Week 38 runs Sunday 2026-09-13 through Saturday 2026-09-19. The day headings below carry their dates.
Best Accomplishments
Helen Edgar stopped being a contributor to this work and became a peer in it. On Sunday the knowledge system learned to hold a second checkout — the search config templated, both commit paths taught to fetch, a floor put under the library, an ingest handoff built for a library that lives complete on exactly one machine — and the same day Held, Not Fixed became the first sheet on Queering.Earth written by her. By Tuesday her Map of Monotropic Experiences had a home of its own, and by Wednesday she had read forty passages of ours written onto her artwork and sent back notes on eighteen. The machinery is the part nobody will see, and it is what lets everything else reach two people instead of one.
The Pattern Library went from 18 published patterns to all 56, and A Pattern Language — the book our whole form comes from — became readable and credited after four months in the library as an image-only scan that no search could see. Recovering it turned up nine more books in the same state, including Helen’s illustrated ebook on monotropism.
A special ed teacher in Portland asked how to start a Neurodivergent student group, and got a page rather than an email. The answer had been in the garden all along, spread across five corpora and a thirteen-file toolkit, with nowhere a person could be sent. The page argues for a multiage group in a section of its own, and it refuses to publish a list of Oregon organisations we could not vouch for, publishing a filter instead.
Alongside them, five domains came into existence or changed hands — penguinpebbling.app, monotropicmap.org, cavendish.space, whysheet.press, and stimpunks.love, now Stimpunks.World — and Our Sibling Sites made stimpunks.org link to them at last. The press opened on one reconciled count of how many Why Sheets exist, where there had been three, and AAC Is Not a Last Resort published as the seventeenth.
Theme
The week a second person arrived, everything that had only ever been checked by one of us turned out to have been checked by nobody. A search config correct on exactly one machine. Nine publishing hazards kept in one laptop’s memory. A scroll threshold that shut out most of the windows the feature was built for, invisible in the tall one it was tested in. Forty passages on Helen’s artwork that she had not yet read. None of them errored. All of them passed.
The other half is the fault no second checkout finds. What a number cannot see sits inside our own stated principle until somebody reads it out loud — a deadname printed four times beside the sentence that forbids it, an age policy read as open because nothing prohibited anything.
Sunday — 2026-09-13
Accomplishments
- Thirty-seven commits across two sites in twenty-two hours, and ★ stuff did not publish a single new page. It opened at 01:57 and closed at 23:56, and every hour of it went into material that was already up: a figure we had published on Saturday re-measured and found wrong by seventy pages, 1,054 dead selectors taken out of 119 pages, 107 copies of one function collapsed into one, 1,015 duplicated rules pulled into a shared sheet, a stylesheet’s last unused selector removed, 353 more audited and deliberately left alone, fifteen headings given addresses they had never had, and two pages of our own caught telling readers different things about how this site is built. The day’s output is a site that says the same thing about itself in every place it speaks, and nothing new to read. Queering.Earth mounted one sheet and spent the rest of the day on the machinery. Held, Not Fixed is Sheet No. 13 and the first written by Helen Edgar — and eleven of the day’s twenty-four commits there are not about it at all, they are about what her arrival broke or exposed: a shared draft branch that produced three faults before noon, a publishing checklist that would have handed one person the other person’s files, a scope command introduced at 15:00 and corrected twice before 16:10, a merge instruction in three skills that was confidently wrong, and two host behaviours that each cost a round trip and neither of which is a fault in the repository. The zine setting Helen’s piece arrived as was built, measured, and thrown away, on her and Ryan’s joint call, and the findings kept so nobody spends that time again. Two sites, one new page between them, and both of them are steadier than they were at breakfast.
- One post went up on stimpunks.org on Sunday morning, and everything after it was either repair or the cost of admitting a second person to the work. The Campfire post published before the session it announces, and two of its faults could only be found after publication — a heading slug the platform generates and nobody can predict from the source, and a link checker reading the rendered page instead of the raw blocks, which reported 33 links where the post has 69. A publication is not finished when it publishes, and this is the second month running that the checklist learned that from the same direction. It also left a queue rather than a fix. One cited domain no longer resolves at all — not moved, gone — and it is carried by nineteen files in our mirror, including our own Play glossary entry. This post’s citation was repaired in place; of five sampled replacements two answer and three do not, so the other eighteen are a page-by-page repair and not a search-and-replace, and they are still sitting there. The rest of the day is seven local commits with nothing to show a reader: the search config untracked and templated, both commit paths taught to fetch, a floor put under the index so a partial library cannot quietly regenerate a smaller one, an ingest handoff built because the library is 2.4 GB and lives complete on exactly one machine, the session sync moved off the weekly pass and caught up ninety-eight sessions, twelve decisions written down with their reasoning, and the wiki counters corrected from 232 to 296 after a quarter of the record had gone unstated for a day. None of it is writing, and all of it is what makes tomorrow’s writing reach two people instead of one.
Insights and Takeaways
- Sunday was a day of things that sounded right, and the two sites failed that test in two different ways — one of which a machine can settle and one of which nothing can. On ★ stuff the same reasoning went wrong three times in a row while moving duplicated rules into a shared stylesheet. A page that declares its own rule overrides the shared one — false, because the cascade merges per property, not per rule, so a page that set colour and not letter-spacing kept its colour and silently inherited a letter-spacing it had never had, across 2,878 elements on 53 pages. Then: a rule that passes that test is safe — false, because one of them outranks a scroll zine’s own rule in a state no reader ever reaches and four of our own checkers deliberately enter in order to measure. Then: the rules inside media blocks are the same problem at a smaller scale — false again, and in the opposite direction, because a media query adds no specificity at all, so an override that works only because it comes later in the same file is silently lost the moment it moves to a file that loads first. Every one of those was settled by rendering and counting — 7,705 elements, then 2,878, then 140,667, then a two-file fixture showing 16px against 48px. Three wrong theories, zero arguments, because the answer was a number the whole time. And then the other kind, on Queering.Earth, where there is no number. Our attributions file says in as many words that reproducing a publisher’s record over a living author’s own name is a database outranking a person — and the entries around that sentence printed a deadname four times, on two published sheets and twice in the ledger. The justification written beside it was findability, and it took one question to dissolve: the DOI is what makes a paper findable, so the old name was never doing that work and the cost was being paid for nothing. Nobody had asked, because the sentence sounded like a reason. Nothing gates this and nothing can — a deadname is not a shape a checker recognises; it is a fact about a person that only a person knows. The lesson is not that measurement wins. It is that the faults a number can catch get caught the same day, and the ones it cannot sit inside our own stated principle for as long as nobody reads it out loud.
- Nothing in the knowledge system was broken on Sunday morning, and four things in it were, and the only instrument that could tell the difference was imagining somebody else’s machine. Every one of those faults was invisible from the only checkout that existed — a search config full of one home directory’s paths is correct on that machine forever; a commit path that never fetches cannot fall behind when it is the only writer; a floor under the library count is unnecessary when the library is always complete; a handoff for ingest is unnecessary when there is one pair of hands. They were not latent bugs waiting to be found. They were not bugs at all until Sunday, and then all four were, at once. We did not find them by auditing. We found them by writing down what would happen when Helen cloned the repo, one step at a time, and noticing that every single failure was silent — a search tool reporting no results rather than no index, a push rejected while the nightly backup goes on reporting success, a partial index regenerating under a plausible commit message because our deletion guard counts deleted files and a gutted index is a set of modified ones. The uncomfortable half is the record we had already been keeping and never reading. Two hundred and thirty-three session summaries, written faithfully every week, and not one of them had ever been cited by anything — which is exactly what you would expect of a record that says what was asked and which files moved and never what was concluded. The summary of the session that chose a share-alike licence over public domain opened with “I’ll look for a log file in the repo.” and never mentioned the licence at all: the preamble kept, the reasoning thrown away. And the newest site had zero of its thirty-two sessions indexed, an entire project absent, because every session on it postdated the last weekly run — with nothing anywhere saying so. A record nobody cites is not a record, and it fails exactly the way a silent error does: by looking complete.
Gratitude
- Helen Edgar, whose zine became Held, Not Fixed — the first sheet on that site written by her, and the one that turned out to be about the cabinet it was mounted in. Her note after publication is what fixed it: Ole Worm comes out of nowhere. He did, and three sentences now say who he was. Alongside him, Ole Worm himself, who kept a unicorn horn in his room after proving it was a whale’s tooth, and did not throw it out — the wonder surviving the correction, which is the whole argument of the sheet.
Monday — 2026-09-14
Accomplishments
- Three sites, thirty-three commits, and one new page between them that a reader can go and read. Penguin Pebbling went from a Custom HTML block inside somebody else’s blog post to a six-page site with its own domain, typeface, palette, dark mode, changelog, privacy notice and published deck text — in two hours and forty-six minutes of commits, starting at 20:53 — and by the end of that stretch its own instruction file was wrong in every load-bearing claim it made, so the last commits of the night went on correcting the documentation the same evening wrote. ★ stuff published nothing new at all: its day went into fifty-four headings that had existed only to the eye, contents lists on eight racks, a margin rail that took four drafts to stop landing on the words, seventeen collection pages whose structured data disagreed with their own card grids, and a 1,402 KB changelog broken into months. Queering.Earth mounted one sheet at 03:25 and spent the rest of the morning on what publishing it exposed. The honest measure of the day is not the sheet and not the new site. It is that most of what got fixed was in the files that describe how faults are kept out — three copies of one wrong sentence about draft deploys, a gate reporting PASS over a page it had never read, a privacy notice that had been false by two for two sheets, a README contradicting itself two screens apart, and an instruction file overtaken by eleven commits in an evening. Nothing errored. Everything passed. The day’s real work was noticing.
- Launching Penguin Pebbling cost more inside the knowledge system than it did outside it, and none of that cost is visible on the site. A new site is not one change here, it is five: the daily log skill had to learn the repository or a day of work on it leaves no trace in this repository at all; the mirror had to learn it in three separate maps, because a site in one map and not the others is synced by nothing and reported by nothing — which is exactly how a sister site spent a day invisible last week; the wiki had to be reindexed, and the concept layer picked the new corpus up on its own with thirteen new backlinks. Two of the night’s corrections were to text written earlier the same day: the site gained a changelog at 22:06 and made false a sentence we had written that morning saying it had none, in two places. Alongside it, the work nobody sees: thirty-seven assets ingested out of the inbox in a single pass, the garden deploy put on a 6:31 PM timer thirty minutes behind the notes backup so the site is built from a day that is finished, and nine facts promoted out of one laptop’s memory directory into the repository. The tally for the day is one new public site and seven local commits with nothing for a reader to look at — and the seven are what make the one reach two people instead of one.
Insights and Takeaways
- Almost every fault found on Monday was the same fault wearing different clothes: a description that had stopped being true of the thing it describes — and not one of them errored. A structured-data block naming a different page. A collection listing seven of its eight members. A privacy notice saying three recordings when the site had five, for two sheets running. Three files insisting that deleting a draft branch takes its deploy down with it, when the deploy was still answering 200 with the superseded draft on it. A README announcing “There is no webfont” two screens above its own section on the typeface. An instruction file written at an import and overtaken by eleven commits the same evening. In most of those cases a gate ran and reported success, because the gates that existed asked the answerable question — does this parse, is every third-party origin named — while the unanswerable one sat directly beside it, untouched. The one that should be uncomfortable is the gate that printed
SKIP not foundand thenPASS, 0 problems. A page that measures nothing reports zero problems, which is indistinguishable in every other line of the output from a page that is clean. That had been true for as long as that gate had existed, and it took a page certified clean at zero tags — from a branch the page did not exist on — before anybody saw it. The lesson is not “write more checks.” It is that a second description of anything is a thing that rots, so the three checks written on Monday all compare a description against its source rather than against a hand-kept list, and every one of them was made to fail first, in nine shapes across two of them, including the two ways a gate can be quietly defeated by somebody editing around it. A gate believed before it has failed is a sentence, not a check. - The knowledge system spent Monday learning that a memory is single-player, and that we had been running the publishing side of this Foundation on nine of them. They lived in a memory directory under one home folder on one laptop, which means Helen Edgar’s Claude structurally cannot read a word of them — so the pair of dollar signs that WordPress silently replaces with a picture of itself, the apostrophe a heading slug turns into a hyphen rather than into nothing, the one surgical editing path on a published page that keeps a revision so there is an undo, and the way our editor backslash-escapes emphasis every time it saves were all sitting where exactly one of two contributors could reach them. Every one of them had already been paid for once, on a live page — a deficit figure that vanished out of the sentence whose whole job was publishing it, two table-of-contents anchors pointing at headings that do not exist, twenty-one stray asterisks in a published reflection. They share a shape, and it is the reason a pre-publication check can never catch them: the draft is clean and only the rendered page is wrong. So each one now names what to run afterwards. The uncomfortable half is not that those facts were unwritten. They were written, carefully, and kept faithfully — in a place that had quietly stopped being the place where the work happens. The same is true of the Foundation’s own legal footing: our accountability page calls us a public charity while linking the filings of a private foundation, and the filings are what is authoritative. That correction, too, had been living in one session’s memory. A fact only one contributor can read is not documentation. It is a coincidence that keeps working.
Gratitude
- Helen Edgar, whose Deep Pressure card settled a design brief nobody had managed to argue their way out of: Penguin Pebbling had launched on the same cream ground as two other sites of ours, and the palette that replaced it was sampled out of her artwork rather than chosen — the page built on the hue of her sage, the five locution tints read straight from her five card pigments. Sage and cream is the pairing she drew into that card herself, so her cream artwork now sits on a sage page as a deliberate combination rather than a clash to manage. The brief was answered by looking harder at work that already existed. Alongside her, Tressie McMillan Cottom, whose distinction between visibility and legibility gave Sheet No. 14 its spine — and whose harder reading of the life on that sheet sits inside it rather than after it.
Tuesday — 2026-09-15
Accomplishments
- Forty-five commits across four sites on Tuesday, and the honest accounting is how much of Tuesday was spent undoing Tuesday. The Map of Monotropic Experiences went from a repository to a live site with twenty area pages in a day — and the same day deleted the view switcher it had built that morning, dropped a changelog section it had published twice, and corrected its own measurement of an image format it had recorded hours earlier as dramatically better. It also spent its entire first day serving a typeface it had only named, because the fallback is a real font and a page with no font file behind it looks fine to everybody checking it. Penguin Pebbling shipped offline support twice before it was offline once: the worker registered, activated, controlled the page and held nothing, and the fix for that could not reach a single browser, because the file it changed is not the file a browser compares. The cost worth naming, though, is the part that is not ours to finish. The forty passages on the Map’s area pages are our words sitting on Helen Edgar’s artwork, and she has reviewed none of them — a packet exists now so she can read them in one place, which is also the admission that for a day they were unreviewable as well as unreviewed. Her watercolour pebbles could not survive the card following the theme; two techniques were tried, both failed, both were screenshotted rather than argued about, and the drawing was left alone instead of degraded. And the re-set deck — the most visible change to her game yet — sat unpushed until Ryan had seen it, because the entry directly above it in that repository’s decision log is about a visible change that shipped before anybody had looked.
- Publishing Our Sibling Sites took an afternoon of work that appears nowhere in the post. Before we could write that the five sites exist together, they had to be reachable at all: a footer column, seven navigation pages, the front page, and the plain-language home page — and the footer was only cheap because the sitewide block had been carrying two literally empty columns on every page of the site for long enough that nobody saw them any more. What the day corrected on the way out is the uncomfortable part. Our own design-patterns page had been publishing four zones while the canonical page publishes five, and that error had already propagated onto a new page on another domain before anybody caught it. The Neurodivergent Knowledge Garden had seven first-level headings and had not been touched since March. The Knowledge Garden field guide — eight days old, careful, thorough — had drifted eleven claims, two of them false rather than merely stale. And the checker built to catch exactly that had been reporting 64 of our own files as a colleague’s writing, because three of our sites had never been added to the list of what counts as ours. What the day left open is smaller and still open. The retitling backlog that field guide called currently clear is 182 of 622 assets. And the Cavendish domain moved in the evening, which meant four separate maps in this repository had to follow it the same night — a sync run before that would have quietly refiled eleven mirrored pages under a folder named for a domain that no longer serves anything.
Insights and Takeaways
- Almost every fault found across the sibling sites on Tuesday was the same fault: a claim that looked exactly like the thing it claimed. A stylesheet naming Atkinson Hyperlegible with no font file behind it renders in a real fallback and looks fine. An encoder’s lossless flag returns a file that decodes to 2.8 million altered bytes of somebody else’s artwork — and silently ignores the parameter that would have fixed it. A service worker reports 39 of 39 URLs held. A hotspot layer positioned with inline styles sits perfectly over the map. Every one of those is correct on the machine you are working on, and three of them were wrong only in production, where the Content Security Policy a dev server never sends is the thing doing the blocking. The fourth ran the other way — a worker that cached everything locally and nothing live, one commit after a worker that cached everything live and blanked the whole site the moment the local server stopped. A development environment is not a weaker version of production. It is a different set of permissions, and a service worker touches both halves of that split. What caught all of them was the same act in every case: decoding the file, counting the entries, rendering the page, and looking. So the tools grew comparisons rather than assertions — an image encoder that decodes its own output and deletes it on a single differing byte, a build that fails when a named typeface has no face serving it, a check that refuses an image source pointing at a file that is not there. That last one is invisible by design, which is why it needed a gate at all: the browser drops silently to the fallback, so the page looks perfect while every visitor downloads the big file.
- We spent Tuesday morning building a guard that measures whether a mirror could silently shrink, and Tuesday afternoon finding that nothing anywhere measures whether a site we published is linked from anything. The morning was the good version of the day: four mirrors, each carrying a minimum page count, none of which had ever been compared against the site it guards — and one of them would have accepted the loss of eighty-nine pages without a word. Real gap, closed, and the measurement took under a minute. Then the afternoon: a fetch of our own front page came back with 419 KB of navigation and footer naming not one of the five sites we had built, three of which were linked from zero pages on stimpunks.org. The commitment to fix that was already written down as an objective — written when there were two sites, and sitting at zero while the count reached five. Nothing noticed, and nothing could have. Every guard in this system counts something: files deleted, pages mirrored, floors against sitemaps, deletions against a threshold. Not one of them asks whether a thing we published can be reached from anything else we published — and unlike a shrinking mirror, that failure has no signal at all, because from inside the repository a site nobody links to looks exactly like a site everybody does. We can build a site faster than we can add a link to it, and the only instrument that found this was somebody fetching our own front page and reading it.
Gratitude
- Helen Edgar, who made two calls on Tuesday that unblocked work waiting on her: Penguin Pebbling is CC BY-NC-SA 4.0, which adds the ShareAlike the old sentence had left open, and the card text under the artwork is approved as built. Her Map of Monotropic Experiences got a home of its own the same day — her artwork, her CC BY-SA licence, twenty areas, and a map that stays a map because the site refuses to score anybody on it. Alongside her, Sylvia Duckworth, whose Wheel of Power and Privilege appears on one of the training slides with her handle printed inside the crop, where it stays.
Wednesday — 2026-09-16
Accomplishments
- Forty-two commits across two sites, and the honest description of most of them is repair of something already live. Cavendish Space ran from 08:17 to 23:47 in twenty-five commits, and the one that should sting is the print PDF: the deck was rewritten for all ages in July, the web version moved with it, and the sheet a facilitator prints and pins up went on contradicting the project’s own settled position for two months — not because anybody decided to leave it, but because rebuilding it needed a virtual environment whose setup lived only in prose. A generated file with a hand-run build step is a file that will be wrong, and it was. The security contact shipped and 404ed on its first attempt, found only because somebody checked the deployed URL instead of trusting a build that reported success. And the group-needs method took four independent test sessions in one evening to shake out, each of which came back with the mapping correct and a rule the page had never written down — including one hole that the same evening’s work had created. On the Map, the day’s fourteen changed area pages are Helen Edgar’s review, not our writing: the packet went out, the notes came back, and the work was applying somebody else’s corrections to our own sentences on her artwork. The cost of the day is not the hours. It is that almost none of this was new ground — it was the interest owed on things shipped earlier that nobody had gone back and looked at.
- The one post stimpunks.org published on Wednesday spent the rest of the day being diagnosed, and the diagnosis is what the day actually cost. The Week 37 changelog went up carrying a hazard note about our own editor rewriting files on save — and shipped with that exact fault in its body, 105 stray asterisk runs, in the one post whose job was to warn about it. Everything a check could catch came back clean: 62 links, both labels, a featured image, thirteen anchors landing on real headings, and zero of the money-figure corruptions the
$discipline was written for. The part nothing caught was the part we had written down ourselves and written down wrong. So the rest of the day went into two documentation commits with nothing for a reader to look at: the hazard bullet rewritten to cover both mechanisms with a separate check for each, a warning to diff before restoring from git because Ryan’s own edits live inside the damaged copy, and the structural fix recorded — the publishing copy should never be the copy an editor has open. Alongside it, a capability this file had recorded as broken for ten days was re-tested and found working, read and write, which means ten days of sessions were steered away from the right tool by our own note. Two commits, no publication, and the day’s real output is that two of our own instructions stopped being wrong.
Insights and Takeaways
- Both sites learned the same thing on Wednesday from opposite ends: the document was never the test, and the person using it was. On Cavendish the method was written carefully, with three rules, a mapping table and an answer key parsed out of the page itself so it could not drift — and then four independent sessions ran it and every round returned the mapping correct and a rule the page had never thought to write. That a spread carries the app’s reflection questions, which name cards nobody laid. That the page said which cards map to which condition and never which realm a card belongs to, so the tool had to infer from tone the very thing the page forbids inferring. That
something softhad been mapped by its name rather than its words and survived two whole rounds that way. Twice the failure was the worked example breaking the rule printed three paragraphs above it — the part nobody proofreads sitting directly above the part a reader copies. On the Map, Helen Edgar opened the builder and reported the board not staying put while she scrolled: not a browser bug, a height threshold set at 704 pixels that excluded most of the windows the feature was built for, invisible to anyone testing in a tall window. Then she did not find the land backgrounds, which were third in a tray of ninety-five and had been moved there for a good reason she herself had given — open with what joins a map, not with separate islands. Order is discoverability, and it beat the argument it was carrying. The uncomfortable shape is that every one of these was correct reasoning producing a wrong thing. Nothing was sloppy. The rules were right, the demotion was right, the threshold had a real justification behind it — and none of that survived contact with somebody actually using the thing. A rule written by the person who built it is a hypothesis, and the only instrument that tests it is a stranger’s hands. - Our own notes about our own faults failed twice on Wednesday, in the two ways such a note can fail, and both of them cost more than the fault would have. The hazard bullet about our editor rewriting files on save named the right culprit and half its method. It documented the backslash escapes and prescribed a grep for them — so the check returned a number, the number looked like the problem, and nobody kept looking for the fifty-one invisible spaces that were the actual reason seven bold runs never opened at all. A check that answers on the wrong signal is worse than no check, because it ends the search. The other failure was the mirror image: a capability recorded here as does not work, confirmed twice, true when it was written — and still being obeyed as a constraint ten days after it stopped being true. Nothing surfaced that. A note saying this is broken generates no error, no failing test, and no second look; it just quietly routes every later session away from a working tool, and it does that for as long as nobody re-runs the thing. So the pair of them is one lesson. A note about a fault is a claim with a shelf life, exactly like the site descriptions that rot and the print PDFs that fall behind their source — except that this kind of claim is the one we consult instead of measuring, which is precisely why it is the one that goes unchecked the longest. The only remedies either of these suggests are unglamorous: give the check a way to fail on the real signal, and re-test a recorded impossibility before trusting it, because the record of what does not work is the part of a knowledge system nothing else in the system ever contradicts.
Gratitude
- Helen Edgar, twice over in one day and in two different registers. She read forty passages of ours written onto her own map and sent back notes on eighteen — including the first name of an area’s namer that our page had never carried and the source article still does not, supplied rather than guessed, which is the whole reason to ask. Then she picked up the builder and told us the map would not stay on screen and that she could not find the land — two sentences that between them corrected a threshold, a tray order, and an assumption about who is looking at a tall window.
Thursday — 2026-09-17
Accomplishments
- Thirty-two commits across two sites, and the honest accounting is that Cavendish Space spent most of Thursday paying for documents that had already been printed and handed to people. Three card names on the facilitator sheet were wrong — a What helps card listed among the lily pads, a phrase that is not a card in any realm, an article dropped off a name — and all three were in the printed starter deck too, from the same original sentences, which means every copy run off and put in somebody’s hands carried them. They were found by reading the sheet line by line against the deck’s own data rather than against itself, which is a thing nobody had done. Six more documents still told a reader to look for placeholder markings on the card faces after those markings came off. The AAC boundary had been stated in four places, one paragraph each, and one paragraph had not been holding it — the deck keeps being mistaken for a communication board, so the boundary needed a page that argues rather than asserts. And the watering hole was described by how it feels rather than what it is for in three separate files, with the first repair making the same mistake in plainer words before Thornburg’s actual definition settled it. The genuinely new work of the day is real — the whole deck prints now, the worn realm can be made into badges, nineteen menu entries stopped being unreachable on a phone — but the shape of the day is the same one Wednesday had. Almost none of it was ground nobody had walked. It was the interest owed on things shipped earlier by people who were sure at the time.
- A stranger’s email got answered on stimpunks.org on Thursday, and the cost of answering it was discovering that the garden had held the answer all along and could not hand it over. A special ed teacher in Portland asked how to start a Neurodivergent student group. Every part of the answer already existed here — across five corpora, plus a thirteen-file toolkit from the University of Edinburgh that has been sitting in
raw/and quoted on our own Alternatives to ABA page, reachable only by somebody who already knew to search for it. The affinity group concept page was a bare link hub with no synthesis, so there was no single place a person could be sent. Writing the page was assembly, not research, and three things went wrong on the way out that a draft check could not have caught: a converter that merged four of the toolkit’s ethos concepts into one list item because paragraph-joining ran ahead of the bullet split; a parent page reported as needing a manual link when it lists its children automatically, invisible in the mirror because a mirror stores rendered output and an auto-generated list looks exactly like a hand-written one; and the age question answered backwards. Separately, the cavendish.app retirement swept thirty-one occurrences across seventeen pages, a glossary term and the sitewide footer — and four of the day’s WordPress edit calls were spent rediscovering parameters that were never written down, every rejection counting against a rolling daily cap. The five dated posts keep the old domain deliberately: they said it because it was true the day they published, and editing them would make the archive claim we wrote something we did not.
Insights and Takeaways
- Every check that failed on Cavendish Space on Thursday was working correctly — each one compared a document to itself, and the errors were all in the gap between a document and the thing it describes. The card errors survived because prose gets proofread against prose and never against the deck’s own data. The nav parity check reported that five copies of the menu held the same links while saying nothing about whether the groups matched, or whether any of them could be reached on a phone — the open menu overflowed a small screen by 81 pixels and two entries could not be tapped at all, which is not a thing a list of hrefs can see. The print stylesheet rule meant to strip the page furniture matched nothing at all on generated pages, and a stylesheet rule that matches nothing fails without a sound: the deck paginated to eighteen sheets instead of fifteen and the badges printed five percent undersized, both found by rendering the pages rather than reading them, exactly as ★ stuff found three figure defects by screenshotting them at the size a reader sees. The pattern is not carelessness and there is no version of it a more careful reader catches. A check knows what it was told to compare. Every one of Thursday’s faults lived on the axis it was not told about — the deck behind the sentence, the screen under the menu, the paper under the page — and each new guard written on Thursday is narrower than it looks: the card-citation check says in its own docstring that it cannot catch an invented phrase in plain prose, because that is how one of this week’s three got in and was found by a person reading. The only honest move is to say what a check does not cover in the check itself, so the next person to trust it knows the size of the thing they are trusting.
- Searching for a rule, finding none, and reporting the absence as a position is the way this knowledge system fails that no amount of searching harder will fix. Thursday’s page needed to say what ages a Neurodivergent student group should hold. The search looked for an age restriction, found none anywhere in our material, and concluded the policy was open — when multiage is a mission-level commitment pointing the other way entirely, our own learning community is multi-age, and the learners in it run the organisation. The absence of a prohibition is not a position, and reading it as one had this page about to publish a shrug where we hold a conviction. The correction earned the page a whole section, because a high school will reflexively split a group by grade and the age spread is the thing that makes a small group work: a senior who has survived three years of that building is more credible to a frightened ninth grader than any adult in the room. The same day showed the twin failure from the other side. Our service directory has twenty-eight state pages, nineteen of which are a wildfire-alert link with a state name on top, and the one page with real substance is Denver, written by a named person at Denver Human Services in their own words. So the new page refuses to publish a list of Oregon organisations and publishes a filter instead, and asks the reader to send local knowledge back. Desk research produces a page that looks like Denver and ages like California, and a stale referral sends a Disabled person to a door that is not there — which is a worse outcome than saying we do not know. Between the two, the lesson is the same shape: what the garden does not say is the least reliable thing in it, and the fix is never to search again but to go and ask somebody who knows.
Gratitude
- Helen Edgar, whose Zone Explorers became Place Explorers — the whole Cavendish model said smaller, for children of about seven to eleven — and whose twenty years of work on Profound and Multiple Learning Disabilities is credited by name in the implementation layer, where Thursday found we had been dropping a word out of the name of the term. And David Thornburg, at one remove: the watering hole had drifted into being a second cave in three separate files, and it was his actual definition — peers acting as learner and teacher at once — that said what was wrong with all three.
Friday — 2026-09-18
Accomplishments
- Cavendish Space did its largest structural day of the week between a quarter to one and a quarter to three in the morning, and three of the things it shipped had to be repaired before the sun came up. Two framework pages, a new card, two new signposts and a PMLD section is a real day’s work by any measure — and the honest accounting is that the first pass at the SPACE-TIME tiles put white on dark-mode cyan at about two to one, that a page went live saying “SSensory attunementThe channels a bodymind takes the world in through” to anyone whose stylesheet had not caught up, and that the person who found it was a reader on the live site, not a check here. The deeper cost is the age of what came up with it:
md_to_htmlhas never had a rule for numbered lines, so the three steps for reading a spread have been shipping on the example spreads page as one run-on paragraph with the numerals stranded mid-sentence, and nothing ever failed, which is exactly why it lasted. The night also spent time correcting itself twice over — a claim logged the day before that novelty had no home in the deck turned out to be wrong, and a decision-log entry written at 01:35 saying nothing checked the two signpost lists was corrected at 01:39 by the check that now does. Four Patterns and two placed cards were left as open questions rather than settled while the file was already open. - Friday finished the Pattern Library at 56 of 56 and made A Pattern Language readable and credited — and both halves of that were interest owed rather than new ground. The book our entire form comes from has been sitting in the library since May as an image-only scan returning 313 characters, one per page break, invisible to every search and every concept page for four months, and the four pages it shaped carried no credit to it at all. Recovering it surfaced nine more books in the same state, including Helen Edgar’s illustrated ebook on monotropism — the concept most central to our work, in a book no sweep of this garden has ever been able to see. The Library side cost a full re-canonicalisation before a single missing page could be written, because the index had been contradicting itself about what was missing. And the day’s verification failed in the open: the live link sweep tripped the site’s own bot protection partway through and started returning 403 for URLs that had answered 200 seconds earlier, so that run proves nothing in either direction and is not being counted as a pass. Eighteen older dead links across the patterns and design tree are still there, several of them “working” only through the guesser, and four crosswalk gaps stayed open.
Insights and Takeaways
- Every naming decision on Cavendish Space on Friday was load-bearing, and the deck’s own note on the hardest one is the thing worth keeping: if a name has to be defended, it is doing work a name should not have to do. The two new signposts could have been Predictability and Novelty, which are the words the field uses. They were rejected because those are principle names, and a signpost has to name something a person can recognise — and because Predictability would have reached over and pulled tell me first, a card about warning rather than sameness. So the signpost was named for the channel instead, Same & new, which holds both directions the way Sound already holds turning it down and adding a steady hum. Then the sting was written down beside it: if tell me first turns out to belong with sameness after all, the channel name was chosen partly to keep it away, and that is a name propping something up. The same shape ran through the rest of the night — something different was left in no signpost rather than inventing a thirteenth, because that is Helen’s call and not one to make while adding a card; four Patterns with real pull from already-placed cards were logged as open rather than mapped, because mapping a placed card is a deck decision that must not ride along on a registry rewrite. The discipline is not caution. It is refusing to let a decision get made as a side effect of a smaller one, while the file happens to be open.
- Three separate faults on Friday were the same fault: a correct instruction that nothing enforced, and nothing anywhere that could tell it had stopped being true. The deck’s Pattern registry carried a comment telling whoever came next to keep its published flags current — 29 of them went stale in eight hours with every single build passing. The installer built to catch drift between the seven unattended task prompts in this repo and the copies actually running on the machine had never once reported more than a single task, because it compares files with
diff,diffexits non-zero precisely when it finds something, and the script runs underset -euo pipefail: it died at the first difference and exited 1, which reads as found one rather than stopped looking. And the concept layer fingerprints each asset by size and modification time, so recovering 365,571 characters of Alexander changed nothing — the book stayed at zero concept hits with its own text sitting in a cache one function away — then the new OCR cache made the same mistake within the hour, because sixteen files marked done had not failed, they had merely gained nothing, and nothing recorded which method had said so. A cache keyed on its input cannot notice that the code changed, and an instruction with no enforcement is a wish. The repair in all three cases was the same: give the thing a version or a check, so that improving the method re-derives what the method covers.
Gratitude
- Helen Edgar, twice over and from opposite directions — her twenty years of work on Profound and Multiple Learning Disabilities is what the new multi-sensory section of the implementation guidebook is built on, and SPACE-TIME is her framework, dual-licensed so it could be carried rather than only rewritten. Her illustrated ebook on monotropism, with Marion McLaughlin and Felicity Goodhall, is also one of the ten books this garden could not read until Friday. And Christopher Alexander, Sara Ishikawa and Murray Silverstein, whose 1977 book about towns said in plain words what we hold about people: in a healthy society there will be as many pattern languages as there are people.
Saturday — 2026-09-19
Accomplishments
- stimpunks.love went from first commit to live in thirty-four minutes, and then ran another six hours — almost all of it on things the launch had just finished listing as not built. First commit 16:35, live at 17:09, last commit 22:57. The changelog we shipped at launch named three absences — the audio room, the polaroids, a tenant for Your Room — and by midnight two of the three were gone, which is a good evening and also the reason a changelog that lists what is missing has to be regenerated rather than written once. What the hours actually went to was rarely the feature. The audio room needed a licence change before a single file could be accepted, the polaroid wall needed the whole consent process written and four refusals verified against fixtures before the first photograph could go up, and the first photograph then caught the plate cropping people the terms promise not to crop. The yell button had to be rebuilt from an oscillator that could not sound like a shout, and the fix was measured by zero-crossings rather than judged by ear, because we cannot hear it. Three of the evening’s pieces were repairs to promises the site had already made in writing.
- The Why Sheets got a press, and the expensive half was not the press. whysheet.press is the logistics — tick what you need, get one correctly paginated PDF — and the logistics are what eats an evening at eleven o’clock the night before a meeting. But the day’s real work was settling what the library actually contains. When the build started there were three answers to how many Why Sheets exist — fourteen files, twelve published pages, nine named on /why/ — and reconciling that to one number, with a check that says so out loud if it ever stops being true, came before anything could be printed. Five older sheets gained a What to Ask For in the Room section, AAC Is Not a Last Resort published as the seventeenth, and 126 non-breaking spaces came out of nine older sheets — Ulysses damage that would have printed as literal asterisks in a document a family carried into a meeting.
Insights and Takeaways
- We held photographs to a privacy rule and never asked whether the rule applied to anything else, and it did — to five files we had already published. Enid’s wall refuses a photograph carrying EXIF, on the reasoning that sending us a picture should not tell the internet where you live. Audio was never held to the same rule. The first community yell arrived carrying the recording app, the device, the OS build and the exact second — and so did all five readings already live on the site, which nobody had checked, because the rule had been written for one medium and the other was never examined. A voice identifies a person as surely as a face does. All six are stripped, re-muxed rather than re-encoded so the sound is untouched, and both audio tools refuse a file carrying any now. A privacy rule scoped to the medium you were thinking about when you wrote it is not a privacy rule, it is a habit — and the tell is that the wall’s own terms were already general enough to cover the case, while only the photograph code enforced them.
- A checker that reports “missing” when it means “shaped differently” points at the wrong repair, and the wrong repair here was to write over better prose than we would have produced. Our tooling read the Why Sheets for a What to Ask For in the Room section and reported Monotropism as empty. It was never missing one — it has had “Questions To Ask In The Room” all along, written as a bold label above each question rather than as a list, and the checker read only lists. Taken at face value, that report would have led to adding a new section on top of one that was already there and already better. The same day ran the other half of the lesson at the press: with checking tools and deliberate care we nearly shipped a Pellegrini quotation under the wrong publication, and Ryan then put the actual paper in the inbox, which turned a secondhand citation into a primary one. So the defence built into the letter prompts is structural rather than an instruction to be careful — the generator throws on a quotation with no attribution, and refuses outright for the school sheets with no what to ask for section rather than let a model invent what to demand of a school. A named backlog beats a quiet degradation, and an assistant’s invented study hands a school a free reason to dismiss a parent who is already outnumbered.
Gratitude
- Michael Masutha and William Rowland of Disabled People South Africa, who gave the movement “Nothing About Us Without Us” in 1993, and Eric A. Meyer and Sara Wachter-Boettcher, whose Design for Real Life our philosophy page has credited as an inspiration all along. Both sayings went into the talking chair on Saturday, and the chair names whose they are aloud rather than leaving the credit to the liner notes — which is the whole argument the zine table next door spends a page making.
Looking Back
Meaningful Moments
One question, and a justification dissolved. Our attributions file says that reproducing a publisher’s record over a living author’s own name is a database outranking a person, and the entries around that sentence printed a deadname four times. The reason written beside it was findability. The DOI is what makes a paper findable. That was the whole conversation, and it was enough, because the old name had never been doing the work it was kept for. Nothing in our tooling could have raised it. Somebody had to ask.
Ole Worm comes out of nowhere. Helen’s note after Held, Not Fixed published was one line long and right, and three sentences now say who he was. He kept a unicorn horn after proving it was a whale’s tooth, and did not throw it out. The wonder survived the correction — which is what the sheet argues, and what her note did to the sheet.
A senior and a ninth grader. Writing the Portland page, we searched for an age rule, found none, and nearly published a shrug where we hold a conviction. The correction gave the page its best sentence: a senior who has survived three years of that building is more credible to a frightened ninth grader than any adult in the room.
Life is Teaching Me
Silence is not a finding.
The week kept producing the same shape. A gate printed SKIP not found and then PASS, 0 problems. A search found no age restriction and read the absence as a policy. Two hundred and thirty-three session summaries were written faithfully and never cited once, while a whole project sat at zero of its thirty-two with nothing saying so. A note recorded a capability as broken, and ten days of sessions obeyed it after it had started working. Three of the sites we built were linked from zero pages on stimpunks.org, and nothing we run counts that.
In every case the quiet was read as an answer. No error became no problem, no rule became no position, no record became nothing happened. The fix was never to search harder. It was to re-run the thing, fetch the page, or ask somebody who knows.
Habit Insights
The habit doing the most work this week had a name, and it was Helen Edgar’s. Six of this week’s seven Gratitude entries open with her, and none of them is a courtesy: she sent notes on eighteen of our forty passages on her map, supplied a first name our page had never carried rather than letting us guess it, found the scroll threshold and the buried land tray by using the builder, and caught Ole Worm. A second reader is not a kindness to the author. It is the check that catches what the author’s own checks were built on.
The documenting habit is strong and the reading-back habit is weak. On Wednesday we rewrote the hazard note about Ulysses rewriting files on save, and recorded the structural fix beside it: the publishing copy should never be the copy an editor has open. The changelog for this very week then published with thirty-two stray asterisks anyway, since repaired, because the fix is written down and not wired in. This document is clean as of this writing — zero backslash escapes, zero non-breaking spaces — and that is a count taken today, not a promise about the next save.
Looking Forward
One Change I Can Make Next Week
Correct the licence line on /together/ before opening anything new.
Week 37‘s one change was to turn the record of what we could not verify into a list with dates on it. Week 38 did not do it. Eighteen files still cite a domain that no longer exists. Eighteen older dead links across the patterns and design tree are still there. The retitling backlog is 182 of 622. And as of this writing /together/ still licenses Kaya Oldaker’s Autism Infinity Dragon CC BY-NC when she released it to the public domain — the page was edited on the Sunday of this week, and that line survived the edit. This is the fourth week running that a reflection has named it.
So the change is not a better list. It is one match-once edit, made first, before the list. A resolution carried four weeks is no longer a plan; it is the record of one.
I’m Looking Forward To
Hearing back from Portland. The page closes by asking readers to send local knowledge back, because desk research ages badly and a stale referral sends a Disabled person to a door that is not there. The first reply from somebody who knows their own town is the one thing that page cannot write for itself.
#changelog #reflection log#