Stimpunks.org Changelog for Week 37 2026: From the Gate to the Instrument, From the Byline to the Label

This week a site that did not exist on Sunday went live on its own domain and published twelve sheets, a mission, a manifesto, a privacy policy, a search, an accession register of its own errors, and ten automated gates. queering.earth is a Stimpunks Foundation × More Realms collaboration reading art, literature, poetry, politics, people, and history through a queering lens — sibling to ★ stuff and built as its deliberate opposite, a herbarium sheet in daylight against a night sky. Alongside it ★ stuff shipped zines numbered 89 through 110, two field guides, a specimen sheet, a Trigger, a Glimmer Wire, and a shelf for four-year-olds. And both sites spent the week being audited against a specification — which is how we found that a great many of our checks had been passing without measuring anything at all. A gate that never launched the browser it reports on. An overlap sweep that had only ever looked at one screen size, while 105 of 108 zine covers were broken on phones. A control whose two green lights both came from code that skipped the thing that was broken. The site never changed across three runs of the tap-target audit; only the instrument did. The same week we wrote a crawler policy saying yes to everything and asking only for attribution, and discovered we had been redistributing two typefaces with no licence and crediting their designers to a search engine. A check that does not run the thing it claims to measure will tell you the site is fine, every time. And an attribution policy is only worth what it costs you on the day it convicts you.

Throughline
The Work
Week 37 Changelog

Throughline

A check that never exercises the thing it claims to measure will report PASS forever, and the PASS is the worst part. On Queering Earth the three rendering gates each polled a debug port until something answered — and an orphaned headless Chrome had owned that port since 7 September, so for five days every typeface run bound nothing, drove a stranger, measured in Chrome 152.0.7977.77 while .84 was installed, and came back green. On ★ stuff the overlap gate had been measuring 1280×900 and only 1280×900: at 375px there are 313 text-on-text collisions across 103 of 205 pages, every one of them on spread one, plus the motif box sitting across cover type on all 98 covers that have one — 105 of 108 zine covers broken on phones since the first cover shipped on 17 July. The Reading panel on Queering Earth shipped with two green lights and could not be opened by a human being, because every pre-flight check called summary.click(), which dispatches straight at the element and skips hit testing — the one thing that was broken — and the screenshot proving it worked was taken after setting open = true from the console. The typeface gate reported passes across 25 pages while rendering the default face and nothing else, and only when nine faces went into the picker did it surface that the wordmark had been laid from the start edge and sitting 34px right of the page’s centre line since the day it was drawn, because text-align distributes leftover space and a line wider than its box has none. The contrast sweep was blind to 218 elements collapsed to opacity 0. The storage check located settings by their on/off classes, so a two-value setting would have gone undocumented on the privacy page while the check reported green. A shallow clone made a generator date a 38-day-old site as “170 pieces · 6 days,” and the tool exited 0 both times. A probe in the decision-log audit reported a converter fix missing from all eleven commits, uniformly and confidently, because zsh ate a colon and a tidy 2>/dev/null hid the fatal error — suppressing stderr converts an error into an answer, and a uniform answer reads as a finding rather than a broken tool. The fixes are never more checks. They are checks that run the real thing: a real pointer, a real viewport, a real browser you launched yourself.
The best decisions in a brand-new site are the things it refused to build. queering.earth refused the tagline it started with, refused a CMS, and refused a byline. Helen Edgar declined the byline on her own essay’s sheet, and she was right, because the site’s own colophon already said so: a label is Maker plus Read by, a specimen is collected rather than written, so sheets have labels, not bylines. The CMS answer is not yet, with the trigger named so it is not relitigated — a CMS earns its cost when content is uniform and layout is fixed, and here the layout is the argument. The accessible copy is not a second copy: the type leans and refuses to sit on the line, all of it in CSS over ordinary semantic HTML, switched off by a plain view control that is one class on one document, because two copies of the same words drift and the accessible one is always the copy that rots. A corner-floated accession stamp was designed, measured and refused partly because nothing there could see it land on the prose — and refusing a design for want of a gate is a reason to build the gate, which then caught the refused design on its first run. A link hidden inside the header illustration was refused on measurement: at 375px one viewBox unit is 0.713 CSS px, so reaching a 44px target would need a 62-unit invisible disc, bigger than anything in the drawing. Doodles around the heading rail were drawn at Ryan’s suggestion and refused on the risk he named in the same breath, because that margin is where the skip link appears on focus. And a treatment that varied tick length without meaning was refused for exactly that: it is the prettiest way to lie.
Wednesday’s accessibility audit published two figures that were both overstated: “282 of 299 <th> carry no scope” counted <thead> in the grep and should have read 201 of 218, and “28 of 83 tables have no caption” was literally true and materially misleading, since 24 carry an aria-label that names them and only four tables on the site had no accessible name by any means — presenting 28 as an accessibility failure counted a naming convention we had chosen as an absence. Twice in one day we published that a paper could not be read and twice it could: Nirenberg & Matthaei was held as page images in the reader our fetcher hit and as an OCR’d PDF at Europe PMC, a route our own working notes already name, and Larivière, Pontille & Sugimoto came back as soon as Ryan put the PDF in the inbox. “We could not read it” is a claim about the world and needs the same standard as any other claim on these pages. And the loudest one ran outward: an entry claimed Queering Earth’s Markdown mirrors ship with no staleness check. False — and it was a public claim that somebody else’s work was deficient, when their gate regenerates every derived file into memory and compares, and its error message for that case even reads “a stale .md is the quiet one.” Their arrangement is arguably better than ours, which is why ours now copies it.

The Work

A Site That Did Not Exist on Sunday

queering.earth went from nothing to live on its own domain in an afternoon, and by Saturday held twelve sheets. The display face is Fraunces, picked for its WONK axis — letterforms that deviate from the norm, exposed as a typographic setting — so a site about queering normativity makes its argument in the type rather than only in the prose. Star Stuff is a night sky; this is a herbarium sheet in daylight: warm vellum, ink, moss, and rust, specimens laid out like a Victorian plate, because “the same site in green” was the failure mode to design against.

The sheets are readings, not essays. The Die Uncast reads Christina Rossetti’s 1861 Promises like Pie-Crust alongside 2026 work on a neuroqueer ethic for Autistic adults’ intimate lives — and the 1896 first printing has a colon at line 20 where every circulating copy has a full stop, read in two separately scanned copies to be sure, because the difference between stopping and setting up what follows is the poem’s whole argument about what is being declined. Wyrd, Weird, and the Invention of Normal is Helen Edgar’s More Realms essay mounted as a zine wall inside a herbarium register. Miranda, To Be Wondered With was laid out and adopted as written, with three devices rebuilt for accessibility rather than reproduced — the correction uses real del and ins elements instead of a line-through on a span. The Swell and the Dwell reads Dickinson and answers her in verse, and its findings are all absences: the poem has no indicative verb in it, poets.org prints line 11 without “but”, and two agreeing scans of page 97 turned out to be one witness, because they came out of one library. Sheet No. 8 is the ghost pipe — a plant with no chlorophyll that feeds through fungi, named in 1753 for the single turn of its stem — following Helen Edgar’s essay for the fifth anniversary of Dinah Murray’s death.

Then the flower codes: violet, pansy, green carnation, lavender, all four worn or sent as a way of saying it when saying it in words was dangerous, gathered as a set by Sarah Prager and credited to her. The reading is ours: a worn code fails by succeeding. It works until enough people use it, and the number who make it useful is the number who make it legible from outside — at which point it is a charge rather than a description, and the last move left is to wear it in the open until it goes back to being a description. Two of the four plates are the wrong object and both captions say so: the violet is Viola pedata standing in for Viola odorata, and no plate of a green carnation exists anywhere, because a green carnation is a white one stood in dye.

Sheet No. 9 reads Carroll’s Snark through Deleuze’s empty square — quoted from the French, because Lester’s English translation was not in the house — mounted with Henry Holiday’s Ocean-Chart, an empty square drawn ninety-three years early. Holiday’s own 1898 article corrected the sheet rather than confirming it: Carroll called the Boojum drawing “a delightful monster, but inadmissible” — he liked it and refused it anyway, and every secondary account drops those three words. Sheet No. 10 reads Octavia E. Butler’s hyperempathy through Sami Schalk, whose chapter turns on its having no real-world equivalent — and the Autistic people who have been describing exactly that under the same name for years. The sheet endorses nearly all of Schalk and argues with one move: her protection against stereotype is bought with unfamiliarity, so it is only available to a reader who does not have the thing, and for an Autistic reader it inverts. The fairest account is that she read a literature that had not caught up with community knowledge, and the 2024 study says so itself.

My Gloomy Sunflowers is Sheet No. 11 — Townes Van Zandt, Vincent van Gogh, Sylvia Plath and Kurt Cobain read against the myth that made them — and the argument is Hannah Gadsby’s and Shoshana Kessock’s before it is ours. Both are credited by name. What the sheet adds is evidence: five primaries read at source, and in three unrelated archives the same shape — a 1964 clinic reading a young man’s gentleness as a symptom, a painter calling his best canvas “painted with calm” in the letter where he says his mind is stupefied, a biographer suggesting that the most famous anguished voice of a generation may be a man’s spine. In every one the work is read as a symptom, and in every one the record says the work was the thing going well. It does not claim the four share a diagnosis, because what they share is having been treated as a diagnosis and read afterwards as one. Nothing is quoted from Polly, because every word of that song is spoken by the man who committed the crime, and a ruled box with a herbarium label would give a rapist’s speech the mounting this cabinet gives to Woolf and to Plath.

Sheet No. 12 closes the week on an asymmetry: attention that pools on one person and will not redirect is a masterpiece when a novelist does it and a symptom when an Autistic person does it. Dorothy Tennov was read first-hand, which corrected the sheet — crystallisation is Stendhal’s, not hers — and also produced the finding it is built on: her 1979 criteria already carry “intensification through adversity” and “inability to react limerently to more than one person at a time.” The barrier measured, and single-channel attention written into a definition a quarter of a century before monotropism had a name.

Around the sheets, the housekeeping became argument. The accession register is drawn as the document a herbarium already keeps for this — Mounted, Re-determined, Label corrected, Cabinet — and twelve of its first thirty-four entries are our own errors, because a site that publishes an attribution ledger owes the account of its mistakes to readers, not just to sessions. /mission is thirteen aims, each ending by naming the page already keeping it, and each paired with an observation that would falsify it — an aim you cannot fail is not an aim — and most are not hypothetical. The manifesto is adapted under its own licence from the Neuroqueer Learning Spaces Manifesto, a nine-author CC BY-SA document initiated by the two people who make this site, with every contributor named at the foot. Sourcing one of its quotations found a drift in our own house: Play Radical reads “a translation of my world to yours,” where stimpunks.org has “into yours” with an undeclared ellipsis. A translation to somebody is offered; a translation into them is done to them.

How we quote was written before the sheet that needed it, and neither argument carrying it is a legal one. A refusal to quote produces paraphrase, and paraphrase with the attribution still attached is the tightened source the fact-check ledger exists to prevent. And “go and listen to the recording” is an audio-only requirement that shuts Deaf and hard-of-hearing readers out of the argument entirely. The sourcing behind it ran the same way: a disputed two words in Plath’s journal were settled against two further editions, and the methodological half is worth more than the fix — our glossary closes “which ever” to “whichever” and a spaced en dash to an em dash, which is a copy-editor’s reflex applied to a woman’s journal in the edition that exists precisely not to do that. It was catchable only because the first copy rendered every dash as a hyphen and so announced its own unreliability. A tidy copy would have agreed with our page and closed the question.

And on Saturday the review workflow got built around one fact: Helen reviews the drafts, cannot use GitHub, and does not edit HTML. Reading a draft never needed an account; giving feedback did, because a pull-request comment thread was the whole answer — so reason three of the CMS deferral is struck and the other two stand. Drafts live at the root of a long-lived branch, which makes a draft byte-identical to the page that will ship. The review layer is three lines in one head block: every block grows a visible focusable Note control, notes live in the reviewer’s own browser and leave only through Copy my notes, as plain text carrying the section, a deep link, the quoted line and the reviewer’s words. No endpoint, no account, no third party.

The Instrument, Not the Thing

Three of Queering Earth’s gates render pages in a browser. Each polled a debug port until something answered, and an orphaned headless Chrome had taken that port on 7 September — so for five days every run bound nothing, drove the stranger, measured in a Chrome version that was not the installed one, and came back green. The launcher now refuses a busy port and a dead child, and removes its profile properly, since killing a process only sends it a signal. It was found while making the checking cheaper: the bill was timed rather than guessed — generators 4.8 seconds, six offline gates 5.6 together, and the width gate alone 4 minutes 22, three quarters of everything. One runner now regenerates and runs every gate, the three browser gates at once, scoped to what git says moved: a one-page change is 24 seconds and the full sweep 4:47.

On ★ stuff the same shape had been shipping since July. The overlap gate measured 1280×900 and nothing else, and its own header had said in advance that it would miss exactly this — a limitation a tool documents is still a limitation. At 375px: 313 text-on-text collisions across 103 of 205 pages, every one on spread one, plus the motif box sitting across cover type on all 98 covers that have one. The fix is one shared rule below 600px rather than 108 per-page ones, on the reasoning that the desktop placement is a composition and the phone failure is the absence of the margin that composition assumed. Then the argument was tested by a page nobody chose as evidence: a zine written the same day by a session that knew nothing about the fix landed on main first — six collisions on its cover at 375px with the shared rule removed, zero with it. The second viewport immediately exposed a false positive in the same tool, which had been treating scrollable ancestors as transparent and calling 256 reachable table cells cut off. Baseline zero at both viewports across all 207 pages.

The Reading panel on Queering Earth is the purest case. It shipped, Ryan pressed it on the live site, and nothing happened. Inside a flex row, a real click over the summary arrives with event.target set to the <details> rather than the <summary>, so the disclosure never toggles. The mechanism is still not established, and the fix records the evidence instead of inventing a cause. Why it shipped is the part worth keeping: every pre-flight check called summary.click(), which dispatches straight at the element and skips hit testing — the one thing that was broken — and the screenshot showing the panel open was taken after setting open = true from the console. Two green lights on a control nobody could operate. The rule written out of it: a control is not verified until a real pointer has operated it. The keyboard path was then recorded as UNVERIFIED rather than passing, and updated to confirmed only after Ryan checked it live, because “I could not check this” is only honest for as long as it stays true.

Three pages on Queering Earth had been scrolling sideways since the day each was mounted — pushed out by DOIs, archive.org identifiers, sha256 hashes and file paths, strings with no space or hyphen in them — and all eight gates passed all three, so a ninth now sweeps every page at 320, 375, 768 and 1280 and against both paper sizes, because the hazard is attached to a routine: crediting a source adds one of those strings every time. The fix then broke the wordmark into Queering.Ear / th, and it had passed nine gates, so every measurement taken had been of the wrong quantity — none of the damage moved a document width. And the next screenshot produced the real finding: text-align distributes leftover space, and when a line is wider than its box there is no leftover, so the wordmark had been laid from the start edge and sitting 34px right of the page’s centre line since the day it was drawn. Nine faces made visible a fault that had been shipping with one. The typeface gate had the same hole the contrast gate had three days earlier: it reported passes across 25 pages while rendering the default face and nothing else. It now sweeps all nine, reading the face list from the file the font build writes rather than a hand-kept copy — and it was made to fail before it was believed.

The rest of the week’s instruments failed the same way and are worth listing, because the list is the argument. A tap-target audit was wrong twice before it was right — testing a link’s immediate parent under-detected the prose exemption and invented eleven failures; testing the nearest block ancestor over-detected it and made a real finding vanish. 5,591 targets, 0 failing both tests. The site never changed across the three runs; only the instrument did. A cache-skew gate’s first attempt at proving itself was a false pass, because the edit meant to loosen the stylesheet had loosened the catch-all with it, so there was no skew to catch. 229 gradient-clipped headings were reported invisible from a computed property and every one of them renders; the rule written for them was a no-op on 198 pages and is gone, because reading a computed property and inferring a rendering is the exact fault that repo is organised against. A count that reconciled the verse three ways had two of the three methods silently wrong in the same pass — a regex that did not survive shell quoting returned 0 for every file, and a hand-written counter blind to fenced code blocks read a shell line continuation as a line of poetry. The peer’s figures were right all along. A cross-check is not independent until it has itself been checked. And a shallow clone made a generator that dates pages from git announce a 38-day-old site as “170 pieces · 6 days” — caught by reading the diffstat, not by an error, because the tool exited 0 both times.

Two Sites, One Specification

★ stuff spent Wednesday being audited, and the first two findings were Level A failures that had been live since the site’s first page. No page had a skip link — 0 of 197, from 2026-07-17 onward — so a Tab-only or switch-device reader walked past all six controls in the nav cluster to reach the first word, on every page. Then the tables: 199 scope attributes across 81 of 83 source tables, captions for the four with no accessible name by any means, and a third defect found only by reading the accessibility tree — 161 of 208 table names read “misreads the The Biggest Mouth in the Sea”, because fourteen field guides label each table from a template ending ${entry.name} and most entry names begin with their own article. Correct English on both sides; the defect existed only in the assembled string, invisible to every gate and to anyone reading the source.

Writing a privacy page then changed the site before the page could be written. The draft sentence “on most pages your browser contacts nobody” was false: every page fetched its typefaces from Google, handing over an IP address and a user agent before a word rendered — on a site with no analytics, no cookies, no accounts, and a search that runs entirely in the reader’s browser. The fonts were quietly the largest thing we were giving away, on pages that had nothing else on them. All four families now come from our own domain, and the verification found a pre-existing fault on eight pages: the front page drew its masthead tagline in Menlo, three zines drew their nav in a system face, four more used a family they never requested — masked the whole time because both families are installed on this machine. The privacy page declines the word anonymous for Netlify’s analytics and corrects the same morning’s entry: calling 66 embeds moved to youtube-nocookie.com “the cheapest privacy win” reads as no tracking, where Google’s own text promises only that a view will not be used to personalise. We read the domain name and not the small print — this site’s own too-good-to-check reflex, aimed at infrastructure.

The supernova on one of the three doors into the collection was a 2.3 MB JPEG displayed 400 pixels wide. Every raster is now a <picture> chain, and a modern browser fetches 600 KB where it fetched 4,216 KB: 86% less. The cost is stated rather than buried — no print rule constrains a figure, so it now prints at roughly 216 dpi against 736, deliberate, because 736 dpi is past what a home printer resolves and every phone reader was paying for it.

Then a class of fault nobody had a check for. Reading all seventeen collection descriptions in a row to write an index for language models is how we found one cut off mid-sentence since the day it shipped: the quotes around star were plain " characters inside an attribute already delimited by ", so the parser closed content at the first inner quote, kept nine words, and turned the rest into stray attributes. Search result, link unfurl and shared card all showed the same fragment. The source reads perfectly, and a meta description has no rendering to look wrong. Shipped with the ninth check for exactly that class — the browser silently handing the reader a different document than the source describes — came /llms.txt, curated at 39 links rather than duplicating the sitemap, and /.well-known/security.txt under RFC 9116, which required switching on GitHub private vulnerability reporting in the same pass, because an issue there is world-readable the moment it opens, so the ordinary route would publish a vulnerability rather than receive one. Its mandatory Expires field is not hand-written: one constant, the file generated from it, the countdown printed on every run, and the build failing 30 days out — a date-stamped claim maintained by memory is this house’s signature fault with a one-year timer attached.

The morning’s entry had said a full Content-Security-Policy was impossible here. Both premises held and the conclusion was wrong, because the spec names a third option in the same sentence and it is hashes: 157 inline scripts is 49 distinct ones, most of them the same pager repeated across a hundred zines. The measurement that settled it was one command nobody had run. It is generated and gated, because a hash list is a second description of the site and its failure is silent — a script whose hash is missing simply does not run. Tested against the site working as well as against the attack, because a policy that breaks a hundred pagers is worse than none. An hour later all 233 inline event handlers were gone and 'unsafe-hashes' dropped out of the policy on its own.

Thursday went to reachability. Windows Contrast themes replace every colour on a page, and Chrome does not force SVG fill — so 4,862 diagram labels across 133 pages kept the light ink they were drawn with while the ground beneath them turned white. And the 24 field guides hid their notes with display:none, which removes text from find-in-page and from the accessibility tree entirely: 278 entries whose notes are the substance of the page, now carried by hidden="until-found". This corrected the previous day’s audit note, which had named the 401 <details> disclosures unreachable and listed fixing them as the job. Measured first: they are native disclosures, the text is in textContent, three accessibility nodes carry it while closed, and the specification itself says prefer <details> because it gives find-in-page reachability for free. No edit was made to any of the 401. The real trap was that four separate things reveal those notes and all four worked through display:block, which does not override content-visibility — one added declaration per guide fixes all four, and search-index.json came out byte-identical, which is the only proof the reveal still yields the same text.

Queering Earth’s own audit returned eight required-tier findings and closed all eight in a day. The first is a duplicate-address defect that four separately correct things could not see: every sheet answered 200 at both /on-being-ill and /on-being-ill.html, byte-identical, same ETag, because Netlify shadows a redirect with a real file and the file is real. Fixed with a forced 301! per sheet — the ! is load-bearing; an unforced rule never fires — and a fourth gate that asks the deployed site instead of the files. Closing a finding about error pages then introduced the exact fault that finding leads with: 404.html is a real file, so both /404 and /404.html answered 200, and the new gate reported it within a minute of the deploy — the first time a gate there has caught a regression in the same pass that shipped it. Bounded caching went on 2.4 MB of scanned lithographs that had been revalidating on every visit, with no immutable, because the spec reserves that for content-hashed URLs and these carry descriptive filenames a better scan could overwrite, so promising a browser a body can never change would assert a fact we cannot enforce. And the audit corrected itself on the way out: it had recorded “no analytics, no tag manager, no third-party script of any kind” — accurate about scripts, misleading as a privacy claim, because every page loaded two families from Google. Nobody had written it down because there was no privacy policy to write it in.

Yes to Every Crawler, Which Makes Attribution the Whole Ask

Ryan’s decision, 9 September, stated as a decision on both sites: read it, quote it, index it, train on it. Thirteen crawlers named individually and all allowed, each group carrying Content-Signal: search=yes, ai-input=yes, ai-train=yes — named rather than left to the wildcard because an explicit allow is a decision somebody made, and repeated in all thirteen because a crawler obeys only its most specific group and inherits nothing from *. Crawler names came from each vendor’s own documentation rather than the specification’s list, which was stale in two places. The alternative we declined — allow retrieval, refuse training, on the CC BY-SA ShareAlike argument — is written into robots.txt itself so the choice cannot read as a default. robots.txt had always opened with a sentence saying everything here is meant to be read, quoted and reused: a sentence addressed to humans in a file read by machines. Nothing about the policy changed; it stopped being an inference.

The answer being yes makes the credit trail the other half of the decision, so the same pass built it. A .md beside every page, /llms.txt and /llms-full.txt, a feed drawn from the register’s own accessions, JSON-LD on every page, and a digested SKILL.md at /.well-known/agent-skills/ whose main subject is how to cite the site and how to tell a checkable fact from a reading — including that a number records when and never rank, that every count in our prose is a lead and not a fact, and the one sentence widely quoted as our motto that is not ours. On Queering Earth the JSON-LD is authored rather than generated, because author against about.author is an editorial judgment — who wrote our reading, against who made the thing read — and the fifth gate refuses a block that names one person as both. That makes it the first gate there that checks an attribution rather than a fact about a file.

On ★ stuff the same question had a harder answer. Add .md to the address of an essay and you get the same words as Markdown — for 57 of 198 pages, and the 141 exclusions are the more important half. 137 are artifact pages carrying part of their argument in figures: 1,021 inline SVGs holding 4,851 text labels. A zine’s .md would be its prose with the diagrams silently gone — a smoothed retelling of our own work, published in a file whose purpose is to be trusted by a machine that cannot check. Where anything is dropped from the 57, the frontmatter declares it. And the last hour of Thursday found every poem on Queering Earth arriving in its Markdown mirror as prose: the generator listed span as a wrapper, which is right for nearly every span there and wrong for the one that is not, since <span class="l"> is a line and the stylesheet has said since the Rossetti sheet that a poem’s line breaks are the object being mounted. Ten pages of poetry had been reflowing into single paragraphs. Two faults on My Gloomy Sunflowers were fixed in the same spirit: the quotations had gone up as bare blockquotes where every other sheet mounts them on the ruled tier, which put Gadsby’s words and Ryan’s on the same visual surface — and worse, the markup put the citation inside the quotation in the Markdown mirror, where an agent cannot tell our attribution from more of Gadsby’s words, on a site whose crawler policy is that the answer is yes and the ask is attribution.

Two faults also ran the other way, and both are worth saying out loud. Our own working notes claimed every relation in the site-wide Link header was IANA-registered; two of the five were not — the registry holds 236 entries and neither sitemap nor security is among them — and RFC 8288 admits an extension relation only as a full URI, so both were invalid rather than merely unconventional, on every response the site sent. And a young-readers page’s JSON-LD had been invalid since it shipped on 2026-09-06, straight quotes around a quoted question closing the string four words early. Live for four days; nine gates passed it and every one of them was right to. The tenth check parses the block rather than pattern-matching it, was tested against the real breakage, and reports 198 blocks, 198 parsing.

The Names That Came Off

Self-hosting the fonts on Queering Earth introduced both faults in one commit, and fixing them took the rest of the day. The SIL Open Font License requires its text to accompany the fonts wherever they are redistributed — for a day there was none — so each family now ships its own licence file with its own copyright line, because a generic OFL carrying the wrong copyright is a gesture at compliance rather than compliance. The colophon had credited both faces to Google Fonts and named nobody; Fraunces is by Undercase Type, Phaedra Charles and Flavia Zimbardi, and Newsreader is by Production Type. They are on the colophon, in the attributions file and in a generated fonts README, fetched from upstream metadata rather than typed, so the credit cannot drift from the files being served — and the tool refuses to write a licence it cannot recognise or a credit with no designer.

Twenty candidates were then set on two specimen sheets in the site’s own palette, at its own sizes, on its own paper, weighted toward Bye Bye Binary, who draw glyphs for inclusive and non-binary French — where a language with no neutral form makes -le, -es, -e and -te endings what an inclusive suffix looks like when it is drawn rather than argued. Reading their licence changed what shipping them involves. The Bye Bye Binary families are under CUTE, not the OFL — a licence stating outright that it is “not technically compatible with any FLOSS license to date” — and three conditions bind us: credit the designers and link the original publication source, share complete files rather than a lone extracted face, and never delete post-binary characters or the OpenType features that activate them, which forbids the latin-only subsetting our two house faces get. It also asks for a donation on a sliding scale as a condition rather than a suggestion, met at the tier for an association, and all of it is stated above the candidates on both sheets so the terms are visible where the choosing happens instead of in a commit message. Nine faces came out of it, and they cost a reader who picks none exactly nothing: 32 faces are declared, three are fetched.

The material kept producing the same lesson. The sentence that travels as Sandburg on Lincoln is verbatim Sandburg and is about Joshua Speed — page 265 opens Chapter 54 with it, and Sandburg extends the figure to Lincoln two sentences later in different words. Nothing fabricated, no word altered, only the referent moved, and the sheet says so without scoring a point. Three Internet Archive copies agreed and were thrown out as witnesses — same scanning centre, one pipeline, one witness — with the passage re-read in a copy scanned elsewhere before publishing. That rule fired before an error this time rather than after one. The credit for Lilac Wine under Helen’s Eliot is James Shelton’s, 1950, written for the revue Dance Me a Song — it reaches most people through Nina Simone and Jeff Buckley, a cover better known than its original being this site’s characteristic failure in its most ordinary form. The Darwin woodcuts on Sheet No. 3 carry no signature and the library’s metadata names only Charles: the artist is his son George, credited in one sentence of the preface. Read the corner. When the object has no corner to read, read the preface.

And on Friday a collision nastier than a synonym. Ryan: the word “monotropy” is associated with Bowlby’s attachment theory, which we do not endorse. He is right. Monotropism is Dinah Murray’s, coined with Jeanette Buirski and set out in Murray, Lesser and Lawson (2005), and it describes attention; monotropy is John Bowlby’s, an infant’s innate drive to attach to one ranked primary figure. Writing the second for a theory of attention hands a reader mothers, ranked caregivers and innate deficiency — on the one page whose whole argument is that two modes are equal and neither is a grade. The prose had it right and the label had it wrong, in a heading, an id, two comments, a stylesheet note and two of the cabinet’s own records. A published address changed with it, because carrying a repudiated term in a URL beats tidiness. And the 2005 paper’s byline deadnames Wenn Lawson: we cite him as himself and say what the record says.

Saturday’s three pieces are one argument about credit. A Bit of Scruff was proposed as “she was robbed and it was unfair”, and Jocelyn Bell Burnell’s own 1977 speech declines that reading with four reasons and ends “I am not myself upset about it — after all, I am in good company, am I not!” So the zine routes through what she does record as absurd — the press who on discovering a woman was involved “descended even faster” and asked how many boyfriends she had — and hands the child the distinction instead: being kind about something is not the same as the something being fair. The primary cut two other things: the famous quarter-inch-of-scruff-in-400-feet image is not in it, and her name is second of five on Nature 217 (1968), so the circulating version in which she was left off the paper is false — the spread insists on her by-line rather than sharpening a true grievance with a false detail.

A Field Guide to Missing Names then sorts eleven pieces of work not by how famous the person was but by how the name came off, because “she was not given credit” is not one event — the work split and one half counted; credit resolving toward the senior person with nobody deciding anything; the material changing hands while the name stayed behind; a correct document degraded by the retelling; the correction arriving decades late. Two entries in it argue against the guide, which is what earns it the right to be believed on the other nine: Bell is on the paper, and Hedy Lamarr did not invent frequency hopping or Wi-Fi — but her patent is real, US 2,292,387, filed under Hedy Kiesler Markey, which is a missing name, and one that is on the document. Three cards carry no reframe table at all, Henrietta Lacks’s first among them, because consent is not attribution and a credit frame proposes a remedy that addresses none of the injury.

Fourteen Kinds of Doing is the constructive half: a byline is a data structure with exactly one field, an ordered list of names with no room to say what anybody did, pushed to its limit by a 2015 Higgs mass paper with 5,154 authors across 33 pages of which only the first nine are the research. The zine refuses the easy reading — five thousand people really did make that measurement, so the number is the field being honest and the format is what is absurd — then refuses its own happy ending, because a vocabulary makes work visible and not valued, and names the failure in our own house: credit here is well kept for sources and poorly kept for labour. Then Ryan put the one paper we could not reach into the inbox and it argued our case better than we had: CRediT splits PLOS’s single writing category in two and the picture inverts, women 6% likelier to have written the original draft where the coarse scheme had shown men dominating. A coarse vocabulary did not merely fail to record something — it produced a false picture, and a finer one corrected it.

Our First Zine Was Wrong About the Thing It Is Named For

Bone Song was made 17 July 2026, before the Star Stuff voice and method existed, and this week it was rebuilt twice. Writing the read-aloud translation required checking the fact underneath. The credit is Fukada and Yasuda’s, Fukada first author on the 1957 paper; we corrected that on two pages in August and never swept it back to No. 1, where it still stood in four places. The larger fault is the claim itself: nothing on the site had ever mentioned that they measured dried bone, that piezoelectric response falls with hydration, or that streaming potentials in living tissue are a second and possibly larger source — while No. 1 and the masthead broadside both said “every step you take is a small current” with no hedge. What No. 1 says is not false. It is stated more strongly than the evidence about living bone supports, which is a different fault and still open.

Then both primaries were read at full text, and the mechanism was wrong. Fukada and Yasuda say the effect “appears only when the shearing force is applied to the collagen fibres to make them slip past each other” — shear, not compression. Our spread said the voltage goes negative where the tissue is compressed; push straight down the bone’s axis and there is nothing to read at all. Not one abstract-level source or review summary we had used said any of this. Yesterday’s fix went to one page; grepping the fault instead of the page found it on seven — the read-aloud’s grown-up box, its colophon, its JSON-LD, two cards, a broadside’s card, and the Manifesto, which asserted “compression, tension, bending” in one paragraph and correctly described collagen slipping under shear two paragraphs later. Our own edit had falsified four cross-references on top of that: rewriting the poem removed two stanzas, after which the refusals cited an argument the zine no longer makes. Every gate passes a reference like that — it stays grammatical and the link still resolves.

Two corrections from Ryan during that revision are the part worth keeping. The draft said the credit fault had stood “for three years”, in four places — it was seven weeks, and the duration was invented to make the correction sound weightier. The exact failure this site exists to catch, committed inside a correction to our founding fact. And a ruling that reshaped the piece: the spreads carry the story, the colophon and changelog own the mistakes, so six passages were rewritten and the new spread now has zero self-references.

The same overclaim then turned up in its strongest form anywhere on the site. The Bone Note entry in A Field Guide to the Star Stuff You’re Made Of hung piezoelectricity on calcium — summary line, tag, and a whole fact row — a step further than the sentence removed from No. 1 the same day. An element guide is exactly the wrong place to hang a composite-material property on a single element. The replacement is calcium’s own property and is a better fact: more than 99% of your calcium is locked in bone and teeth, and the part that does anything is the under-one-per-cent that moves, because cells spend energy continuously pumping calcium out of themselves, holding it inside at roughly a ten-thousandth of what is immediately outside. A signal made of scarcity, held in place by work.

The refusals broadside then broke its own top refusal, on the back of the same sheet. 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”: a pattern recurring in nature is not why difference deserves welcome, since difference would deserve welcome in a universe that made exactly one shape. Side B then read “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. Because no grep matches “derives an ought from an is”, all ten printable sheets were read, and two more carried it. A rule came out of the day, from Ryan: no hedges and no negations on a broadside. It is read at three metres off a wall, and we had printed the footnote instead of the claim — so the hedge belongs in the sources block, the zine and the ledger, and the accuracy now comes from stronger claims rather than weaker ones. Silence is not an error where a wrong mechanism is.

A Shelf for Four-Year-Olds, Held to the Same Bar

Young Readers is the sixteenth collection and the first defined by who a piece is for. Why a shelf and not a tag: all three founding read-alouds pass the Star Stuff register test outright, and that is the problem — they pass and stay unfindable, because nobody looking for something to read to a five-year-old will locate three read-alouds inside a grid of thirty-nine adult zines. The page opens addressed to the reader it is for, in plain language at read-aloud size, and only explains how the collection works under a heading that says it is for grown-ups. The cost was paid deliberately: a page belongs to one collection, so the three moved out of Star Stuff — 44 cards to 41.

Every read-aloud carries two layers on every spread: large plain type for the child, and a boxed grown-up layer holding species names, figures, sources and the places the evidence runs out, so a caregiver can answer “is that really true?” without leaving the page. Same eight gates, same ledger row, same grading, because “for kids” is exactly the shelf where a verify-everything standard quietly lapses. And in almost every one, the frame it was proposed with did not survive its own sources.

The Colours You Cannot See exists because of a correction to the sentence it arrived with: “bees see more colours than we do” is in nearly every popular account and is a ranking wearing a fact’s clothes — ultraviolet is “merely one component of colour vision”, and humans are more sensitive to red than bees are. The window is shifted, not widened. Almost All of It Is Holes was proposed with a demonstration the paper itself refutes: dropping a same-size paper circle beside a dandelion plume is wrong, because a solid disc supplying the same drag is 38% smaller than the plume, so a same-size circle has more drag and, at roughly twenty times the mass, lands first anyway. Rebuilt on steadiness rather than speed, and the zine states the correction to the reader on a spread instead of quietly benefiting from it. It Stays Green, and Then It Goes was to be called The Yellow Was Already There, the standard unmasking account, and the only source that separated the individual carotenoids refused it — so the uncontested half stays, and the contested half is its own spread in the child’s own layer, ending “nobody has finished this one either.” Its real finding is better anyway: chlorophyll in each marked leaf “remained constant until a phase of rapid degradation commenced” lasting about a week, and the leaves are not synchronised, so a tree turning gradually over weeks is hundreds of fast changes, staggered, and the smooth fade everybody pictures is an artefact of averaging.

One Ear Higher Than the Other is the one where the demonstration is the published experiment at kitchen scale: cover one ear, point at a clap. Your two ears are level, so up-and-down is the direction two level ears cannot do — a barn owl’s are not, and that is the direction it needs most, because prey direction “usually forms a shallow angle with the ground.” Where three papers give three answers to how many times ear asymmetry evolved independently, the zine prints all three: the independence is not in dispute; the counting is. The Cuttlefish That Cannot See Colour holds two measured results that should not both be true — a checkerboard matched for brightness against the animal’s own single pigment, and hyperspectral imaging showing the camouflage works in the eyes of whatever is hunting them — and offers the best thing in the literature: the null hypothesis, that maybe there is no colour perception here and the animal predicts colour from light and dark. The child’s half carries that without the vocabulary: if I showed you a grey photograph of grass, what colour would you make it?

Then Ryan read the draft. “I’m confused between spreads 5 and 6. Humans can see the cuttlefish against the special checkerboard but fish cannot?” No fish was involved in the checkerboard experiment at all — spread five is an artificial board in a tank probing only the cuttlefish’s vision, spread six is natural sea floor and the vision being modelled is the predator’s, and the two spreads described two experiments without ever saying they were two. No gate can see a fault like that. A reader landing on a question neither paper poses is a defect in the zine, not in the reader.

Saturday’s genetics arc is three read-alouds where each exists to keep the last one’s promise. The Same Four Letters is built on the half of Watson & Crick (1953) nobody quotes: the famous line fixes the pairs, and a few lines later the same paper says “the sequence of bases on a single chain does not appear to be restricted in any way.” Every constraint runs across the two strands and none runs along one — the alphabet is shared, the sentence is not. More Than One Way to Spell It keeps its promise that the code is “not in this zine at all”, with every codon checked against NCBI’s translation tables rather than written from memory, because this is the first zine there whose demonstration fails if one row is wrong — and the exception is inside the reader, a body running two tables at once. And The Onion Has More Than You closes it by paying a debt: Pigliucci 2010, cited as “a lead, not read” the day before, read at full text and blunter than we were — blueprint and programme are “not only woefully inadequate but positively misleading.” An onion carries ~16,400 Mb against a human ~3,200 Mb; counting instructions instead does not rescue it; and the strongest form needs no definition of complexity at all, since within genus Allium alone genome size varies about 4.5-fold — so onions disagree with onions by more than onions differ from us. The objection is handed to the child before it is answered, on the ground that an argument presenting only its best evidence is a sales pitch and not a check.

And when the collection’s own convention got named, it was named from finished work rather than invented. This was going to be written up as the newest zine introducing something new; checking the other five first found that four of the six already hand the child something to do, in three distinct kinds — a prop, a thought experiment, a demonstration. That is a form observed in finished work, which is this site’s bar for naming a convention. Two limits came with it, both self-critical, and it is explicitly not a quota: two pieces have none and are not worse for it, because you cannot do a seed’s decade of waiting on a table.

The Blood Was Made of Ants

The Blood Was Made of Ants was commissioned to check a claim, the check came back negative, and the zine is the animal instead. Its argument is that nothing in a Texas horned lizard is its own. Not the weapon: fractionated plasma run past coyotes and mice shows the aversive compound comes from the harvester ants the lizard eats, so losing the ants costs the animal the weapon and not merely the meal. Not the venom resistance — 0.119 µg/g kills a mouse, 28 a spiny lizard, 162 this one. And the tongue lands on the mesosoma in 76% of fifty filmed strikes, “never touched the ants at their rostral or caudal end, even if it was the closest position to strike at” — given a nearer armed target and a further safe one it reaches further, in sixty milliseconds. The venom never became safe. The lizard became specific. Then fire ants arrive and are not a substitute, leaving a perfect apparatus, intact, pointing at nothing — a support withdrawn, not an animal too fragile to cope.

And the claim it was commissioned for does not survive. A video essay says horned lizards are recovering in protected limestone corridors and cutting fire ant numbers through “grazing pressure”; ten papers were read and none contains the phrase, the mechanism, or the recovery. The arrow runs the other way in the literature — suppress fire ants, vertebrates recover. The geography is backwards too. Fire ants really are vanishing in parts of Texas — tawny crazy ants are wiping them out, which is worse news. The true half of the headline is the bad half.

The one source in the zine we never went back to was our own, and it was wrong within the hour. Spread ten read “the thing those children did is now illegal” and quoted a statute. Those children did not handle it — they took turns looking into a tree hollow and left the animal where it was, and the law names three acts: it is illegal to “pick up, touch, or possess” one. Looking is on none of them. The spread was reaching for the sharpest version of a real point and overshot into a claim its own source does not support — and the refusals two spreads later already said “looking is not extraction,” so the zine contradicted itself inside its own closing argument. The founding anecdote was the one source in the piece we never checked, because it came from us. Ryan caught it from his own account of the afternoon.

The corrected version is better. The law draws its line at the next step because that is where the harm was, so those children were on the right side of it — and spread seven gained the load-bearing half: a lizard being looked at and not handled is on rung one of its own defence ladder, and rung one is to hold still. The stillness those children remember was not tameness. It was the defence, working.

Whales, Sexes, Dogs, and a Qualifier Nobody Lost

Nobody Set Out to Do This is built on twenty-eight constant-course segments by sixteen humpbacks, fifteen of them held to better than one degree across up to 2,230 km and 28 days — through currents that deflected their headings by 25 degrees without moving their tracks, and through a tropical storm. Neither the Sun nor the magnetic field, alone, accounts for it. Describing how a humpback turns took twenty-nine distributions. On the same fasting crossing those whales move 3,784 tons of nitrogen a year into water that has none. Companion to No. 30: the vertical whale pump stays Helen’s, and this takes the horizontal one.

Strike “His” and A Field Guide to the Sexes are the week’s largest piece, and the spine of both is an argument we refuse to make. A bluehead wrasse begins behaving as male within hours of the dominant male vanishing and makes sperm inside twenty-one days, by rewriting the methylation marks that stored what it was — and the 1977 paper names the mechanism plainly: sex change by aggressive dominance, which is a ladder, not a freedom. Two is not the number of sexes: Schizophyllum commune runs past twenty-three thousand mating-type combinations. Male pregnancy in seahorses was paid for with the MHC class II pathway and three Toll-like receptors — a price, not a superpower. Against that, Ursula K. Le Guin, who built a world with no fixed sex, found the default she had not noticed she was obeying, and in 1987 reprinted her 1976 essay entire with the corrections showing in brackets rather than smoothed away: “hiding the evidence that one had to go there to get here.” The title is one of those corrections, and she asks that anyone quoting the essay include the reconsiderations — most don’t, and this zine does.

The refusal runs in both directions. Bagemihl’s 450 species were cited in the Lawrence v. Texas amici and it helped, and it is still the wrong argument, because the same sentence licenses traumatic insemination and sexual parasitism — both named on the page rather than curated away. Nature does not authorise us. It only fails to convict us. The guide’s sort is its finding: by taxonomy you get fish, molluscs and a fungus, which tells you nothing, and by strangeness you are building a freak show — so what divides the nineteen is which part of the bundle each one takes apart, because “there are two sexes and you are one of them” is five claims. No organism here breaks all five, which is why the bundle held together so long. Two cards carry no reframe table: one reports an open problem, and the other is about people — both published figures for human sex variation, the note that the gap is almost entirely one condition, and no adjudication, because a sounds-like/actually table turns people into specimens.

Three pieces came out of one research stack about dogs. The Channel You Cannot Mask starts from Ryan’s own account — that like many Autistic people he got on better with dogs than with humans — held on its own spread, marked an account and not data. The settled fact carries the rest: four dogs, breath and sweat from 36 people before and after a stress task, double-blind three-alternative forced choice, 675 of 720 trials and 34 of 36 on first exposure to a stranger. The authors’ conclusion is chemistry rather than sympathy, which puts the dog on a different channel from the one a room grades: masking is face, gaze, prosody and turn-taking latency, and metabolic exhaust is authored by nobody. The honesty is on spreads rather than in footnotes, including 18 dogs that got significantly more cautious under human stress odour, which has its own spread and figure. It closes on an Autistic professional dog trainer whose standard for a dog is higher than the one behaviorism holds for us, with an explicit note that the analogy indicts ABA and never describes us as animals to be handled.

Nobody Knows Where the Dog Came From carries the lowest documented share and the highest contested count of any chain on the site — six documented joints, five contested, two leaps. That is a fact about the subject and not about the sourcing: the question is open and the most recent paper says so. Two rails, bone against genome, because the dispute was never one field disagreeing with itself but two methods returning different answers about the same animal. The paper that closes it turned up while chasing a radiocarbon date and was not in the source drop: “our results cannot resolve where in the world, and how many times, dogs were domesticated.” Without it the chain would have ended at 2020 and been six years stale on its central question. And The Dog Breathes Out Sideways starts where no explanation of a dog’s nose ever starts — with where the air goes. “A ventral-laterally directed air jet is ejected from the nose”: two sideways puffs that stir the scent up instead of blowing it off. The paper’s comparison with us runs the way nobody expects — human sniffing is “undesirable for consistent odorant deposition” — so your nose is not a broken dog nose. It was never trying to be one.

And Too Narrow is the week’s answer to last week’s throughline, from the other side. Schrödinger proposed that a chromosome carries a code-script in 1944 and wrote four lines later that the term is “of course, too narrow”, because chromosomes are “architect’s plan and builder’s craft — in one.” Watson, Crick and Wilkins all credited the book. The limit was on the page the carriers were reading. The chain’s account of why it fell off is the opposite of a conspiracy: one half became testable within a decade and the other did not — base pairing gave a copying mechanism, the code, tables, prizes and a three-billion-dollar project, while how a sequence becomes a body gave almost nothing measurable for fifty years. A qualifier survives only if something can be done with it. Nine documented links, two contested, one leap, and the leap is refused rather than crossed, because we cannot show the metaphor caused the searching and both directions of the arrow fit. Two weaknesses are declared rather than hidden, including link 5 graded contested for our own sourcing failure, with the spread asking to be corrected. And Spectrum 10K — opened 2021, paused within weeks after we objected, consulted on, closed 2025 — is deliberately not a villain story, because that sequence is the thing working.

Money Is What the Filter Eats

The August newsletter went up on Sunday with 92 of 93 links resolving, and the paragraph it is proudest of is the one WordPress broke. WordPress.com’s LaTeX filter reads a pair of dollar signs as inline math, so “revenue of approximately $120,000 … actual 2025 revenue was $32,599” rendered as an image of the words between the two amounts, and both figures disappeared into it. Three passages were affected and every one of them is a money passage. The newsletter’s own claim is that we published the deficit instead of the flattering version — and the deficit is the number that did not render. And it had happened before and nobody noticed: a reflection had carried the same fault since 17 August, where the swallowed span ran 352 characters and took three Markdown links with it, never converted to HTML at all because by the time the converter ran they were already inside a shortcode. A bullet about restating our financials was published with its three citations sealed inside an image of themselves. Both posts are repaired and verified, every restored figure goes back as $, and a sweep of the whole posts table found no third instance. A filter that eats a dollar sign eats the sentence around it, links included, and the only place the damage is visible is the rendered page.

Then on Saturday the Week 36 reflection published with all forty-three links resolving — and twenty-one stray asterisks scattered through the body. The cause is not WordPress. Between two nightly backups the running file gained twenty-eight backslash-escaped emphasis markers in a hundred and forty-five bytes of change: Ryan typed one sentence into the file, and Ulysses rewrote the whole thing on save, escaping ** into *\*. The escapes published as literal asterisks and dragged the surrounding spans onto the wrong words. Riding along with it, the LaTeX fault fired again, in the same paragraph class as last time, so the deficit figure vanished from a sentence whose whole job is publishing the deficit. Money is still the thing this filter eats, and accountability writing is made of money. The repair was rebuilt mechanically rather than retyped, with the converter validated byte-for-byte against twenty-eight undamaged paragraphs before it was trusted with the eight damaged ones. A tool that edits your prose while saving it is a publishing-path dependency, and ours was undeclared.

One more page told the truth badly. Our own Now page had gone six weeks and four changelogs stale, and the tell was not a wrong number — it was a list still describing work that had finished. It went from 24 flat bullets to 35 in 8 groups, and the two items we cut are the more interesting half: dark mode had shipped in Week 30 and sat in a list of work without a finish line ever since, and a landing page had been “in development” for six weeks while the mirror showed it linked from three pages. An Ongoing list that keeps finished work is not out of date, it is inaccurate — it spends the reader’s trust to look busy. Four more items came off because nothing in the garden could confirm them, and Ryan’s call was to cut rather than publish unverified. Each can go back in one line when its status is known.

Week 37 Changelog

Published the August newsletter, the Week 36 changelog and the Week 36 reflection — and found the same WordPress LaTeX fault eating money figures in two of the three, plus a second, separate fault where Ulysses escaped twenty-eight emphasis markers into the running file on save. Both repaired, both verified live, and every dollar figure now published as $ so no dollar-pairing filter can reach it again.
Rebuilt the Now page after six weeks and four changelogs of drift — 24 flat bullets to 35 in 8 groups — cutting two finished items and four unverifiable ones, on Ryan’s call to cut rather than publish unconfirmed.
Launched queering.earth, a Stimpunks × More Realms site reading art, literature, poetry, politics, people and history through a queering lens, live on its own domain in an afternoon and holding twelve sheets by Saturday.
Published Queering Earth sheets on Rossetti, Helen Edgar’s Wyrd, Weird, and the Invention of Normal, The Tempest, Dickinson, the flower codes, the ghost pipe, Carroll’s Snark, Butler’s hyperempathy, four artists and the myth that made them, and attention that will not redirect.
Published Queering Earth’s mission, manifesto, design colophon, privacy policy, quoting position, search, what has arrived and accession register — twelve of the register’s first thirty-four entries are our own errors.
Refused the tagline the site started with: a paraphrase that generalised Nick Walker’s object while keeping her name on it. If we changed the words, they are ours is now a repo rule with an attribution ledger under it.
Helen Edgar declined the byline on her own sheet and was right — the colophon already said a specimen is collected, not written, so sheets have labels, not bylines.
Discovered we were redistributing two typefaces with no OFL text and crediting their designers to Google Fonts. Fixed: per-family licence files, Undercase Type, Phaedra Charles, Flavia Zimbardi and Production Type named, credit fetched from upstream metadata so it cannot drift. Added nine faces to a picker, weighted toward Bye Bye Binary, whose CUTE licence forbids the subsetting our house faces get and asks a sliding-scale donation as a condition — met, and stated above the candidates on both specimen sheets.
Audited both sites against The Website Specification. ★ stuff: no page had a skip link, 0 of 197, since 17 July; 199 scope attributes added; 161 of 208 table names read “the The Biggest Mouth in the Sea”; self-hosted all four typefaces after finding every page was handing an IP address to Google before a word rendered; cut image bytes 86%; shipped a hash-based CSP, .md siblings for 57 of 198 pages, /llms.txt, /.well-known/security.txt, favicons, a web manifest and an RFC 9264 api-catalog. Queering Earth: eight required-tier findings, all eight closed the same day.
Made text reachable that never had been: 4,862 SVG diagram labels now honour Windows Contrast themes, and 278 field guide entries moved from display:none to hidden="until-found" — while the previous day’s claim that 401 <details> were unreachable was measured and withdrawn, with no edit made to any of them.
Ryan’s crawler decision, stated as a decision on both sites: read it, quote it, index it, train on it. Thirteen crawlers named and allowed, Content-Signal on every group, and the declined alternative written into robots.txt itself so the choice cannot read as a default — with the credit trail built in the same pass.
★ stuff published zines numbered 89 through 110, A Field Guide to the Sexes, A Field Guide to Missing Names, The Glimmery, Advice to a Young Marine Biologist, and Glimmer Wire‘s third edition, whose median source lag fell from 36 and 40.5 days to 14.
Opened Young Readers, the first collection defined by who a piece is for — nine read-alouds by Saturday, every one two-layered so a caregiver can answer “is that really true?” without leaving the page, and every one held to the same eight gates as the adult shelf.
Rebuilt our first zine twice. Bone Song credited Yasuda alone where the work is Fukada and Yasuda’s, and said the voltage appears under compression where the 1957 paper says shear — “push straight down the bone’s axis and there is nothing to read at all.” Grepping the fault instead of the page found it on seven pages including the Manifesto, and the calcium entry in the elements guide was replaced with a better fact.
Corrected a correction: a draft said the credit fault had stood “for three years” in four places. It was seven weeks, and the duration was invented to make the correction sound weightier.
The Blood Was Made of Ants was commissioned to check a claim and the claim did not survive — ten papers read, none containing the mechanism, the phrase or the recovery. And the one source we never went back to was our own: those children looked and did not touch, and looking is on none of the statute’s three verbs.
Found 105 of 108 zine covers broken on phones since 17 July — 313 text-on-text collisions at 375px across 103 of 205 pages — because the overlap gate had only ever measured one screen size. Fixed with one shared rule rather than 108, and tested against a page nobody chose as evidence.
Found three rendering gates measuring in a browser they never launched, five days green, driving an orphaned Chrome that had taken the debug port on 7 September. Launcher now refuses a busy port; full sweep re-timed and rescoped from 20 minutes to 4:47, and 24 seconds for a one-page change.
Shipped a control with two green lights that nobody could operate: every pre-flight check called summary.click(), which skips hit testing. A control is not verified until a real pointer has operated it — and the keyboard path was recorded UNVERIFIED rather than passing until Ryan checked it live.
Fixed three pages that had been scrolling sideways since they were mounted, and the fix broke the wordmark in half — which found that text-align had never centred it at all, 34px right of the page’s centre line since the day it was drawn.
Named the monotropy/monotropism collision after Ryan caught it: monotropy is Bowlby’s, and writing it for a theory of attention hands a reader mothers, ranked caregivers and innate deficiency on the one page arguing that two modes are equal. A published address changed with it, because carrying a repudiated term in a URL beats tidiness.
Corrected ourselves in public on two overstated accessibility figures, two papers we said could not be read and could, a public claim that a neighbouring site’s gates were deficient when theirs are better than ours, 229 “invisible” headings that all render, and a count that had been doing two jobs. Their arrangement is arguably better than ours, which is why ours now copies it.

This week two sites were audited against the same specification, and the most useful thing either audit found was not a defect in a page. It was that a great many of our checks had been reporting PASS while measuring nothing: a gate driving a browser it never launched, an overlap sweep that had only ever seen one screen size while 105 of 108 covers were broken on phones, a control certified by code that skipped the one thing that was broken, a typeface gate rendering a single face, a tap-target audit that was wrong twice in opposite directions before it was right. The site never changed across those three runs. Only the instrument did. And the fix is never another check — it is a check that runs the real thing, with a real pointer, at a real width, in a browser you started yourself. The other half of the week was attribution, and it ran the same way. We wrote a crawler policy saying yes to everything and asking only that the work be credited, and in the same days found ourselves redistributing two typefaces with no licence, crediting their designers to a search engine, generalising Nick Walker’s sentence while keeping her name on it, and inventing a duration inside a correction to make the correction sound weightier. A byline is a data structure with exactly one field, and credit here is well kept for sources and poorly kept for labour. Both findings are the same finding: a claim you never tested is not a claim, whether it is about a page or about a person. Broken systems, not broken people.