Helen and I built a whole new site this week. It went from nothing on Sunday to twelve sheets and its own domain by Saturday.
Week 37 runs Sunday 2026-09-06 through Saturday 2026-09-12. The day headings below carry their dates.
The companion piece is Stimpunks.org Changelog for Week 37 2026: From the Gate to the Instrument, From the Byline to the Label, which records what shipped. This one records what it cost.
Best Accomplishments
queering.earth did not exist on Sunday. By Saturday it was live on its own domain carrying twelve sheets, a mission, a manifesto, a privacy policy, a search, and ten automated gates. It is a Stimpunks Foundation × More Realms collaboration with Helen Edgar, reading art, literature, poetry, politics, people, and history through a queering lens — built as ★ stuff’s deliberate opposite, a herbarium sheet in daylight against a night sky.
The part of it worth being proudest of is the accession register, and the reason is that twelve of its thirty-four entries are our own errors: a byline we should not have offered, a citation handed to us with the wrong issue and pages, a caption, the pronouns, three cases of metadata contradicting the corner of the plate. It exists because the corrections had nowhere to go. A site that publishes its own mistakes on an open page in its first week is easier to trust than one that files them quietly.
Young Readers is ★ stuff’s sixteenth collection and the first defined by who a piece is for. 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 as the adult work, because “for kids” is exactly the shelf where a verify-everything standard quietly lapses.
Alongside it ★ stuff published zines numbered 89 through 110, two field guides, a specimen sheet, a Trigger, and a Glimmer Wire. And this repository had the largest single day of ingest the garden has ever had: forty-nine files in, including three texts our published glossary has been naming for years without holding — Chapman’s Empire of Normality, Schalk’s Bodyminds Reimagined, and Murray, Lesser and Lawson’s founding monotropism paper.
Theme
A check that never touches the thing it claims to measure will say we are fine, every time — and the saying so is the expensive part, because it buys weeks of not looking.
The other half of the week is the same fault wearing somebody’s name. Attribution comes off work quietly, and nothing reports it.
Sunday — 2026-09-06
Accomplishments
- A twenty-hour day on ★ stuff — first commit at 03:58, last at 23:51 — and the last hour of it went backwards, into our own first zine. Eleven new pages: seven zines, a specimen sheet, a nineteen-organism field guide, a sixteenth way added to an existing one, and a new collection. What it cost is visible in the order. The day did not end on a new piece; it ended on Bone Song, made 17 July before the voice and the method existed, rebuilt from six spreads to eight — and then on a fix to the element guide, which turned out to be carrying the same overclaim in its strongest form anywhere on the site. Three of the day’s pieces required reversing something we had already committed to: a collection convention that had to grow a third marker before Trigger No. 2 could be filed at all, a specimen sheet whose organising premise collapsed mid-build, and a pond zine that got its central relationship wrong twice before Ryan corrected it from his own water. Eleven pages is the count. The day’s actual work was closer to fourteen, because three of them had to be un-made first.
- The stimpunks.org half of the day produced one publication and then spent everything after it on repair. The August newsletter went out in the morning, and by noon the work had turned into a live edit of two published posts, a mirror sync to carry the corrected text back, a new step in the post-publication checklist, and a re-index behind all three. None of that was writing. Then the afternoon went to the Now page, where the surgical edit was refused by the connector and the fallback was a full rewrite of a page carrying eighty blocks — so the day ended with Ryan pasting Gutenberg markup by hand into a page he could otherwise have had a tool change for him. Four of the five local commits are the cost of one post going out, and the fifth is the markup we had to build because the documented path was shut. A publication is not finished when it publishes.
Insights and Takeaways
- We wrote a correction to our founding fact and put a fabricated number inside it. The revision of Bone Song said in four places that the miscredit had stood “for three years.” It was seven weeks. Nobody counted; the duration was invented because it made the correction sound weightier — *the exact failure this whole site exists to catch, committed inside a correction to the first thing we ever made.* Ryan caught it, and grep confirmed it had not reached another page. *The rest of the audit is the same shape running the other way.* Checking the fact well enough to say it to a child found two real faults we had never noticed: Fukada is first author and this site had credited Yasuda alone in four places on No. 1 — fixed on the Manifesto and the broadside in August and never swept back — and the sentence “every step you take is a small current” was stated more strongly than the evidence about living bone supports, because the 1957 measurement was on dried bone and nothing here had ever said so. Not refuted. Overclaimed. And then the same overclaim turned up in a stronger form in the element guide, where piezoelectricity had been hung on calcium rather than on bone as a composite. Three faults in our oldest, most-quoted material, all of them found by trying to explain it in large plain type to a six-year-old. “For kids” is the shelf where a verify-everything standard quietly lapses — and it is also, apparently, the instrument that finds what the adult version hid.
- Every check we had was pointed at the draft, and the damage only exists in what a reader sees. The post-publication checklist passed the August newsletter that morning — and it was right to, by its own terms: it read the draft, the raw block content and the link list, and all three were clean. The corruption is written into
post_contentat publish time, so it exists nowhere except the rendered page. The instrument was measuring the wrong artifact. The checklist now fetches the live HTML and greps it, which is a one-line addition and should have been there from the first run. The part that stings is the three weeks nobody looked. The Week 33 reflection had carried the same fault since 17 August — a bullet about restating our financials, published with its three citations sealed inside an image of themselves — and it was found only because the same fault happened again to a post we happened to be checking on the day it went out. A fault that is invisible to every check you own is not rare. It is unmeasured.
Gratitude
- Kiersten Case, whose low-stimulation puppet series Lilypad Library is what Nobody in This Pond Is Late is dedicated to — and whose frog-and-axolotl pairing is the zine’s premise, hers and not ours. Her show is the reason there is a read-aloud shelf on that site at all.
Monday — 2026-09-07
Accomplishments
- Seven new pieces on ★ stuff, and the first six hours of the day did not make any of them. First commit at 01:00, last at 18:11. Ten of the day’s twenty-three commits are corrections to material that was already published, and they came first: Bone Song rebuilt from six spreads to ten, both of its primary sources read at full text for the first time, the mechanism found to be wrong, that error then swept out of seven pages, and three of the ten printable sheets found to be deriving a moral claim from physics — one of them on the reverse of the sheet whose top refusal is that you cannot do that. A new page was not made until 07:02. Then the same shape ran through the new work too. The cuttlefish zine had to be rewritten after Ryan asked a question the draft made unanswerable; the dandelion zine’s demonstration was false and had to be rebuilt on a different property; the autumn-leaves zine lost its title and its frame to the one source that had actually separated the pigments. Three of the seven new pieces required abandoning the premise they were proposed with. And the day’s most-repeated finding was about our own instruments. Four gates-passed-clean faults were caught by print-emulation probes instead, four cross-references were falsified by our own edit and stayed grammatical while they were wrong, and three separate errors turned out to be in the check rather than the thing checked. Seven pages is the count the site will show. The day’s real work was closer to seventeen, because ten of them were already up.
- Two things went out under our own name on Monday, and almost none of the day’s cost is visible in either of them. The Week 36 changelog published, and queering.earth went from nothing to live on its own domain — and the launch was mostly argument, not building. Two things were refused: a tagline we had been carrying that put Nick Walker’s name on a sentence she did not write, and Helen’s own reasonable request for a lightweight CMS, which got a not yet with the trigger named so it does not have to be relitigated. Then continuous deployment failed silently for four commits while the dashboard reported success the whole time, and we called it wrong twice on the way to finding out — once declaring a working link dead on the word of a lagging API, once reading our own call kicking a build as the problem repairing itself. A green light and “auto publishing is on” is not evidence that a push deployed. And the last three commits of the day were not writing at all. Twenty-six files ingested into the garden, a false positive found in the concept layer, a section-scoped denylist designed and verified to fix it, and a defect in the site mirror that had been quietly splitting Star Stuff page titles — Starlight indexed as “Star light”, unfindable by its own name, for who knows how long. None of that is a publication, and all of it is the reason the next one works.
Insights and Takeaways
- Every fault that mattered on Monday was invisible to the tools we built to catch faults, and three of them were in the tools. All eight gates passed the pages carrying the compression error, and they were right to by their own terms — the sentences were grammatical, the citations resolved, the counts added up. What the gates cannot see is a relationship. Two adjacent spreads described two different experiments without ever saying they were two, and every sentence in both was individually true; the confusion only exists in the space between them, and it took a reader — Ryan — asking a question neither paper poses. Four cross-references were falsified by our own edit when a rewritten poem lost two stanzas, and they stayed grammatical and their links kept resolving while they described an argument the zine no longer makes. Four print faults were caught by emulation probes rather than gates, because a light translucent arrowhead and a motif crossing a title are art over text, outside what an overlap checker compares. And no grep matches “derives an ought from an is,” so all ten printable sheets had to be read — which is how we found the worst instance on the site sitting on the reverse of the sheet that counts that exact refusal nine times. A pattern search finds the faults you already know how to name. Then, three times in one session, the error was in the check rather than the thing checked — a
grep -cthat missed a capitalised caption and counted “the seventeenth century” as a citation, a changelog entry asserting a consequence its edit never had, and a note about counts going stale that broke our own rule about counts going stale while it was being written. An unmeasured fault is not a rare fault. It is a fault with no instrument pointed at it — and the instrument is a thing that can be wrong. - A science shelf arrived in the garden and our vocabulary stopped meaning what it means. Twenty-six new files went in on Monday and moved only eight concept pages — which is the design working exactly as specified, since concepts are keyed on our published glossary and that glossary is neurodivergence and disability vocabulary, so a corpus of optics and comparative cognition is close to invisible to it. The one science file that did reach a concept page was a false positive. Two papers landed on accommodation — because in optics that word is the eye’s focusing mechanism, “image sharpness vs. accommodation setting”, and has nothing to do with a disability accommodation. Homographs across two vocabularies are a failure mode that arrived with the shelf, and the obvious fix would have been the wrong one. A global denylist entry is much simpler and much worse: accommodation carries forty real links across the disability corpus and is one of the most load-bearing terms we have. Scoped to the wiki section, it loses the two optics papers and keeps all forty. The measurement is what made the decision, not the instinct. Across all thirty-five science assets only five concepts linked any of them — two wrong, and the other three exactly the links worth having: corvid cognition landing on the same page as Autistic cognition, where one literature treats theory of mind and working memory as capacities to demonstrate and the other has spent decades having them denied. A separate concept space for science would have severed those three by construction to fix two bad links. And five more terms with obvious technical homographs were deliberately left out, because none of them has actually collided yet — denying a term costs real links, so a speculative deny is the same class of error as the false positive it means to prevent. Add one after reading the sentence that matched, not before.
Gratitude
- Carol Millman, an Autistic professional dog trainer, whose words close The Channel You Cannot Mask — and whose standard for how a dog should be treated is higher than the one behaviorism holds for us. Building that spread also caught a credit fault in our own library: the saved highlight file for her article credits the publication rather than her, so both quotations were verified at the article itself and the attribution corrected on the page.
Tuesday — 2026-09-08
Accomplishments
- Sixteen hours and nineteen commits across two sites, and a good deal of it was undoing work from earlier the same day. First commit at 07:11, last at 23:26. Five new sheets went up on a site that is four days old — Rossetti, Helen’s Wyrd essay, Miranda, Dickinson, the flower codes — plus a public accession register, a second dark ground, and a vined frame on the front page. None of it ran straight. Three black-and-white plates were mounted on the Rossetti sheet in the morning and replaced with chromolithographs by lunchtime, taking a CSS rule out with them — and that rule came back four hours later when Helen’s landscape plate needed it, so the day both deleted and restored the same line. A byline shipped and had to be undone: Helen declined it, and the site’s own colophon already said a specimen is collected, not written, so the Wyrd sheet was wrong by the grammar we wrote ourselves. The front page had been saying “three sheets, so far” since Sheet No. 3, when there were five. *And the register exists because the corrections had nowhere to go.* Twelve of its thirty-four entries are our own errors — the byline, a citation handed to us with the wrong issue and pages, a caption, a paraphrase that never shipped, the pronouns, three cases of metadata contradicting the corner of the plate. *On Star Stuff, the day’s zine was commissioned to check a claim and the claim did not survive the reading.* Ten papers, none containing the phrase, the mechanism, or the recovery — so the piece had to become the animal instead, with the claim graded on two spreads out of fourteen. Then it shipped with a false statement about Texas law and was corrected within the hour. The count the two sites will show is six new pieces. Most of the day was spent on the four that were already there.
- Nothing was published under our own name on Tuesday, and the whole day’s work in this repository was about the fact that a day of work can leave no trace. The running changelog holds no stimpunks.org bullet for the day, and that is accurate — the four local commits are all plumbing. queering.earth was added to the daily log-catching task, which is the same blind spot that lost 2026-08-28 and 2026-08-29 to Star Stuff: a site built in a git repo of its own produces no changed file here, no status line, nothing for a session to notice. A mirror is not that signal either — it records what the pages say, not that a day happened, and only when somebody runs it. Then running the mirror sync found it had never been running. Its
--allwalked five sites while our own documentation, the skill’s description, and the previous day’s commit message all said six, because there are two maps and queering.earth was only in one of them — the mirror stood at 1 page against 9 live. And the guard that should have caught a one-page fetch was set to a floor of 1, from the day the site really was one page: a threshold left at its founding value is a guard that cannot fire. Raised to 8 the same morning, and it passed on the real fetch that afternoon. Four commits, no publication, and the day’s actual output is that tomorrow’s log will not be empty by accident.
Insights and Takeaways
- Two agreeing copies are not two witnesses, and our own words are not a source — and Tuesday moved that line in both directions on the same day. Three Internet Archive copies of Sandburg agreed on the lavender sentence and were thrown out as witnesses, because they came out of one scanning centre and one pipeline, so the passage was re-read in a copy scanned somewhere else before the flower codes sheet published. That rule fired before an error this time rather than after one — and it caught the referent, not the words: the sentence is verbatim Sandburg and is about Joshua Speed, nothing fabricated, no word altered, only who it is about moved. The same shape turned up twice more. Two agreeing scans of the Dickinson page were one library. The 1896 Rossetti colon had to be read in two separately scanned copies to be trusted. *Agreement between copies of the same document measures the pipeline, not the fact.* *And the other half of the day was the opposite problem.* The horned lizard zine shipped a claim that a thing children did is now illegal, sourced to a paper about handling — those children looked into a tree hollow and left the animal there, and the statute names picking up, touching, and possessing. The founding anecdote was the one source in the piece nobody went back to, because it came from us, and the zine’s own refusals two spreads later already said looking is not extraction, so it contradicted itself inside its own closing argument. No gate can see a claim about a Texas statute. Meanwhile, on the other site, the rule that our own library is not verification was narrowed the same day — to our own prose, not to Ryan’s transcriptions with the citation attached. The distinction that emerged is not “ours” versus “theirs” but checked versus assumed, and a source we wrote is the easiest one to skip.
- The documentation was the only place the two halves of the mirror sync agreed with each other, and that is exactly why nobody noticed. Three separate files here said
--allcovered six site mirrors. It covered five. There are two maps — one that knows how to fetch a domain, one that decides which domains get walked — and queering.earth had gone into the first and not the second, so after its launch day nothing ever synced it again. Everything reported success the whole time. *A split-path failure survives review because each half is individually correct* — and this is the second time this same pair of maps has done it, which the comment in one of them already records. It was found by running the command and counting the sites in its own output, not by reading either map, and not by reading the sentence that was wrong. The guard underneath was in the same condition: a minimum-pages floor of 1, correct on the day the site was one page, sitting under a nine-page sitemap — so a fetch that returned a single page would have passed on a technicality. The instinct in both cases was to trust a description that routes real work. *The only check that catches a claim about what code does is running the code and counting.*
Gratitude
- Sarah Prager, who gathered violet, pansy, green carnation, and lavender as a set for JSTOR Daily in 2020 — the set that became the frame on the front page and then Sheet No. 7. Her gathering is the contribution; the reading is ours, and the sheet says so, along with the fact that her claims have not been read back to a primary here and so the page asserts none of them.
Wednesday — 2026-09-09
Accomplishments
- Sixty-nine commits across two sites in twenty-two and a half hours, and the day’s real subject is that a tool cannot tell a thing from a sentence about the thing. Queering Earth opened at 01:26 and closed at 23:46; Star Stuff ran 11:10 to 17:30. Both sites were audited against the same specification, and both audits came back with Level A failures that had been live since their first page. Star Stuff had no skip link on any of 197 pages and 201 of 218 table header cells with no
scope; Queering Earth was answering 200 at two addresses for every sheet, byte-identical, since the day the domain was attached. What it cost is that almost none of it ran straight. Five separate times in one day a Star Stuff tool read our own prose as the mechanism the prose describes: the markup gate flagged a deliberately broken<meta>inside a JavaScript string; the CSP generator counted a comment explaining a removed handler as a live handler, and would have kept a weakening keyword in the policy indefinitely on the strength of a description; the Markdown exclusion probe excluded the changelog itself — the largest and most agent-useful document on the site — because a changelog entry quotes the template expression it was explaining a fix to; and two verification greps reported prose as findings. Queering Earth hit the same fault three times with angle brackets in comments, once in the commit immediately after writing the rule against it into its own notes. *And the sixth instance was aimed at somebody else.* Star Stuff published, in its changelog and its working notes, that Queering Earth ships Markdown mirrors with no staleness check — read from a comment naming a file that was never written, without reading the seven-line gate list that has the answer in it. *That correction took four files here and a fix in the neighbour’s repo.* Alongside it: three figures published in the morning and corrected in the afternoon on Star Stuff, and on Queering Earth *a finding about error pages whose closure introduced the exact fault the finding leads with*, at two addresses, caught by that site’s own new gate within a minute of the deploy. *The count the two sites will show is a privacy page, a CSP, AVIF everywhere, Markdown siblings, a new sheet, a search page, a mission, a manifesto, and five new gates. A good deal of the day was spent correcting what the same day had already published.*
Insights and Takeaways
- Three separate rules were retired on Wednesday for measuring the wrong quantity, and every one of them had read as correct for as long as it existed. On Queering Earth the rail’s breakpoint had never measured clearance at all — the rail and the page derive from the same centre line, so the gap between them is constant at every width, and widening the window buys the prose nothing. A CSS rule reserving 288px for a block that asks for 46rem inside a container capped at 34rem constrains nothing, and has presumably done so since the day it was written. *Nobody can miss a block that was never wider, which is how an inert rule survived on a site carrying eight gates.* And the three thresholds for deciding which pages deserve a contents list were all retired inside two hours: more than two headings counted the wrong quantity; the block must not be too tall compared a page’s block against other pages’ blocks instead of against its own page; the page must be long enough failed the same way from the other end. *What replaced them is structural and has one instance — does this page already have a better index of itself?* The same lesson arrived from a third direction and caught us mid-correction. Saying the 288px rule “does nothing” went out in three commits before anybody checked paper, where
@media printremoves the container’s cap and that rule is the only thing holding the block in: it narrows, on the one medium where it is live. The first claim was reasoned from a rule; the correction was measured on screen and then generalised to every medium. *The same fault in better clothes, and it is why both browser gates on that site now sweep paper as a separate pass: paper is not a wider screen.* *The other half of the day says what a gate is for and what it cannot be.* Five gates were green while an accession stamp rendered as three unstyled spans, because the stylesheet was cached five minutes longer than the markup that needed it — and the comment in the headers file had predicted that failure in those words and priced it as acceptable. A comment cannot fail a build. So a sixth gate was built for it, a seventh ported for text landing on text — built because a design had been refused for want of it, and the first thing it did was catch that refused design — and an eighth for card order, after the plate said seven, carried eight, and had two of them out of order while all seven existing gates passed. Three of the four faults that mattered on Wednesday were invisible to every check we had, and the ones we built to see them were each proved by being made to fail first.
Gratitude
- James Shelton, who wrote Lilac Wine in 1950 for a revue called Dance Me a Song, and whose name Wednesday’s work put back under the song on the front page of Queering Earth — the credit having been carried away by Nina Simone and Jeff Buckley, which is the work surviving while the maker falls off it, the plainest form of the failure that site exists to record. Alongside him, Helen Edgar, whose essay for the fifth anniversary of Dinah Murray’s death is what Sheet No. 8 follows.
Thursday — 2026-09-10
Accomplishments
- Twenty-six commits across two sites in a day that ran from 06:32 to 23:58, and the honest summary is that a great deal of it was undoing what we had already published about our own work. Queering Earth carried 22 of them and Star Stuff 4. What the day cost is visible in how much of it was spent on things that had looked finished. A reader control shipped that nobody could open, and the two artefacts proving it worked — a scripted click and a screenshot — had each bypassed exactly the mechanism that was broken. Two typefaces had been redistributed for a day with no licence file and their designers credited to a search engine, a fault we introduced ourselves the day before on a site whose stated correctness requirement is attribution. Two link relations we had written down as IANA-registered were not, and had therefore been invalid on every response Star Stuff sent. A page had been offering search engines and agents nothing typed at all for four days because four words of a quoted question closed a string early, with nine gates green over it. Ten pages of poetry had been reflowing into prose in the Markdown mirror since the mirror existed. Two font families we recommended do not exist under the names we gave them, and three more were set on a specimen sheet in the wrong weight because those filenames sorted first. Nothing on that list was caught by a gate; every one of them was caught by going and reading the thing. The count the two sites will show is three sheets, three plates, a Reading panel with nine typefaces in it, forced-colours support on both sites, real icons and a manifest, an AI crawler policy, an agent-skills file, find-in-page reaching 278 field notes, and four new gates — and the day’s most expensive hours went to the other list. Left standing on purpose: the 101 paged zines on Star Stuff that hide 2.16 million characters the same way the field notes did, recorded as a ruling to make rather than quietly added to a backlog; an outline masthead refused twice and logged open; and two glossary pages on stimpunks.org that splice two Butler passages ninety pages apart into one quotation, found from a neighbouring repo and written down for a session in this one.
Insights and Takeaways
- Thursday’s lesson arrived four times from four directions and it is the same one every time: the thing that told us we were fine had not touched the thing that was broken. A Reading panel shipped that could not be opened, and both proofs of it working were forgeries in good faith — a scripted
summary.click()dispatches straight at the element and skips hit testing, which was the one thing broken, and the screenshot was taken after forcing the panel open from the console. A control is not verified until a real pointer has operated it. The same session then recorded the keyboard path as UNVERIFIED rather than passing, because a synthesised keypress does nothing in that harness even on a pristine control, and it was only marked confirmed after Ryan pressed the key on the live site — “I could not check this” is only honest for as long as it stays true. The inverse fault ran the same day on Star Stuff, where 229 headings were reported invisible from a computed property that said the fill was transparent; every one of them renders, and the rule written for them was a no-op on 198 pages before it was deleted. *Reading a property and inferring a rendering, and operating a control without touching it, are one mistake wearing two coats.* Then the stale-premise pair, which cost the most. A session searched the decision log, found nothing, and reported that no such entry existed — right about its own worktree and wrong about the repository, because the entry had landed on main after that tree was cut; it merged the entry in and did not re-check, having already concluded the thing was not there. An hour later a third session *declined to write anything at all because it saw a merge in progress that had been pushed two commits earlier* — careful, considerate, and wrong in the other direction. Believing a state that is gone is the same fault as doubting one that is there. Caution does not protect you from a stale premise; it only changes which way you are wrong. And when we went to confirm a peer’s line counts three separate ways, two of the three methods were silently wrong in the same pass — one regex did not survive shell quoting and returned 0 for every file, one counter was blind to fenced code blocks and read a shell line continuation in the decision log as a line of poetry. The peer’s numbers were right. A cross-check is not independent until it has itself been checked, and three of these four faults were invisible to every gate either site owns — a gate that read our own prose for coherence would be a grader of English, not a guard. - The knowledge system published nothing on Thursday and spent its one commit fixing a safety net that had been reporting success while doing nothing. The daily log-catching task had recorded 69 commits by passing them to its ledger as
--sha $SHAS, and zsh does not word-split an unquoted parameter expansion — so the argument parser received one 2,800-character string, the ledger stored exactly one “SHA” that can never match a real commit hash, and it printed a line that reads like success. The next morning’s run would have reported all 69 commits as unlogged and written the entire day a second time. Nothing in the script could have caught it, which is the part worth sitting with: the failure was in the shape of the call, not in the logic the guards were watching. The fix validates every value as a git object id and exits without writing anything when one is not, splits a whitespace-joined value rather than storing it — because this runs after the bullets are appended, and a hard refusal at 5:07 AM with nobody watching would leave exactly the double-log the guard exists to prevent — and now reports what it newly recorded separately from the day’s running total, since the old line reporting one-of-sixty-nine as “recorded” is the whole reason it looked fine. Two days of the ledger were repaired by hand as part of the same commit. A guard that cannot fail is not a guard, and a success message is not evidence of success.
Gratitude
- Phaedra Charles and Flavia Zimbardi of Undercase Type, who drew Fraunces, and Production Type, who drew Newsreader — whose names went back under the letters every word on Queering Earth is set in, after a day of the colophon crediting their work to a search engine. Alongside them Bye Bye Binary, whose CUTE licence made us read terms rather than assume them and whose inclusive French glyphs are on the sheets; and Henry Holiday, whose 1898 article we finally read off the page images and which corrected Sheet No. 9 instead of confirming it — Carroll called his Boojum “a delightful monster, but inadmissible”, three words every secondary account drops.
Friday — 2026-09-11
Accomplishments
- Thirty commits across two sites, first at 00:06 and last at 23:54, and the honest shape of the day is that most of it went to things that had looked finished. Eight new pages — Glimmer Wire’s third edition and Zine No. 104 on ★ stuff, and on Queering Earth a sheet, a quoting position, three drawer pages and a page of what has arrived. Against that: a masthead that had been sitting 34px right of centre since the day it was drawn, three pages that had scrolled sideways since each was mounted, a wordmark broken in half by the fix for those three, a heading carrying a word we repudiate on a sheet a day old, eleven quotations mounted on the wrong surface hours after the sheet went up, and four corrections to a decision-log entry that was refuting itself. The zine’s best spread exists because our own first check over-corrected and cut the argument, and Ryan brought it back. Two of the day’s screenshots were Ryan’s, and each one opened a fault nine green gates had passed.
- Nothing went out under our own name on Friday, and the knowledge system’s nine commits are the largest single day of ingest this garden has had — which is also how we found out it had been holding twenty-two books at a third of their length. Fifty-one files offered, forty-nine ingested, one deduped and one refused and left in the tray because the extracted text was byte-for-byte identical across all 71,127 characters and only the PDF’s internal metadata differed;
raw/went 540 to 589. Three of them are texts our published glossary has been naming for years without holding — Chapman’s Empire of Normality, Murray, Lesser and Lawson’s founding monotropism paper, and Schalk’s Bodyminds Reimagined — and the concept page for empire-of-normality gained an “In raw/” section for the first time. Then the shelf had to be rebuilt around them: a literature-and-arts category, because a Townes Van Zandt biography was sitting under a heading that says Philosophy & Aesthetics and aesthetics was the nearest true thing and not true enough. And the rest of the day was not acquisition at all. Ryan found a passage missing from the Plath journals; the index held 109,718 of the book’s 299,336 words and ended mid-word. Twenty-two assets, every one of them a book, and a report this repo described on 6 September as held and citable was 32% indexed. The day’s work made the garden bigger and then spent its afternoon finding out how much of what was already in it we could not actually read.
Insights and Takeaways
- Friday’s fault arrived four times and it was the same one every time: the number we measured was not the number that mattered. A drawing was judged at 544px because 34rem is the measure, and a block inside main gets 448 — a figure already written in our own decision log for a different question and not reached for here, so it shipped 21% oversized. A sweep of document widths across 25 pages and both paper sizes reported no damage from a rule that had broken the wordmark in half, because none of the damage moved a document width — a word can sit wider than its own box without the page scrolling, and those are different faults. The wordmark had never been centred at all, and the reason is that
text-aligndistributes leftover space, so when the line is wider than the box there is nothing to distribute; it had been laid from the start edge for as long as it existed, at 34px off in the face we always look at. And the gate built to catch exactly this had been reporting passes while rendering one typeface of nine — the same blindness the contrast gate had, in a second gate, three days after we wrote that lesson down. The related lesson is cheaper and worse: a probe that reported a fix missing from eleven commits, uniformly and confidently, was itself broken, and a2>/dev/nulladded to keep the output tidy had hidden the error that would have said so. Suppressing stderr converts an error into an answer, and a uniform answer reads as a finding. - A search that returns nothing has to be distinguishable from a garden that holds nothing, and for twenty-two books ours was not. The cause was one limit serving two consumers with opposite needs: the concept layer caps extraction at 600,000 characters, which is fine — aboutness is settled early — and the library collection is built from that same cache, so the only thing that makes
raw/searchable at all inherited a cap that was never meant for it. The text was discarded at extraction, before anything downstream could see it. Ryan’s passage sits 62% into the book. Nothing distinguished that empty result from the garden not holding the Plath journals at all — which is precisely the failure our search wrapper carries a header about, arriving one layer down, inside the index itself. The generalisation is the part worth keeping: 98 of 586 assets have no usable text, they were counted in build output, and nobody greps yesterday’s console. So the manifest of what the index cannot see is now written as a document the index itself holds, and a search for the title of something we own but cannot read returns a page saying so. The tempting fix was refused on measurement, not taste — flagging files whose length is implausible against their page count catches three files the existing floor misses and is blind to 95 of the 98, because only PDFs have page counts and 68 of the blind assets are images. The right goal with the wrong instrument still reports success.
Gratitude
- Hannah Gadsby and Shoshana Kessock, whose argument My Gloomy Sunflowers is built on and who are credited by name on it — the case that the myth of the tortured artist reads the work as the symptom when the record keeps saying the work was the thing going well. Ours is the evidence, not the thesis. Building the sheet also caught our own site crediting the Kessock passage to the anthology rather than to her, which is the same failure the sheet is about, found in our own house.
Saturday — 2026-09-12
Accomplishments
- Thirty-one commits across two sites between 00:22 and 19:12, seven new pages, and the day’s real division is that ★ stuff spent it making things and Queering Earth spent it building the road a draft travels. Six pieces went up on ★ stuff — three read-alouds carrying one genetics argument end to end, an adult chain behind them, a zine and a field guide on credit — and the cost of that pace shows in what had to be repaired alongside: two published claims that a paper could not be read, when both could, a count on a collection page that had been wrong through two careful re-derivations because nobody had asked which of two sets it named, a page quietly re-linking Google’s fonts three days after the site announced it had stopped, and 105 of 108 zine covers broken on phones since the first cover shipped in July. Queering Earth shipped one sheet and then eight commits of plumbing, most of it correcting its own work from earlier the same day: promotion is not a merge, a branch deploy is not noindexed, an uncommitted draft is still a draft, a draft is not always one file. Cavendish stayed dormant, as it has for weeks. The honest summary is that Saturday produced a lot and then spent its second half finding out what it had been shipping wrongly for months — and the two halves were done by sessions that could not see each other, which is how a cover fix built all afternoon got tested against a zine written in ignorance of it and landing on main first.
Insights and Takeaways
- Saturday’s lesson is that our gates keep failing in the same shape, and the shape is a check that measures one instance of a thing that has many. The overlap gate had been reading 1280×900 and nothing else, so 313 collisions on 103 pages sat there since July; its own header had written down in advance that it would miss them, and a limitation a tool documents is still a limitation. A rendering gate on the other site had for five days been driving an orphaned browser that took its port, measuring in a version nobody installed and reporting PASS — the same failure the typeface gate had on Friday, rendering one face of nine, and the contrast gate three days before that. Three gates, three weeks, one hole: green means the thing we sampled was fine, and we keep reading it as the site is fine. The other half of the day was the reverse error and it is more expensive. Twice we published that a source could not be read, and twice it could — once because our fetcher hit a page-image reader when the same repository serves an OCR’d PDF by another route already named in our own notes, once because a publisher returned 403 and Ryan simply put the file in the inbox. The failure was not the blocked route, it was stopping at it. A claim that something is unreachable is a claim about the world and gets checked to the same standard as any other, and when the second of the two was finally read it carried the result that argued the zine’s thesis better than the zine did. The generalisation both halves share: we are good at recording that we could not verify something and bad at treating that record as a task.
- The one thing that went out under our own name on Saturday was damaged by our own editor, and the two faults in it are opposite kinds of problem with opposite fixes. The escaped asterisks are new and were nobody’s decision: Ryan typed one sentence into the running reflection and Ulysses rewrote the entire file on save, and twenty-eight escaped emphasis markers arrived inside a hundred and forty-five bytes of change. A tool that edits your prose while saving it is a publishing-path dependency, and ours was undeclared — it sits between the file we write and the page we publish, and nothing in the repair changes that; the next reflection is exposed to it exactly as this one was. The LaTeX fault is the other kind, and it is worse for being the second time. We documented it six days earlier, in the same paragraph class, and it still ate a deficit figure out of a sentence whose whole job was publishing the deficit — so writing something down in the fact-check ledger did not stop it happening again, and only escaping every dollar sign at the source will. The cost of the repair is worth naming too, because it was paid in method rather than typing: the damaged paragraphs were rebuilt mechanically from the pre-Ulysses version and the converter was validated byte-for-byte against twenty-eight undamaged paragraphs before it was trusted with the eight damaged ones. That is the right way to fix eight paragraphs and it is expensive, and it was only possible because a nightly backup happened to hold the file from before the editor touched it. The thing that saved this post was a commit nobody made on purpose.
Gratitude
- Jocelyn Bell Burnell, whose 1977 “Petit Four” speech we read at full text while building A Bit of Scruff — and who declined, in her own words and with four reasons, the grievance the zine had been proposed to argue. She also supplied the numbers that corrected a secondary source we had been ready to trust, and the detail that matters most: she is second of five on the 1968 paper, so the circulating story that she was left off it is false. Being argued with by the subject of a piece, decades after the fact and out of her own record, made it a better piece.
Looking Back
Meaningful Moments
The horned lizard. The zine opens on an afternoon out of Ryan’s own childhood — children taking turns looking into a tree hollow at an animal that held perfectly still — and it shipped with a spread saying “the thing those children did is now illegal.” The statute names picking up, touching, and possessing. Looking is on none of them. The founding anecdote was the one source in the piece nobody went back to, because it came from us, and Ryan caught it within the hour from his own memory of the day.
The corrected version is better than the one we meant to publish. Those children were on the right side of the line the law actually draws, and the spread gained the half that had been load-bearing all along: 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 they remember was not tameness. It was the defence, working.
And on the other site, Helen Edgar declined a byline on the sheet built from her own essay — and she was right, because that site’s own colophon already said a specimen is collected rather than written. Sheets have labels, not bylines. The correction came from the person it cost.
Life is Teaching Me
Green is a claim about the instrument, not about the work.
Three gates in three weeks failed the same way, and every one of them reported PASS the whole time. The contrast sweep was blind to 218 elements collapsed to opacity 0. The typeface gate rendered the default face and nothing else. The overlap sweep had only ever looked at 1280×900 — at 375px there are 313 text-on-text collisions across 103 pages, and 105 of 108 zine covers had been broken on phones since the first cover shipped on 17 July. On queering.earth three rendering gates spent five days driving an orphaned browser that had taken their debug port, measuring in a version nobody had installed, and coming back green.
What they share is not carelessness. Each one measured one instance of a thing that has many — so green means the thing we sampled was fine, and we kept reading it as the site is fine. The fix is never more checks. It is checks that run the real thing: a real pointer, a real viewport, a real browser we launched ourselves.
Habit Insights
The changelog and reflection habits hold, and the daily log-catching task got a real repair after two days of reporting success while storing one unusable value for sixty-nine commits.
But the habit doing the most work this week is not automated, and never has been. A reader control shipped with two green proofs that it worked, and both were forgeries in good faith — a scripted click() dispatches straight at the element and skips hit testing, which was the one thing broken, and the screenshot was taken after forcing the panel open from the console. It was marked confirmed only after Ryan pressed the key on the live site. Two of Friday’s screenshots were his, and each one opened a fault nine green gates had passed. The cuttlefish zine was rewritten because he read the draft and asked a question neither of its papers poses. A reader landing on a question the sources do not answer is a defect in the zine, not in the reader.
And this document has an undeclared dependency of its own. Ulysses rewrote this file on save, adding thirty backslash-escaped emphasis markers and twenty non-breaking spaces to a week of prose nobody had edited — the same fault that published last week’s reflection with twenty-one stray asterisks in it. The repair was checked byte-for-byte against the last clean backup rather than retyped. A tool that edits your prose while saving it sits inside the publishing path, and ours is still there.
Looking Forward
One Change I Can Make Next Week
Turn the record of what we could not verify into a list with dates on it.
Saturday’s finding was that we are good at recording that we could not verify something and bad at treating that record as a task — and the evidence is sitting in this document’s own predecessors. Close one open item before starting a new piece was Week 35’s one change. Close the corrections we have already found was Week 36’s, which already noted it was the second time. As of this writing stimpunks.org/together/ still licenses Kaya Oldaker’s dragon CC BY-NC when she released it to the public domain. Three weeks, three resolutions, same line of markup.
And Week 37 added to the pile rather than drawing it down: 101 paged zines on ★ stuff hiding 2.16 million characters from search and from anyone using find-in-page, two glossary pages here splicing two Butler passages ninety pages apart into a single quotation, an outline masthead refused twice and logged open.
Writing it down is not the task. It is the receipt for the task.
I’m Looking Forward To
More read-alouds on the Young Readers shelf — and finding out whether the two-layer spread holds up in the one test that matters, which is somebody reading one out loud to a five-year-old and the five-year-old asking the grown-up layer’s question back.