen whose title did not match, so that check has been inactive on it since it shipped. Title canonicalised; the check is armed again.

That is the sentence No. 94 arrived carrying, and it is in nearly every popular account. The review it is built on reverses it: ultraviolet is “merely one component of colour vision” for a bee, not a separate channel bolted onto a human-like system — and humans are more sensitive to red than bees are. A bee cannot see red at all. The window is shifted, not widened. Two comparably sized pieces of one continuous spectrum, sitting in different places, each blind to what the other has. The true version is better and harder to sell, and writing it the popular way turns a difference into a ranking, which is the move this whole site exists to refuse.

The first draft held that frogs sit on lily pads was mostly artistic convention, on the strength of finding no primary source for it. What we had actually found was that content farms are not sources — and we reported the shortcoming of our own search as a fact about frogs. A null result from a bad corpus is not a finding, especially for something so ordinary nobody publishes a paper about it. Ryan Boren corrected it from his own pond, and then from a photograph: a southern leopard frog fully up on a mature lily pad, taking its whole weight. The picture was never the thing that was wrong.

Draft two had settled into the pad is a lung the frog sits on — which sounds like a finding and quietly asserts a relationship. The second correction killed it: the frogs perch on everything in the tank. Pennywort, elephant ear, rocks, the steel rim. What an adult frog wants is geometric rather than botanical — a flat surface, at a height, in the sun, above the strike zone, one jump from the water. Take the lilies out and they would sit on whatever was left. The lily has no claim on the frog, and that leaves the actual finding standing alone.

Couch’s spadefoot is everywhere reported as metamorphosing in eight days, the shortest of any North American amphibian. Newman’s 1989 paper gives eight to sixteen with sufficient food, and the range — one species, one desert, twice the time — is what the spread is actually built on. Separately, secondary accounts of the axolotl’s decline run a density series ending “35 per square kilometre in 2017.” The Red List assessment contains no such number. What it says about 2017 is starker and differently shaped: monthly sampling at Xochimilco did not record the species at all. We cut the tidier series and printed the assessment’s own sentence. And an American bullfrog figure would not resolve to a primary, so it appears with no number at all and the grown-up box says why.

The Chlamydomonas card was drafted as look, an organism with no sexes at all. That is not what isogamy is. The gametes are the same size, so sperm and egg genuinely have nothing to attach to — and there are still two mating types, distinguishable by where the mating structure sits relative to the flagellar beat plane and the eyespot, under the control of MID. Equal size is not sameness. Printing the easy version would have been this guide flattening a real distinction because it did not match the expected shape, which is the exact thing it accuses the two-sex model of doing.

Spread ten prints a correction to the author’s own long-held reading rather than dropping it quietly. Bonobos as a liberatory model — sex in place of conflict, food shared, females allied — is built from attested behaviour and is not a society anybody has observed. Mouginot and colleagues followed twelve male bonobos and fourteen male chimpanzees for two years and counted: bonobo males were involved in roughly 2.8 times as many aggressive encounters, and the most aggressive sired the most offspring. The real difference survives, and it is not the one in the story: lethal aggression is common in chimpanzees and absent in bonobos. The error was not believing a false fact. It was believing a selected one, and stopping.

Chatterjee and Ghosh’s abstract says nine mixture components for both species. Their discussion says nine for blue and twenty-nine for humpback, in two separate passages. We could not reconcile that from the published text, so spread nine states both and says the discrepancy is the paper’s. A second caution is ours rather than theirs and sits in the colophon: the models assume independent directional changes across environmental variables — the authors’ own third limitation — while the headline result is a correlation between two of them. We report the finding and the caveat and draw nothing from the tension. Two conversions on spread three (1° over 2,230 km ≈ 39 km; 10° ≈ 393 km) are our arithmetic, not the paper’s, and are labelled as ours on the spread.

All ten claims trace to a primary or to a named peer-reviewed review. Three did not and were cut, and the page names all three: that airglow is the brightest natural component of a moonless sky; that Joshua Childrey first described the zodiacal light in 1661; and the light-pollution atlas’s regional splits (60 per cent of Europeans, ~80 per cent of North Americans), which are the paper’s own but reached us only through news coverage. One date is left open on the page’s face: Cassini’s zodiacal work is popularly dated 1683, and the review we actually read lists it as 1693 — the plate prints what the source prints and says why. And the single-photon result carries P = 0.0545, on the wrong side of the line those papers conventionally draw. It is on the plate in that form rather than swapped for the stronger confidence-restricted figure. A sheet about the edge of detectability has no business being confident about where the edge is.

Four IPCC Summaries for Policymakers were read directly, six papers read at source, and Pauly’s 1995 page read in full. Three sources were out of reach and the page said so on its face: the Armstrong McKay 2022 Science paper refused the request, the Global Tipping Points Report 2025 had been used only through institutional releases, and Marshall et al.’s Reef Grief sat behind a publisher wall with its figures deliberately left off the page. All three arrived mid-build, and the tipping-point section was rewritten from the papers themselves. Stating a gap is not a consolation for having one; here it is the thing that closed it.

Two of the three gaps were not merely sourcing gaps. The near-term North Atlantic tipping element is not the AMOC. Armstrong McKay et al. put Labrador–Irminger Sea convection at a best estimate of ~1.8 °C (1.1–3.8), high confidence, over about ten years — against the AMOC’s ~4 °C (1.4–8), low confidence. A lower threshold, a faster transition and better confidence, on the system nobody has heard of. The full report added a marine tipping system the draft had missed entirely (kelp forests, local-scale, high confidence, urchin barrens as the mechanism) and revealed that the headline coral threshold moved from ~1.5 °C to 1.2 °C between 2022 and 2025, which the page now prints as a movement rather than as a number. “Confirmed at the abstract” turns out not to be a small discount on “read in full” — none of that was in an abstract.

A summary indexed as “new research on the mental health consequences of working in marine conservation” was about to be cited for a striking figure. Read properly, it describes a review with no original survey, a study of residents, tourists and tourism operators rather than professionals, and a study of atmospheric and environmental scientists. There is no survey of marine conservation professionals behind that headline. Reading the paywalled paper closed the sourcing question and confirmed the absence rather than filling it — so the Reef Grief figures are now printed with their real populations named (53.5 per cent of residents, 22.9 per cent of fishers, and the authors’ own caveat that it rests on a single survey item), and the missing study stays in the held section. The title of an aggregating page is a claim about its contents, and has to be checked like any other claim.

Glimmer Wire, edition one was published on 2 September and its file was created on the 4th, when the single page was split into one page per edition. Git can only see the file. The tool carries exactly one date override, with the reason on the line above it, and prints the count of overrides every run — the same shape as the watermark exemption in the contrast checker and the hook list in the dead-class checker. An override should be a decision somebody wrote down, not a mechanism to fall into.

Every tagline on What’s New is already indexed on the card it was lifted from, so the listing is a duplicate presentation and belongs in CHROME_SEL alongside the playlist row list and the front page’s jump strip. Stripping it exposed a gap that had been there all along: textOf() removes chrome from its own clone, which handles chrome inside a chunk — but the chunker selects from the live document, so a chunk inside a chrome container was still indexed. The listing’s 33 <section> elements were excluded from the coverage denominator and indexed anyway: 34 records at 3000% coverage, a figure only visible because the run prints coverage per page. Step 3 now skips a chunk whose ancestor is chrome, which segmentByHeading() had always done. Verified as a no-op elsewhere by diffing every record against the committed index: 174 pages byte-identical, the only change the front page’s new masthead link.

Al Jazeera reports the resolution as saying it “in no way challenges the use of the Mercator projection for maritime and aerial navigation, for which it remains fully suited.” That sentence does not appear in A/80/L.104. The document says without prejudice to the established practice with regard to the use of cartographic material for navigational purposes. The gloss is close in spirit and is not the text, so it is not quoted on our page as resolution wording.

Ukraine’s stated reason for abstaining is characterised from summaries of the meeting record rather than quoted, because we have not read the verbatim record. Said on the page, not only here.

The page first reported only that the US representative “called the resolution ‘superfluous’”, following the Guardian. Ryan flagged the fuller remarks mid-build, which led to the primary: the United States Mission publishes its explanations of vote, so Yaryna Ferencevych’s statement can be read in full. The word “superfluous” is not in it, and it made a substantive objection sound procedural — the statement calls the resolution part of “a much larger and more radical ideological project” and such resolutions “barnacles on our work here”. Two paragraphs are now quoted verbatim instead.

Which turned up something checkable and worth having on the page. “Cognitive justice” is in scare quotes in the US statement, and it is the resolution’s own phrase — its preamble says correcting these distortions “constitutes an act of cognitive justice and memorial reparation”. The objection lands on language the document really uses. What it gets wrong is the premise that this was “presented as an anodyne effort” hiding something larger: the resolution states its reasoning on its own face, in its preamble. Both halves are on the page, because a piece about summaries losing what they were not built to keep should not summarise the one statement against.

A draft paragraph had Ryan clipping the “more accurate” framing on Friday night as his own belief, and then correcting himself. He had quoted it from another source as an example of wrongness. He caught it mid-build; the paragraph now says so. Recorded because the ledger is for our own errors first.

The New York Times obituary and Rolling Stone both have Wray poking a pencil through the speaker cone, and that is the sentence that has travelled — into Wikipedia, into a hundred articles, and into our own first draft of the page.

Wray, quoted in Guitar magazine via the Registry essay: he used a pen, on the tweeters, and “I didn’t mess with the big speaker.” A more precise act than the version everyone prints — it is the high end he was after and he deliberately left the woofer alone. The smoothed version is easier to say and slightly wrong, which is this collection’s subject arriving inside its own sources.

Rule three of the Sound collection says every embed is resolved through YouTube’s oEmbed endpoint before it ships, which returns the real title and channel for a video ID. That catches the fault it was built for: a card quietly pointing at a different recording. It is structurally blind to this one.

Both failing IDs return HTTP 200, with the correct title and the correct channel, today, while the player refuses to play them. The check was verifying that a video ID is what we said it is and never that it still works. All eight gates were equally blind, and correctly so — the markup is valid, the contrast is fine, the words are all indexed, the card is in the right place in the list. The rule has been corrected on the collection page rather than quietly patched.

The rack is built from Al Carroll’s Indigenous Punk (Microcosm, 2026), read in full. It states that Wray entered the Rock and Roll Hall of Fame in 2007. He entered in 2023, in the Musical Influence category, per the Hall’s own inductee record — eighteen years after his death, not two.

The same book names Dead Pioneers’ guitarist “Gregg Rivera” where he is Josh Rivera, and spells the singer “Greg Deal” at first mention where he is Gregg Deal. Neither error changes a claim the page makes. Both are on the page rather than quietly applied, because the sixteen-year gap is the argument for checking and not a footnote to it.

The distinct-title figure could not be reproduced by any method we tried. The page published 182; our own fact-check ledger records 181 for the same set; case-insensitive matching gives 193. The 2026-08-27 ledger row names the cause outright — the number had been carried forward by delta from a base our normaliser already disagreed with, twice. It has now been replaced rather than incremented, and the page states its method: 231 distinct titles counted case-insensitively, 264 distinct song-and-artist pairs. The pairs method was validated by reproducing the existing figure exactly before it was trusted.

“The shelf is 146 embedded videos” was never an embed count. 146 was the card total on 26 August, and it was two racks stale besides. Derived: 269 embeds across 268 cards — the odd one out is a video inside the AURORA essay rather than on a card. And rule two’s “a hundred and twenty-one reasons” was the four-rack collection total from 12 August, used in a sentence about how long a single rack may get; it is now the longest rack, forty-four.

Every figure on the page was re-derived from the card markup rather than incremented: eight racks, 268 cards, replacing seven racks, 230 cards in six places including the og: and twitter: descriptions that a link unfurl quotes. No gate can read a number written out in a sentence, which is why this keeps happening and why the method is now printed next to the number.

The brief asked how do we re-home climate refugees. The report answers by rejecting the framing: “Pacific Islanders have explicitly rejected the framing of ‘climate refugees’”, and both Pacific and Caribbean island communities have consistently expressed a preference to remain in place, with relocation a last resort.

Where movement does happen, the IPCC finds with high agreement that the degree of agency people retain over if, where, how and when they move is what determines whether mobility is adaptation or an additional harm — while consultation “remains very limited in practice, including with populations whose movement is being planned.” The report turns that into its own question, and the entry adopts it: “whose relocation is it?”

The Science macrophage entry said the coverage’s “reverses ageing” overreached, because “the paper’s own title claims the impairment limits organ ageing.” The title is “Restored clearance of senescent neutrophils by tissue-resident macrophages limits organ aging.” The subject is the restored clearance — so the title says that restoring the clearance limits ageing, which is a good deal closer to the coverage than we allowed. The abstract then closes on impaired clearance as “a reversible driver of organ decline in aging.”

We had the grammar backwards and reprimanded a science desk for it. The entry now says so in place of the complaint. Two things came out of the same abstract: the paper has a human arm after all — elevated EP2 and senescent neutrophils observed in aged and diseased human tissue, with no human intervention — and the mouse ages and dosing duration we printed appear nowhere in it, so they are struck from the entry rather than carried on a summary. The withheld author list is resolved: the ordering we distrusted was right, and all thirteen names are printed.

The Nature Materials entry opened: run a current through a conductor, put it in a magnetic field perpendicular to the sample, and a voltage appears across the sides — that is the Hall effect, discovered by Edwin Hall in 1879, hence “century-old.” Then it reported this paper as finding that voltage with the field in-plane. It does not. The paper is about the anomalous Hall effect, which arises from the material’s own magnetisation and from symmetry, not from an applied field. Nothing in it touches Hall’s 1879 result or the sensors built on it. The paper stacks a low-symmetry semimetal on a ferromagnetic insulator so that a single mirror plane survives, breaks that symmetry with in-plane magnetisation, and gets a Hall response tracking both magnetisation components — offered as “a pathway for engineering tunable, symmetry-driven Hall effects.”

The flagged open question is answered, and in our favour: the prior in-plane literature exists in force. Papers reach back to 2022, with independent observations published in 2024 and a run of further work through 2025 and 2026, including a unified symmetry framework at the end of July — all now cited by number in the entry. So “century-old assumption overturned” is release language and appears nowhere in the paper. Being right about the release did not stop us reproducing its error one paragraph earlier.

The heliosphere entry printed 13–14, 6–7 and 2–3 million years ago as the times the solar system crossed cold interstellar clouds. Those are not cloud encounters. They are stepwise cooling intervals read out of deep-sea sediment cores — the independent half of the argument — and the review says their driving mechanisms are “subject to ongoing debate” and that such cooling might be triggered by heliosphere collapse. Getting it backwards turned a proposed correlation into a reported observation, which is the specific move this page exists to refuse.

Two more, from the same abstract. It is not a group: the review has a single author, Merav Opher, who directs NASA’s SHIELD centre — we had written “a NASA-funded group.” And the superflare and nitrous-oxide strand is not in this paper, which we had filed as a “companion strand” of it; the review’s stated span is the Sun’s last ten million years, and almost certainly a second piece of work was folded in by the summaries. It has not been identified, and the entry says so. The title, author, volume, pages and DOI were all filed as not obtained and are now printed. On the figure: nitrous oxide is 273 times as warming as carbon dioxide over a hundred years, per IPCC AR6 — not 300, and the real defect was never the digit but the missing timescale.

The Asgard entry said the archaeon was imaged in contact with bacteria “joined by nanotubes and apparently exchanging material.” The imaging shows the connection — through tubular fibres assembled by the bacterium, not jointly. The exchange is inferred from complementary genomes: the archaeon can make hydrogen, acetate, formate and sulfite, the bacterium makes amino acids and vitamins, and the paper’s verb is that these “could be exchanged in a syntrophic partnership.” We had a genomic inference standing where an observation should be. We also said the group reported isolating the archaeon; it reports a highly enriched culture at 89%, which is not an isolate.

Everything the entry had withheld is now obtained: the DOI, the full fourteen-author list, the volume and pages, and both species names, which are printed for the first time — the archaeon Nerearchaeum marumarumayae and the bacterium Stromatodesulfovibrio nilemahensis. The 9 April 2026 date we filed from a summary is right.

The Conotubus entry printed its lag as “1 d, a floor,” because no online-first date could be found and the issue date had to stand in. The online-first date is 24 August 2026, so the lag is a measured 9 days. Three claims corrected with it. The proxy is molybdenum — extremely low δ⁹⁸Mo in pyritised fossils — and the sulfur isotopes do a different job, showing that pyritisation preserved the original signal and that microbial sulfide oxidation was going on in the habitat; we had both isotopes doing the same work. The bacteria drawing down a lethal sulfide load is not in the paper; that was a summary’s addition and we repeated it. And the paper’s modern comparison is cold-seep tubeworms, where our own frame reached for a hydrothermal-vent animal. The full twelve-author list is obtained. The paper’s hedge, which the entry said was the thing it could not read, turns out to be stronger than ours was: Conotubus “may have engaged in chemosymbiosis.”

Attention patterns in ADHD as a creative advantage. It goes straight at the paradigm and we are not filing it. The sourcing traces to a congress presentation and a cross-field synthesis rather than a study; the summaries conflate a 2025 congress item with a separate 2026 design-task study, so a digest could not tell you which finding it was describing; and the framing on offer — creative activity as a therapy that channels defocused attention while strengthening focus and self-control — sits one inch from regulate yourself out of your material conditions. Here the superpower frame is not a risk downstream. It is the headline.

“Each extra hour of prolonged sitting linked to 9% higher cancer death risk.” Held because the sources disagree on the journal — the summaries name both PLOS One and PLOS Medicine for the same paper, and a scan that cannot say who published a result has not traced it. Held also because an observational hazard ratio about how a day is broken up quietly converts a question about who is permitted to stand up at work into personal advice.

A single dose of engineered immune cells said to protect aged mice for a year is declined as cure framing, and named on the page rather than dropped, because the filed macrophage item is the same field seen from a better angle and a reader is owed the neighbouring claim we chose not to repeat. Octopus, whale and corvid cognition is held for undated round-up sourcing — the same fault as edition one’s cow with a brush, held a second time on purpose so the pattern is on the record twice. That category is where our sourcing standard comes under the most pressure, because the items are the most fun.

Checking where a bats entry would go meant re-deriving the there is no standard X figure. It is fifteen field guides. CLAUDE.md says twelve; the guide to untranslatable words says thirteen in its own refusals. Both predate three more guides — bird, rainbow, and way to be moved — and those three were read here to confirm they arrive at the formula rather than merely mention it. The grep returns sixteen files; the sixteenth, the unfinished-animals guide, matches while explicitly declining the formula, which is why the answer is fifteen and not sixteen.

The standing rule held exactly as written: a count in the working notes is a lead about the site, not a fact about it. Both stale figures are named on the digest rather than quietly corrected, because a count that was wrong in two places at once is evidence for the rule.

What we said. The crow card listed self-recognition among the things a corvid does without a neocortex, and its field notes said crows and ravens “use tools, plan for the future, and recognise themselves.”

What is right. Prior, Schwarz and Güntürkün (PLoS Biology 6(8): e202, 2008) marked five hand-raised magpies. Two showed statistically significant mark-directed behaviour, one a single spontaneous instance that did not reach significance, two none — with the authors’ own printed caution that the sample is too small for a definitive estimate of the distribution in the population. Soler, Colmenero, Pérez-Contreras and Peralta-Sánchez (Journal of Comparative Psychology 134(4), 2020) then failed to confirm it in a close, larger replication.

Hedged narrowly, in both directions. The replication did reproduce more social and more self-directed behaviour at the mirror than at the cardboard control; it failed specifically on the mark test. So “debunked” would be as wrong as the original overclaim, and the card says so on its face. Tool use and next-day planning are untouched, and were strengthened rather than left vague — Kabadayi & Osvath and Gruber et al. are now named on the card.

What we said. The elephant card listed self-recognition alongside memory, grief and culture.

The first thing wrong is one we had never thought to check. Every figure on that card is the African elephant (Loxodonta africana). The mirror study — Plotnik, de Waal and Reiss, “Self-recognition in an Asian elephant,” PNAS 103(45), 2006 — is of three adult female Asian elephants (Elephas maximus) at the Bronx Zoo. And that is not pedantry, because the study’s own control depends on it. The paper notes that male African elephants often touch their heads owing to temporal gland secretion, whereas in female Asian elephants that gland is bilateral, vestigial and nonsecretory — which is precisely what let the authors rule out a general tendency to touch the marked area. Carry the finding across to an African elephant and you reintroduce the confound the species choice removed.

The second is the strength: one of three. Happy passed the mark test on the first day of marking, with a significant shift in her head-touching distribution. Maxine and Patty did not. Re-marked two months later, none of the three touched mark or sham. The paper’s title is singular for a reason.

Graded more gently than the magpie, deliberately. There is no failed independent replication here, and the authors answer the one-of-three objection themselves: even in chimpanzees, the most extensively tested species with mirror self-recognition, fewer than half of individuals may pass on some studies — and repeated mark tests on one animal are considered uninterpretable, because they compromise the mark’s salience. All three elephants reached the third and fourth stages of the standard progression and showed a total absence of social responses to the mirror image, which is evidence in its own right. So the card now says what is actually true: mirror self-recognition has been demonstrated in one Asian elephant, it is a real and careful result, it is not a general fact about elephants, and it is not a fact about this card’s animal.

A figure tightened while we were in there, and its source cited for the first time. Herculano-Houzel and colleagues (Frontiers in Neuroanatomy 8, 2014): 257 billion neurons, 97.5% of them (251 billion) in the cerebellum — the card had said ~98% in four places. The cerebral cortex has twice the mass of ours and holds 5.6 billion neurons against our roughly 16 billion. That last figure is new to the card and makes its point better than the wording it replaces.

What we got wrong. Field Guide No. 11’s corvid card lists self-recognition among the things a bird does without a neocortex — and it lists it in the column a reframe table uses to say what is actually the case, which is the worst available place for a contested result. Its field notes say crows and ravens “recognise themselves.”

What is right. Prior, Schwarz and Güntürkün (PLoS Biology 6(8): e202, 2008) marked five hand-raised Eurasian magpies and gave them a mirror. Two showed statistically significant mark-directed behaviour. One showed a single spontaneous instance without reaching significance. Two showed none. The authors’ own caution, in the paper, is that the number of birds tested is too small for a definitive estimate of the distribution within the population. Then Soler, Colmenero, Pérez-Contreras and Peralta-Sánchez (Journal of Comparative Psychology 134(4), 363–371, 2020) ran a close replication with a larger sample. The mirror-versus-control difference in social and self-directed behaviour reproduced; the mark test did not — self-directed behaviour was more frequent in front of the cardboard control.

Held in both directions. That is not a debunking either: what failed is the mark test specifically, not every mirror response, and the new guide’s magpie card says so in the table. Tool use and next-day planning are unaffected and stand.

Who caught it and how. Nobody wrote in. It came out of checking a favourite fact while sourcing the new guide — which is the whole reason for this page. A first web search returned the birds’ names and a “two of five” framing partly from what look like content-farm pages; the figures above are from the open-access primary, whose own sentence is three out of five birds showed at least one instance of spontaneous self-directed behavior. Not the same claim, which is why the card gives the breakdown rather than a ratio. FG 11’s own wording, and a second instance the audit turned up — its elephant card lists self-recognition too, on evidence we have not yet checked — are queued as their own pass rather than patched where noticed. Fixing only the instance you spotted is the failure mode this file keeps recording.

The 2026 framework’s own copyright page says it “takes effect from 1 November 2026.” BACP’s Preparing for the 2026 Ethical Framework page says it becomes mandatory at midday on Tuesday 3 November 2026. Both are BACP’s, both are published, and they are not the same date — so both are on the page, per the house rule about printing an ambiguity rather than resolving it silently.

The word counts are the essay’s central claim, so the method is printed with them: both PDFs converted with pdftotext -layout, counted with grep -oiw, extracted lengths 6,825 and 4,124 words including front matter. Stated so a reader can reproduce it, or contradict it. A count you have not run yourself is a claim somebody made.

What we published. That a critic’s reading suggested “a central nervous system disorder with drug targets” was “well past what was measured” — framed, in our summary, as headlines outrunning a paper we admitted we had not opened.

What the paper says. Its abstract: “This study provides robust genetic evidence defining fibromyalgia as a central nervous system disorder.” It says so again in its closing paragraph and names two therapeutic targets. This is not a science-communication failure. It is a scientific claim by the authors, which may still be too strong — but arguing with a paper and arguing with a press release are different jobs, and we filed the wrong one.

The rest of the critique, adjudicated. Three figures hold and every one is the paper’s own: OR 1.090 for the HTT variant (“about 9% increased odds”), polygenic score AUC 0.59, and prevalence ranging sixfold across cohorts, 1.2% to 7.5% — which the paper discloses in the second sentence of its Results rather than concealing. One figure is wrong: heritability is 10.4% (95% CI 9.8–11.0), not “less than 10%.” One is unresolved: the one-billing-code claim needs the Methods supplement.

The Kanzi card in Glimmer Wire’s seedbed was written from a single search summary and never checked. Checked now. The citation holds — Bastos & Krupenye, Science 391(6785):583, DOI 10.1126/science.adz0743 — and so do the 68.9% grape figure, Kanzi’s death in March 2025 aged 44, and the six-to-nine-million-year claim.

What was missing is the part that matters. The study shows Kanzi could follow a pretence a human set up; it does not show he could start one. Comprehension, not production — and that is exactly the distinction a card like this gets flattened into “a bonobo played pretend” downstream. The skeptic we had gestured at vaguely is Daniel Povinelli, and his alternative is specific: that Kanzi tracked which cup the experimenter had touched most recently. Both are now on the card, with the real-juice figure (77.8%) we had given only in words.

One thing stays unresolved and is marked as such. The institutional record dates the paper 5 February 2026; the journal’s own citation block says 10 February. An online-first versus issue split we did not settle, so the card cites volume and issue instead of a day. A lead written from a summary inherits whatever the summary left out — including, here, the sentence the authors were most careful about.

What we said. Glimmer Wire filed this paper that morning describing it as finding that “the limbic system and neocortex do not stack; they trade”, and gave its date as “vol 12 no 17, 2026, approximate.”

What the paper says. The inverse relationship is specifically between the olfactory system and the neocortex. The limbic components — amygdala, septum, schizocortex, hippocampus — covary positively with the olfactory system and with each other, and sit on its side of the trade. Calling the whole limbic system the thing that trades against neocortex inverts where half of it belongs. The date is 22 April 2026. Both were caught by reading the paper, which the digest had said plainly it could not open.

One figure on the new page is second-hand and flagged there. The Cesario, Johnson & Eisthen textbook counts — 20 books sampled 2009–2017, 14 discussing brain evolution, 86% carrying at least one inaccuracy, two matching consensus — come from summaries of that paper rather than the paper, which would not open. Its abstract, quoted verbatim, supports the claim in kind but carries no numbers. The colophon says to treat the counts as provisional, and the open item is logged.

Glimmer Wire filed this story that morning graded plausible, sourced only to the NASA release, with a standing notice saying no paper on the page could be opened. The paper was then obtained and read in full, and the entry is now graded verified — the first on this site to make that move. Two things we published are corrected by it.

We wrote “nobody knows why it arrived now.” That overstates the authors. They name two candidate mechanisms and say they cannot discriminate between them, which is a different and more interesting condition than not knowing. And the formation window is bounded at mid-2012, when the southern hemisphere stopped being observable from Earth — not 2017, which is circulating in coverage. A third thing is worth naming without calling it an error: the widely-quoted 16,782 km is the wave’s mean zonal wavelength. For a ten-sided figure that is within about 2% of a side length, so “one side exceeds 16,700 km” survives — but it is not the measurement the paper reports.

The digest’s own counts were updated rather than left to rot: the standing notice, the hero dateline, the edition tally and the colophon all said none verified, and all four now say one. That is exactly the class of stale prose figure no gate on this site can read.

What we said. That a <br> with no space beside it makes a heading announce as one run-together word — that No. 84’s cover read as “There Have BeenSo Many Earths”, that Bone Song’s title read as “BoneSong”, and that 163 headings across 48 pages were affected.

What is actually true. textContent does run the words together — that part is real and checkable. The accessible name does not. Chrome inserts the whitespace itself when it computes the name, and it does so in every shape we could construct: a bare <br>, a <br> between two <em>s, between two <span>s, after a comma, and inside a link and a button as well as a heading. Measured directly with the browser’s own accessibility tree, against copies of the pages from before and after the change: “Bone Song” both times.

How the error was made, which is the part worth publishing. The finding came from a tool that reports what it calls the accessibility tree, and the line it printed was the element’s textContent. We treated the label on the output as the thing itself. This site already has a rule for exactly this shape — aggregators and summaries are leads, not citations; go to the primary — and it had not occurred to us that the rule covers our own instruments. A tool that says “accessibility” is a lead. The accessibility tree is the primary.

The sweep is not shipping, because it was measured and it does nothing. The 347 edits were written and checked before being thrown away. They were pixel-identical at 2× on the worst-affected covers; the search index came back byte-identical, 2,069 records either way, because its builder already pads these boundaries and had always indexed the heading as “Bone Song”; and the accessible names were unchanged because they were already right. Three fragments that looked like proof of harm in the search index were false positives: nevertheless, everythingisaremix.info, and — the one that should have been the tell — our own changelog entry, quoting the mistaken claim.

The count was wrong too, in the other direction. The first pass reported 163 breaks on 48 pages. Re-derived properly — by reproducing textContent semantics rather than pattern-matching the source, which cannot see </em><br> and mishandles a heading with two breaks — the real figure is 348 across 74 pages. More than double. Both numbers are in this changelog now; the first one is wrong and is left standing above with a pointer here, which is what this page is for.

One thing the exercise did establish. The related fault written down in our notes — two adjacent <span>s with nothing between them producing “NeurodiversityParadigm” — is a real accessible-name defect, and it reproduces. It is a different fault from the one we misdiagnosed: no <br>, nothing for the browser to turn into a space. And a footnote we cannot fix from here: because Chrome inserts that space, Underground’s cover, which splits the single word across two lines, announces as “Under ground”. That is the browser being helpful in the one place we would rather it were not.

Scope of the check. Chrome only. We could not test WebKit or Gecko here and are not claiming they behave the same, only that the defect we published does not exist in the engine we can measure.

The famous “thirty seconds later and it would have hit ocean” is not a research finding. It traces to the 2017 BBC documentary The Day the Dinosaurs Died, made around Jo Morgan and Sean Gulick’s drilling of the Chicxulub crater — excellent work that does not contain the number — and the popular-science article that put the figure in its own headline states plainly that there is no causal evidence for it. Green flags his scepticism about it in his source list. The zine refuses it rather than using it. What the papers actually say is better: Kaiho & Oshima (Scientific Reports 7:14855, 2017) found the killing mechanism depended on the target rock being rich in hydrocarbons and sulfur, present under only about 13 per cent of Earth’s surface — their paper’s own title is the low probability of mass extinction — and Collins et al. (Nature Communications 11:1480, 2020) put the trajectory at 45–60° from horizontal, from the north-east, a range that lofts more climate-altering gas per kilogram of impactor than either a grazing blow or a vertical one. A stopwatch is not a smaller version of thirteen per cent. It is a different kind of claim — and it is the kind that makes contingency feel like a near-miss in a film. Filed alongside Too Good to Check. We have also declined the press framing that rounded Collins’s range to “sixty degrees, the deadliest possible angle”.

The video’s citation for the earliest evidence of life is disputed, and the dispute is not in its source list. It cites Nutman et al. (Nature 537, 2016), reporting 3.7-billion-year-old stromatolites at Isua in Greenland. In 2018 Allwood et al. went back to the same rocks and published in the same journal, reading the structures instead as deformation features in carbonate-altered metasediments formed long after burial — and the plainest piece of evidence is that some of the bumps point downward, which stromatolites, growing toward the light, do not do. The floor the field broadly accepts is younger and further south: the 3.48-billion-year-old Dresser Formation in the Pilbara, where stromatolites, microfossils and fractionated carbon and sulfur isotopes agree — and even that has had a volcanogenic-pseudofossil challenge. The joint is marked contested, and the page says on its own face that the date it is least sure of is the one everything downstream sits on.

Three things are hedged on the page rather than smoothed. The isotopic crisis — models say the Moon should be made mostly of the impactor, while its oxygen isotopes are almost indistinguishable from Earth’s — is stated as unresolved instead of skipped, so the zine says a collision took place without claiming to know what hit us. The claim that the Moon stabilises Earth’s tilt enough to matter for complex life is given as genuinely contested (Laskar, Joutel & Robutel, Nature 361, 1993, for; Lissauer, Barnes & Chambers, Icarus 217, 2012, much weaker) and nothing in the argument rests on it. And the photosynthesis-to-oxygen lag is written as something like a billion years, because the figure moves with whichever origin date you accept.

The title is ours, compressed from Green’s, and the colophon says so rather than letting a paraphrase pass as a quotation. His verbatim line, quoted in full on the page, is “there have been a lot of earths and this is just one of them and there could have been a lot of other earths”; he also calls the history of the planet “a sequence of impossible worlds.” And a guard spread carries the naturalistic-fallacy problem in the open: the early appearance of life here is not evidence that life is easy, because we could only ever have found ourselves on a planet where it went this way and went this way early enough. That is a selection effect, not a plan. Rare and inevitable look identical from inside the one case you get to observe.

“A fine carpet of green grass stamps the inhabitants as good neighbors, as desirable citizens” is not Elizabeth Kolbert’s. It is regularly credited to her, including in the essay that prompted this zine. She is quoting it. The line is Abraham Levitt’s, of Levittown, by way of Ted Steinberg’s American Green (2006) — one of the two books Kolbert’s 2008 New Yorker essay reviews. This is the same shape as the misattribution this site shipped once before, when a line of Martin Luther King Jr.’s was credited to Baldwin because a secondary summary was trusted. We have not opened Steinberg at primary, so the line is graded plausible rather than verified, and the zine prints the chain instead of collapsing it.

Etta Kralovec’s Schools That Do Too Much is 2003, not 2002. The 2002 date is widespread, including in the source essay; Beacon Press’s own record gives 2003. And the sentence usually quoted from her — that how schools spend time and money are not facts of nature but decisions that can be unmade — we could not confirm against the book, so we did not print it. Spread 10 paraphrases the argument and says on its own face that it is a paraphrase. If the book comes to hand the line can be restored with quotation marks it has earned.

A number printed as a range because the headline is disputed. Turfgrass cover in the United States is often given as ~163,800 km² from Milesi et al. (Environmental Management 36(3), 2005), usually attached to the claim that lawns are the country’s largest irrigated crop. That figure is a modelled satellite estimate rather than a census, and the “largest irrigated crop” framing is contested — including by the turf industry’s own trade press. Kolbert gives nearly fifty thousand square miles. The zine prints both, takes the range, and declines the headline.

We also looked for the study that would support the popular claim and did not find it. The zine says that, rather than implying no such study exists. It is not an argument against gardeners: taproots do open channels. The dispute is about the size of the claim.