Stimpunks.org Changelog for Week 32 2026: From the Pile to the Garden, From the Claim to the Count

This week we built the thing that remembers for us. The Stimpunks Knowledge System went from nothing to a working system in seven days — a shared, self-maintaining home for our research library and our organizational knowledge, mirrored to two places the whole team can reach. Then we documented it in the IT Field Guide, argued it as curriculum, and gave it away as a template anyone can take. Along the way it learned to index ideas instead of filenames, told us which books we had been arguing from without owning, and caught us being wrong in public more times than we would like. Bus factor 1 is a design failure, and the mechanism is almost never a bus.

Throughline

Bus factor 1 is a design failure, and the mechanism is almost never a bus. It’s burnout, a flare, leave, or six weeks where reading a PDF is not going to happen — normal human variation, and a system that treats it as an exception is the thing that’s badly designed. So we built a knowledge system that no single laptop and no single person is a single point of failure for, then published how it works, why it’s a learning practice, and how to start your own. Interdependence means building so that any one of us can be absent.
A filename index tells you what you own. It cannot tell you what you understand. Our library could list files; it couldn’t answer what the garden holds on monotropism. Now it indexes ideas, keyed to the glossary we’ve maintained in public for years — and the first thing it did was tell us we were standing on sources we didn’t have. Four books later, that’s less true. A pile of files is not a memory until something knows how the pieces connect.
A check nobody has watched fire is not a check. We learned that six separate times this week, and twice it was expensive: a sync that deleted 1,014 of 1,466 mirrored files on a variable that meant the wrong thing, and a staleness alarm that was not miscalibrated but structurally incapable of sounding. Then we counted the animations on this site — we believed there was one, and there were 80, none of them honoring reduced motion. A practice you cannot count is a practice you are not actually running.
Having a page and having a path to it are different achievements. The Design Field Guide went from five pages to nine, How We Tell Our Stories became three pages that each say what they are, and 29 glossary links we cite in good faith turned out to go nowhere. Only one of those achievements is finished when you hit publish.

The Work

The System That Remembers For Us

Behind the scenes this week we stood up the Stimpunks Knowledge System: a shared, self-maintaining home for our research library and our organizational knowledge, mirrored to two places the whole team can reach — our GitHub organization and our team Google Drive. It indexes our source documents into a browsable, auto-generated wiki, carries 12,878 highlights across ~1,460 sources alongside our living notes, and keeps itself current on a weekly schedule: pulling in our ecosystem data, regenerating the index, auditing itself for broken links, and backing up to both homes on autopilot.

Then we built the plumbing that keeps it healthy — a drop-here tray that reads each file, shelves it in the library, and empties itself; a family of small, narrow tools that each do one job; and a guard that sweeps stray artifacts out of every automatic backup. We distilled our House Style Guide and its companion guides into one canonical statement of the Stimpunks Editorial Voice, in force for everything we publish, and wired it in so it actually holds. A voice is only shared if it is written down where everyone — and every tool — can reach it.

We also taught it to answer in its own voice and only its own voice. Ask our Knowledge Garden Spider a question and it walks the trails of the garden first, then answers only from what it finds there, citing the file or link next to every claim and refusing to fill any gap from training data. When the garden is silent it says so and names what we’d need to add. The scope is the safeguard: our corpus is neurodivergent- and Disabled-authored, identity-first, and counter-deficit by construction, so an answer grounded in it stays counter-deficit too. An entrance, not an oracle.

The public versions landed midweek. The Stimpunks Knowledge System in the IT Field Guide documents the motivation, the architecture, the rules that hold it, the 3-2-1 backup channels, and a schedule with no streak to keep and no dashboard turning red — because urgency cues are hostile design in a system meant to hold knowledge for people whose capacity varies. It also says what this isn’t: we’re wary of techno-solutionism and technoableism, and a folder structure fixed nothing structural. It doesn’t create capacity; it stops us spending the capacity we have on finding things we already had. Two sections earn their keep hardest — Honest limits, which admits the library is readable by everyone while the machinery is still understood in depth by one person, and Steal this, which gives away the portable ideas, because a resilience practice you keep to yourself isn’t much of one.

Then The Stimpunks Knowledge System as Curriculum asked what any of it has to do with learning. It opens on a design failure we don’t name often enough: school produces knowledge work for thirteen years and then throws nearly all of it away. An essay is written, submitted, graded, returned, and lost. We don’t do that to professionals — a researcher keeps a bibliography, a lawyer keeps a file, a developer keeps a repository. So the page makes two arguments. Help students cultivate a knowledge garden — digital gardening, the old commonplace book with hyperlinks — which turns out to be a constructionist project that never ships, all four of project, problem, passion, and purpose at once, and one of the few structures that treats monotropic depth as an asset rather than a scope problem. A special interest becomes a region with real expertise in it — cognitive liberty made concrete. And teach change management and revision control. A commit history is documentary evidence of iteration — the attempts in order with the reasoning attached, produced as a byproduct of the work instead of performed afterward in a reflection paragraph. School assesses the last frame of the film and calls it the movie. Version control is psychological safety built into the tooling rather than promised by an adult: nothing you try can destroy what already worked, so learners who cannot lose their work take bigger swings at it. The page is equally clear about what it can’t claim — we haven’t tested this with students, grading a garden would kill it, and any plan assuming every learner has a device and a subscription is a plan that widens a gap.

A Pile of Files Is Not a Memory

Our wiki could tell us what we have and almost nothing about what we know. Ask it what the garden holds on monotropism and it could offer three files with the word in the name, while the actual answer sat scattered across the website, the highlights, and a decade of notes. So we built a concept layer: one page per concept, each gathering everything we hold on that idea from every corpus at once.

The best decision was the one we didn’t make. We were a step away from inventing a taxonomy when it turned out we had published one years ago. Our glossary is 445 terms deep, so the concepts are keyed to it — the vocabulary we already maintain in public, not a private scheme a machine dreamed up. Two problems were worth solving properly. Counting mentions is useless: 335 of our site pages “mention” monotropism, nearly all of it navigation furniture, so matching now weighs a title far above a heading and a heading far above a passing mention. And our own library was unreadable to us: most of it is PDFs, so a paper filed as 40614_2022_Article_338.pdf was invisible to every search we could run. We extract text now, and that paper turns out to be a scoping review of applied behaviour analysis.

Then the layer told us the truth about ourselves, and none of it was flattering. We are standing on sources we do not have filed in our knowledge system. Deficit ideology is load-bearing across our site and rested on exactly one asset that so much as cited Paul Gorski. Damian Milton was referenced inside 79 of our files and we held not one paper of his outside of bookmarks. Anna Stenning’s Narrating the Many Autisms quotes our Cavendish Space framing — our best evidence the idea travels beyond us — and the book was not in the library.

By that afternoon four of those gaps were filled, which is the first time our own tooling has sent us shopping. Stenning’s book reads better than the glossary quote suggested: she takes the whole spatial vocabulary — the instructor-led campfire, the peer campfire, the watering hole — and attaches our framing to a universal claim, that intermittent collaboration benefits all learners. That is universal design, not accommodation — the distinction we keep insisting on, made by someone else, in print, without us in the room. The next day we got the Milton arc end to end: the 2011 Autscape talk that argued the ground before the term existed, the 2012 paper where the double empathy problem was named, and the 2022 ten-years-on editorial. His argument is not “Autistic people have empathy too.” It is that cognitive-behavioural accounts decline to discuss relationality at all, and that a supposed deficit in theory of mind is better understood as a question of reciprocity — an argument made through Harold Garfinkel’s sociology rather than on psychology’s own turf. Read it as an empathy shortfall shared between two people and you have quietly reinstalled the deficit model it was written to escape. The paper declares identity-first language in its own opening, citing Jim Sinclair’s 1993 “Don’t Mourn For Us” — so that commitment is original to the concept, not a courtesy added later by people like us.

We also made the library searchable for the first time. We installed qmd — keyword search, vector search, and a reranker, running entirely on this machine with no vendor, no network, and no API key, which is the only reason it belongs anywhere near our provenance. The clearest proof was a query we could not have answered before. “Why does being interrupted mid-task hurt so much” finds nothing by grep — not one file, anywhere, because none of those words are the words we use. Search it properly and back comes our own Tendril Theory, a post about protecting flow, and a paper on transitioning in and out of Autistic flow. That is the whole point: a person in distress does not arrive knowing our vocabulary. They arrive saying it hurts. Something has to carry them from their words to monotropism, and until now that translation lived only in the head of whoever happened to be asked.

Nine things in the library had extensions but no contents — a .gdoc on a Mac is a 173-byte stub holding a Drive id, so nine research assets sat there indistinguishable from corruption. Seven are open now, including a whole anti-behaviourism school behaviour policy, Human Restoration Project’s case for eliminating homework, and a long winter crip survival guide by Leah Lakshmi Piepzna-Samarasinha and Tina “constant tt” Zavitsanos. Two we could not open, and left alone — their Drive targets are gone, so those dangling links are the only surviving record that we ever held them. They’re relabelled and they stay. The fix for a dead pointer is a legible label, never a deletion.

And the biggest correction of the week was ours. We had described those documents as privately shared, inferred from nothing but a Drive timestamp, and stamped five public community resources with do not quote without permission. Everything in our library was found on the open web; private and pre-published material never goes in there at all. So we tested it properly and rebuilt the rules on what we could verify instead of what the metadata implied. A timestamp records that a file arrived. It says nothing about why.

The Garden Was Never Silent

A Knowledge Garden Spider sweep went looking for what we hold on curiosity and came back empty. We nearly published a glossary draft saying our own library could not support the entry. Our Raindrop library had Susan Engel’s Children’s Need to Know: Curiosity in Schools the whole time — along with the Evo-Stress reframe of dysregulation, a Frontiers study of Autistic inertia, and Melnyczuk’s apraxia caveat on “behaviour is communication.” Four of the exact terms being drafted that morning. Two drafts had been written on a false premise.

The cause is worth naming precisely: our bookmark ingest stored records — title, url, tags, collection — and no page text whatsoever. So grepping ten thousand bookmarks searches ten thousand titles, and a tool that cannot see a corpus does not report thin coverage. It reports silence, with total confidence.

The zero-cost half of the fix shipped in ten minutes: our tools now require both sweeps before either is allowed to say the garden is quiet. Then the expensive half — 4,105 pages of real page text, 135.8 million characters, now fetched and indexed as a sixth searchable collection, each carrying the canonical URL so a hit cites the original and never our copy of it. What it refuses to store is the load-bearing part, and every refusal is counted out loud. Coverage is 38.5%, and the report says 38.5% rather than quietly rounding itself up. 4,685 bookmarks sit on hosts we skip deliberately — twitter.com alone is 4,156, where a status URL returns about a thousand characters of login wall. Another 1,163 come back blocked by publishers who refuse robots outright, and 210 are too thin to keep. Each one is recorded with its reason, because a coverage number you cannot inspect is indistinguishable from a coverage number somebody invented.

The query “curiosity engine of intellectual development school children questions” now returns Engel’s paper as the top hit at 96%. At breakfast that query returned nothing. We came within one paragraph of publishing “the garden holds nothing on this” about a shelf we simply had not built a door to — and an honest 38.5% you can audit is worth more than a hundred percent nobody counted.

Give Away the Thing That Works, Not the Thing That Flatters Us

Helen wanted her own version of our knowledge system, and the obvious answer was a copy of ours. That was wrong: twelve of our shelves were hardcoded, our own name was baked into the index, half the tools are welded to this website, and the library is a gigabyte she doesn’t want. A fork would have arrived broken and she’d have spent her first day un-Stimpunksing it. So the shelves became data instead of code, and we proved it twice — our index regenerates byte-for-byte identical, and a scratch garden with three invented shelves builds correctly with zero mentions of us anywhere in it.

Then the real find, which cost nothing and saved the most work. The on-ramp already existed. We’d written it in The Knowledge System as Curriculum: one folder of untouched originals, one index with a line about why each thing is there, a habit of adding to both. That’s it. That’s a garden. No app, no subscription, no AI. Three sentences, no tools, the whole idea — published on a page about teaching while everyone got pointed at the page about our system. The answer was never to write an on-ramp. It was to stop hiding the one we had.

There’s a template anyone can take now — knowledge-garden-starter — and a page to arrive at, Start Your Own Knowledge Garden. Five steps that degrade gracefully, and we say so out loud, because the opposite framing is how these things die: a tool that only pays off once you’ve finished configuring it gets abandoned during configuration, on a bad day, by exactly the person who needed it. Step one needs no software. Step two is git, and since Helen was learning the terminal this week, the walk-through is written for someone who has never opened one — what pwd and ls do, how to drag a folder from Finder instead of typing a path, a table of the errors you’ll actually hit, and an explicit list of the three commands that can destroy work so you know the four you’re learning cannot. Step five we left out on purpose: the concept layer keys on a glossary we spent years publishing, and asking a machine to invent your vocabulary is how you end up with a taxonomy nobody believes, including you.

Then a fresh clone told us four things we’d never have noticed from the inside. A README explaining “never edit this folder” was being catalogued as an item in it. An empty garden printed a dangling separator that looked like breakage. The banner announced the index was “generated by Claude” to people who may never touch Claude. And forgetting one flag printed a raw Python error at someone whose first command it was. None were functional. All were the first impression, on a kit whose entire audience is people not yet sure this is for them. You cannot test a welcome from inside the house.

The starter tells you to put your originals in one folder and never touch them, and then says nothing about how a web article gets into that folder at two in the morning while you’re reading on your phone. Filling the Garden is the missing half — the Obsidian Web Clipper as an intake pipe, Obsidian or Ulysses for writing, and not one terminal command anywhere in it, because the people most likely to want a knowledge garden are not the people most likely to enjoy a command line. The interesting problem turned out to be an honest one: a clipped page is not an original, it’s somebody’s HTML run through a readability parser. So why does it belong in a library whose one rule is originals, unaltered? Because for a web page there is no original you can keep — it lives on someone else’s server and can be rewritten or deleted tomorrow. That’s a good reason to keep it and a terrible reason to pretend it’s something it isn’t. Which makes the frontmatter non-negotiable: a clipping with no source URL and no date is worse than no clipping, because it’s a wall of text that looks like a source, cites nothing, and cannot be checked. The last honest limit is the one that will bite hardest, and it isn’t technical: one-click capture makes hoarding effortless, and the bottleneck was never saving things. A library of unread clippings is a to-do list wearing a costume.

Everything We Built to Watch Our Work Caught Us

Our website mirror ran its routine sync and quietly deleted 1,014 of its 1,466 files. The diff looked wrong, so we looked. It was wrong. The sync removes files that no longer exist upstream, and it guarded that removal on a variable named complete, which turned out to mean “I reached the last page of this query” rather than “I saw the whole library.” An incremental sync only asks for what changed, so its last page held seven records — and every other file got filed as an orphan and deleted. The comment directly above the check described the correct rule. The code underneath never implemented it. Nothing was lost, and that is not luck — the mirror lives in version control precisely so a bad delete is recoverable and visible instead of silent. The repaired check was proved three ways, including a planted decoy to confirm the genuine cleanup still catches real orphans, because this prune earns its keep and needed narrowing rather than switching off.

The staleness alarm — the one whose whole job is noticing when our own writing goes out of date — failed twice in two days while two pages sat there asserting we didn’t own books we’d just filed. The first excuse was a bad threshold: it watched for a big change in how many links a concept had, and gaining one book moves that number by less than one percent. But a synthesis isn’t refuted by volume; it’s refuted by a single specific asset arriving to contradict a single specific sentence. The second excuse was worse and more interesting: every ingest regenerates the wiki, and regenerating reset the baseline it was measuring against, so the alarm could never sound. It wasn’t miscalibrated, it was structurally incapable of firing.

Our nightly backup fired mid-edit and committed a half-finished change under the message “Daily backup snapshot” — five newly filed papers and a code fix, but not the paragraph rewrite that made them make sense. No data lost, but the history now has real work signed by a cron job. An automated snapshot is a safety net, not an author. Narrowing it created a new hazard, so the run now ends by naming what it left behind, and both versions are told that line is information, not an error.

Then we built a check for the page that describes all of it, because The Knowledge System page was wrong three separate times in a single day — “no database” after we installed a search index, a claim about our reading that turned out to be a slur on it, and a description of the mirror as our own published words after 300 pages of Helen Edgar’s writing moved in. A page that describes a living system will always drift, because the system keeps moving and the page doesn’t. The new drift check counts eleven things the page asserts and refuses to pass when any no longer match. It found three more within a minute of existing. Two design decisions carry the value: hedged claims get a small tolerance and bare numbers get none, because a checker that cries drift over a deliberate “~10,700” is a checker everybody turns off by Thursday. And Helen’s page count is measured as every mirrored domain that isn’t ours, never as her two domains by name — hardcoding the pair would rebuild the exact failure the check exists to catch. And it says plainly, in its own opening lines, that it checks counts and not arguments. The “no database” error was prose; no counter would ever have caught it. A clean run means the countable facts agree — it does not mean the page is honest, and it says so rather than letting a green tick imply it.

The best catch was the funniest. The tool whose entire job is clearing the backlog of files with unhelpful names still had “33 opaque” written in its own instructions hours after clearing that backlog to zero. A maintenance tool made stale by its own maintenance run, still describing a pile it had already finished.

And our search index turned out to be lying by omission: the embedding step exits successfully with work remaining. Twice in one day it stopped at an internal batch limit, returned success, and left 1,836 and then 2,001 documents unembedded while the rebuild printed All stages clean. That failure is invisible from outside the machine — keyword search stays complete and correct, so nothing looks broken, while vector search silently cannot return chunks that were never embedded. It lands on the one claim the spider exists to make carefully. It now reads the pending count and refuses to call a run clean while any remain. We keep learning the same lesson in new costumes. A check nobody has watched fire is not a check, and a description nobody re-measures is just a story we liked.

Rooms With Doors

The Design Field Guide was doing double duty — an introduction to our design method and a hub to our design resources — and the obvious fix was to move the method somewhere new. That fix was wrong, and catching it was the useful part: the method already had a home. The Stimpunks Design Method carries forty subpages. Every poster, every step of the Practice Loop, every one of the Five Questions already lived there. A new URL wouldn’t have cured the duplication — it would have minted a third copy and set it competing with the real one for the exact search it should own. So the page went from 386 lines to 37. The second question was better than the first: should The GRADE EtD Framework move under design too? No — read end to end, it isn’t a design page at all. It’s nine criterion sections indicting mainstream autism intervention and a framework for funders. It had been shelved under design since April, which meant the readers most likely to want it had no path to it. It’s part five of our research series now, landing after the indictment and after what’s possible.

“How We Tell Our Stories” was 1,459 lines holding a reference manual, four separate essays, a design system, and a title that promised none of them. That mismatch is the whole story. A page whose name says essay while its bulk is a manual cannot tell you what belongs in it — so for years, everything with nowhere else to go went there, and nothing could be argued out. It’s three pages now: the short practical version, Page Conventions as the manual it always was, and Rolling Our Own as the essay underneath — why this site is built from scavenged parts and other people’s words, why it’s a garden rather than a stream, and why some of it is deliberately silly. Splitting it turned up three byte-identical duplicates against the House Style Guide. None was flagged by any tool we have; all three were found by a person reading. And each pair had drifted independently, so in one case the better wording was in the copy about to be deleted — when you resolve a duplicate by keeping one side, diff before deleting, because the other side has usually become right about something.

Two more standards came out of the style guide, because they aren’t writing rules. Form Accessibility is the minimum standard for any form we use, and its audience is whoever builds the form, who has no reason on earth to open a style guide. Meeting Your Needs is that standard made real: a request form with no justification field, because asking for what you need should not require building a case for deserving it. Accessible Typography publishes our own type system for the first time — Montserrat at an 18px baseline for hyperlexic reading, Atkinson Hyperlegible for character disambiguation, Lexend, Noto Sans, Roboto as the deliberately ordinary baseline, Literata for a slower mood, and Playpen Sans and OpenDyslexic because people ask for them and preference is legitimate. The caveat travelled with it. The research on dyslexia fonts is mixed. What is supported is larger text, generous spacing, and letting the reader choose — a 2022 study found the font a person reads fastest in varies from person to person, that most of us aren’t reading in ours, and that speed and preference don’t reliably agree. So the fonts are offered for agency, never as clinical claims. Lily Pads documents our signature grouping pattern, including five rules written down for the first time — the one worth keeping is that the six tints are decorative, not semantic. No colour means anything, which is a real accessibility property that was recorded nowhere. And Designing for Memory gathers what we take from Lynne Kelly’s Memory Craft, including the contradiction every previous version quoted straight past: Kelly argues for loud pages, wild and grotesque and decorated until they shout. Our palette is deliberately quiet, warm, and biophilic. She’s writing about private memory work seen by nobody; we’re building a public site read by people with sensory sensitivities, migraine, and overload, where a loud page isn’t memorable, it’s unusable. So we take the structure and decline the volume.

Ryan also deleted the hand-kept subpage lists from the Field Guide index, because a hand-maintained list of a growing tree is a list that lies. What was left was eleven headings and eleven stubs — “Fundraising stuff.”, “Editorial team stuff.”, “Belong.” — a table of contents pretending to be a welcome. Every section has real prose now, with fifty-eight pages named inside it, surfacing things buried two levels down: ear-reading, Meeting Your Needs, the interaction badges that let you say how much socialising you have in you without explaining yourself to strangers, and Cope Is Not an Insult, which we think is the single most useful page we host. Orienting and Welcome Aboard got intros too, and the onboarding one now says two things at the door that used to be buried hundreds of lines down: nobody will ask you to explain, justify, or disclose anything to be here, and we will not hand you every key on day one. Meeting that as a surprise, deep inside a page that opened with warmth, is how a welcome turns into a bait-and-switch.

Then we counted our own links and found 29 glossary links, cited across 41 published pages, that return 404 — terms we send readers to in good faith and never wrote. Why nobody caught it is the useful half: our own link checker carried /glossary/ on an exemption list from back when that part of the site had no local copy, so the one prefix holding the links most worth checking was precisely the one it skipped, and it exited 0 every single time. An exemption that was correct when it was written becomes invisible the moment it stops being true. All 29 are drafted, and the recommendation is publish six and redirect two, which is the part worth keeping. Deficit Ideology already carries Gorski at length; Flow already carries the hyperfocus research. A second, thinner entry beside either one fills no gap — it makes the glossary worse, because two near-identical entries mean a reader lands on the weaker one about half the time. A missing page is not automatically a page that should be written.

Access Is Infrastructure, Including in the Footnotes

The field labels on our join form — and on every other form across the site — had never learned about Dark Mode. They sat near-black on a dark background at 1.4:1, which isn’t so much low contrast as invisible. Fixed, along with the neighbors that were just as bad: descriptions at 1.3:1, and the required-field asterisk and error text at 3.2:1 — a failure that lands hardest exactly when someone has made a mistake and needs to be told what it was. Labels read at 5.6:1 now, descriptions at 4.8:1, errors at 6.2:1. The lesson worth keeping is that we never wrote a rule against the plugin’s markup: it publishes its colors as CSS variables, so we re-pointed the variables and every rule that reads them followed. Dark Mode isn’t a coat of paint; it’s a promise the whole page has to keep — and a form nobody can read is a door that doesn’t open.

We added an ear-reading guide walking readers through turning on free text-to-speech in their own browser or device, and recommending the Read Aloud extension. We retired our old built-in player years ago: it was costly to maintain and, worse, forced a single voice on everyone. The new approach hands readers their own tools instead — the same defer-to-the-reader principle behind why we don’t use accessibility overlays.

We also used to translate this site with a plugin. It was slow, it cost money, and keeping every page current in every language was work that never ended — so we retired it and said nothing, which left anyone who doesn’t read English with no path at all. Now there’s Read This Site in Your Language. The reason it needed writing is the part worth keeping: browsers only offer to translate when the page’s language disagrees with the browser’s, which quietly excludes everyone reading on a shared, school, library, or work computer set to someone else’s language — the people most likely to need it, structurally guaranteed never to be asked. Two sections we didn’t plan and won’t cut. Privacy, because some browsers send the page text to a translation company while Firefox translates on your own device, and someone reading about their own diagnosis on an employer’s laptop deserves to know which is which. And an honest caveat, because machine translation is good at ordinary prose and unreliable at ours — neurodivergent, stimming, masking, and monotropism have no settled equivalent in most languages, so the machine reaches for the nearest clinical word and turns an identity back into a disorder.

That gave the Accessibility Field Guide a reason to stop being a bare list of links. It’s a landing page now with a single thesis: your browser and your operating system already ship an accessibility toolkit, and most people never find it. The Aa Reading panel gained a second exit to match — Prefer to listen? still goes straight to text-to-speech, and More ways to read this page opens the whole guide. One small refusal is the part worth keeping. The ordinary way to mark a second link as secondary is to shrink it or fade it, and we did neither. Dimming text is exactly the low-contrast failure we spent a day undoing on our forms this week, and a panel built for people adjusting how they read is the last place on the site to make text harder to see.

Then the Design Field Guide went from five pages to nine in a day, and every one started as something we already believed. Designing for Motion exists because motion was the one access question that section never answered — prefers-reduced-motion returned zero hits across every mirrored page while our overlays page had been telling readers for years that we respect it. The argument is that motion is a threshold rather than a rule, and the threshold moves. The same person here gets vertigo from parallax and also scrolls our front page up and down for an hour to regulate — stim-watching at page scale, stimming as regulation rather than noise. Our own Sensory Fluid page had already said the physiology out loud. What dissolves the contradiction is asking who holds the throttle: scrolling is reader-driven, animation and autoplay and parallax are site-driven at a rate you cannot change. The difference is not how much the pixels move; it’s who decided. Then we counted. We believed one animation was left on this site. There were 80, across 41 pages — quotations, song lyrics, photographs, video embeds, and a button asking people to crowdfund a member of our community. Not one honored the reader’s motion setting. The worse finding is that every animated block hid its own content behind opacity: 0 and waited for plugin JavaScript to reveal it, so when that script didn’t arrive the content simply was not there. Some had quietly stopped working across upgrades and nobody noticed, because a thing that does not appear looks exactly like a thing that was never there. The Website Specification marks reduced motion required, not recommended, and we’d been failing it while believing otherwise. “Sparingly” is not a number. Eighty is a number.

Controls That Defer turns a single clause on our overlays page into a test with three questions, including the one nobody asks: what happens when our code fails? A control that defers and breaks leaves the reader in their own settings; a layer that overrides and breaks strands them in ours. Ask what your control does when it is broken, not when it works. It also sets out the three moves in order, because most teams start at the wrong one: build access into the structure, then point at the switch the reader already owns, and only then build a control. Rhythm and Pause Points moves ten kilobytes of purely spatial argument out of a writing-style page and into design, organised around the line that carries it — a page expresses time through space. And we argued with our own sources: webtoon spacing is largely a tool for controlling a reader, building suspense, timing a reveal. That’s fiction craft. A webtoon withholds; we front-load. The technique transfers, the suspense does not. Zoom and Reflow finally answers the third access question that section had been asking since March. Measured at 320 CSS pixels — the WCAG 1.4.10 target — prose reflows perfectly, sixteen of nineteen code blocks on one page did not, and on monotropism three third-party embeds were silently clipped, unreachable and producing no symptom whatsoever. The correction we nearly published is the most useful thing on the page. The browser reported 1,356 pixels of overflow. Asked to actually scroll as a reader would, it moved 76 and stopped — the number was wrong by a factor of thirteen, and we had a whole paragraph written around it. Measure what the reader can do, not what the document reports. Then the fix we shipped broke something else, and every code block on the site went full-bleed until Ryan spotted it by eye. That went onto the page too, because the alternative was a quiet patch and a guide that looks more competent than it is. We would rather publish the failing numbers than a page claiming we pass. A guide that only describes the parts that already work is marketing.

A Shared Mark Is Not a Shared Byline

Autistic Realms and More Realms are now mirrored locally — 300 items of Helen Edgar’s writing searchable offline alongside ours, with her enthusiastic permission. The near-miss is the part worth writing down. WordPress publishes a directory of what a site contains, and the obvious way to build this is to mirror whatever it lists. We were about to. Then we checked it against a site we could count by hand — and the directory for Autistic Realms does not mention that the site has posts. It has 167 of them. A tool built the obvious way would have copied 70 pages, skipped every single post, and printed a cheerful summary. It would have looked like it worked, for months. A tool that reports success over the corpus it could not see is worse than a tool that crashes.

The second half of the job was a boundary, not a feature. Our mirror folder used to mean our own published words, and four separate places said so — including the Knowledge Garden Spider, which treats a hit there as the strongest source we have. All four went quietly false the moment Helen’s writing landed. Left alone, our own tool would have started quoting a colleague’s work as a Stimpunks position, fluently and with a citation, which is a worse failure than having no source at all. So the folder is organised by domain now, and the domain is the attribution: ours is ours, hers is hers, attribute by name, never fold it into we say. And it’s written down that mirroring someone’s site is a favour they granted rather than a licence — if she asks us to stop, we delete the folder. Her monotropism pages now outrank ours in our own search, which is exactly right — she runs monotropism.org.

★ stuff is now formally a Stimpunks Foundation × More Realms collaboration. More Realms has been one of the collection’s sources and most of its weather for months, and it was credited the way you credit a citation rather than a co-publisher. Both organizations now appear together on all 49 pages, equal weight. What did not change is per-piece credit, and that is the actual decision. Three zines are Helen’s outright and in her own voice. A shared mark is not a shared byline — the same rule we wrote into our own house when her sites came into our mirror: the attribution travels with the work, not with the folder it lands in.

Cosmic Connections is Helen’s coinage, it had been sitting on the front page as a tagline, and it’s now a piece explaining how the three sites actually work together — with receipts in both directions, because a claim of mutual influence with no ledger is just a nice feeling. Two histories were already inside the phrase and neither had been named: James Burke’s Connections (BBC, 1978), where the house method of following a thread across centuries comes from, and Carl Sagan’s The Cosmic Connection (1973), where we are made of star-stuff first reached print. And Kirby Ferguson is credited by name for Everything Is a Remix, a phrase the house had been using as a section heading without ever saying whose it was. An uncredited phrase is not a phrase you coined, it is one you borrowed and forgot.

Two new pieces shipped alongside it. Zine No. 30, Always Migrating, a guest zine by Helen: a bird crossing an ocean on a compass sense that may run on a pair of electrons whose spins stay linked, the Milky Way–Andromeda collision now closer to a coin-flip than the settled fact it was taught as, and the turn inward — the tug that arrives before thought is monotropism’s gravity. And Field Guide No. 8, The Ways of Migrating, fifteen entries breaking the word open, because almost nothing does the one shape everybody pictures. A relay no individual finishes. The largest migration on Earth going a few hundred metres straight up, every night. The centre of it is partial migration — populations holding both migrants and residents, which the literature calls the commonest form. Variation within a population is not the anomaly; it is the mechanism. The care note is said twice, in the intro band and again in the coda: this is migration as a way of being, and it is not a frame for human displacement. People forced from home by war, persecution, poverty, or a heating planet are not following a tug; they are surviving decisions other people made, and reading that as natural or seasonal would launder it.

The collection is searchable now — one record per spread and per entry, running entirely in your browser with no server, no analytics, and nothing you type leaving your device. It also got a public changelog backfilled to day one across all 121 commits since launch, with the errors named, and a written design system that publishes the value that fails as well as the ones that pass. Same position we took on zoom and reflow. Then the fact-check, which is the one worth keeping. The palette was remembered as derived from Solarized, and it is not: not one of Solarized’s sixteen values appears anywhere in the collection. Credit now goes to Solarized’s method and not its colors. The reason the memory was so confident is that it was true somewhere else — our own palette really is Solarized down to the hexes. A belief that is correct about one project will happily walk into another one and stay.

The rest of that work was five failures that produced no visible symptom. 44 of 46 pages printed blank or illegible, so a poem printed an empty sheet. The Manifesto had been publishing with its own title invisible, nine cyclic custom properties resolving to nothing. Most visitors had never actually seen Atkinson Hyperlegible — the font request asked for weights the release doesn’t publish, so the whole stylesheet came back HTTP 400 with no font rules at all and every visitor silently fell back to system fonts. It went unnoticed for the dullest possible reason: the font is installed on the machine where the pages get built, so it looked right there and nowhere else. Every field guide had been printing without its field notes — 14,040 words across eight guides had never once reached a sheet. And the new search index turned out to be indexing 12% of the About page, because the check asked did we get at least three records? rather than how much of the page did we get? Every one of those is the same shape as the eighty animations. A thing that does not appear looks exactly like a thing that was never there — so you cannot find these by reading the page. You find them by printing it, requesting it, and counting what you actually got rather than how many times you got something.

What a Week Costs

The changelog you’re reading records what shipped. It’s written for readers, in public-facing prose, and it has no room at all for what a week cost — what went wrong, what got abandoned, what tomorrow is carrying. So there’s a second habit alongside it now: a daily and weekly reflection, kept in Ryan’s own notes rather than published. We filled in the first one by reconstructing seven days from the commit history, this changelog, and our website mirror — and the reconstruction is the finding. Every hour of work was recoverable. Almost none of the experience was. Anything that produced a commit survived. Anything that didn’t is simply gone: the dead end that made the eventual fix obvious, the ten minutes of checking that caught a tool about to mirror 70 pages and skip 167 posts, how it actually felt on Wednesday when three of our own published claims turned out to be false. A commit log is documentary evidence of iteration — the argument we made publicly this week — but it documents the work, not the worker.

Two small decisions carry most of the weight. The template ran Monday-first while our week starts Sunday, which quietly filed the biggest day of the week at the bottom of the page, underneath Saturday. A container that disagrees with the thing it holds will keep producing that error every week until someone fixes the container, not the entry. And the personal sections are Ryan’s to write, not a machine’s. A machine can source what shipped from a git log; it cannot source how a Tuesday felt, and a confident guess at somebody’s gratitude is worse than an empty heading. There is also deliberately no tool for this yet. Our own rule is that anything done twice should become one, and we didn’t build it. Automating a practice before you know its shape is how you get a beautifully engineered container for a habit that never happened. This is a house with no streaks to keep and no dashboard that turns red, because urgency cues are hostile design in a system meant to hold knowledge for people whose capacity varies. Toolbelt Theory puts it plainly: people assemble their own tools, for their own capabilities at that time — and those change. The rhythm has to exist before the tool that carries it.

Week 32 Changelog

Published “Stimpunks.org Changelog for Week 31 2026: From the Map to the Directory, From the Warning to the Welcome.”
Published “Stimpunks Monthly Newsletter — July 2026,” pulling Weeks 27 through 31 into one arc: June named the method, and July made the work walkable. The synthesis names a distinction the individual weeks only imply — we spent the month making our systems legible to readers, citers, machines, and funders, while publishing a word for the right not to be, relational opacity. Institutions owe you transparency; you don’t owe them decoding.
Published “Campfire Learn Together: Code for the People” for our August 9 watch-along of Bao Nguyen’s documentary about open source, the open web, and what platform decay costs the people who live inside those platforms. The post says plainly what we usually leave implicit: our philosophy, processes, and practices came out of that work rather than out of disability services or education. Default to open, everyone can contribute, communication is oxygen, our communication stack, intermittent collaboration, self-organizing teams, transparent competency networks, and even the team name Flow Patrol are open source practice wearing different clothes. So it adds the argument the film never quite makes — distributed, text-first, open-by-default work is an access practice, and the open web is disability infrastructure nobody labeled as such — then argues back in three places, because we do not applaud things we helped build.
Published “Infodumplings: The Systems That Get Us Through the Day” for Thursday’s Infodumplings, where the Knowledge System gets ten minutes and then the floor opens for the actual subject: what systems have you created to help you work, live, and cope? The frame is Toolbelt Theory, and the important words in Ira David Socol’s definition are their own. So the post refuses the two failure modes this kind of night has: it is not a best-practices session, and it is not productivity. The argument we care about is that bus factor 1 is a design failure at home too — the freezer meal, the written-down phone tree, the note left for a self who won’t be well enough to reconstruct it. A sticky note in the right place is a system. All 37 links in the published post resolve, no redirects.
Published The Stimpunks Knowledge System in the IT Field Guide — motivation, architecture, and implementation, written for whoever maintains it next, including future us. Four layers, the rules that hold them, the small tools that maintain it, the 3-2-1 backup channels, and a schedule with no streak to keep and no dashboard turning red. Documents what it isn’t: we’re wary of techno-solutionism and technoableism, and a folder structure fixed nothing structural. Honest limits admits what isn’t fixed. Steal this gives away the portable ideas.
Published The Stimpunks Knowledge System as Curriculum in the Learning Space — two arguments: cultivate a knowledge garden as a constructionist project that never ships and treats monotropic depth as an asset, and teach change management and revision control, because a commit history is documentary evidence of iteration and version control is psychological safety built into the tooling. The two meet on one sentence — revision control is what makes AI collaboration safe — because a diff is how a human supervises a machine. Own the garden, don’t rent it: the LMS is the anti-garden.
Published Start Your Own Knowledge Garden and the knowledge-garden-starter template. Five steps that degrade gracefully, with step one needing no software at all, plus a terminal walk-through written for someone who has never opened one. Steal this now sends people somewhere instead of listing principles and stopping.
Published Filling the Garden — how things actually get into a knowledge garden, with the Obsidian Web Clipper as an intake pipe and not one terminal command anywhere in it. A clipping with no source URL and no date is worse than no clipping. A library of unread clippings is a to-do list wearing a costume.
Published Read This Site in Your Language and rebuilt the Accessibility Field Guide as a landing page with a single thesis: your browser and your operating system already ship an accessibility toolkit, and most people never find it. The translate page covers Chrome, Edge, Firefox, and Safari on desktop and phones, names which browsers send your page text to a translation company, and carries an honest caveat — neurodivergent, stimming, masking, and monotropism have no settled equivalent in most languages, so the machine reaches for the nearest clinical word and turns an identity back into a disorder.
Published an ear-reading guide walking readers through turning on free text-to-speech in their own browser or device, and recommending the Read Aloud extension. We retired our old built-in player years ago because it forced a single voice on everyone; this hands readers their own tools instead. The Aa Reading panel now ends with two links rather than one — Prefer to listen? and More ways to read this page — with the hierarchy carried by word order rather than by dimming, because dimming text is the exact low-contrast failure we spent a day undoing on our forms.
Made our forms readable in the dark. Field labels on the join form and every other form sat at 1.4:1 in Dark Mode — not so much low contrast as invisible. Descriptions were at 1.3:1; the required-field asterisk and error text at 3.2:1, which fails hardest exactly when someone has made a mistake. Labels now read at 5.6:1, descriptions at 4.8:1, errors at 6.2:1. Fixed by re-pointing the plugin’s own CSS variables rather than fighting its selectors — we tried the overrides first and its field rules held. All written down in our light/dark maintainer notes, contrast math and dead ends included.
Split The Design Field Guide into a hub and a method, and re-shelved GRADE EtD. The page went from 386 lines to 37, because The Stimpunks Design Method already carried all of it and a new URL would have minted a third copy competing with the real one. GRADE EtD is part five of the research series now, after the indictment and what’s possible. Onboard Translator moved to Operations the same afternoon.
Four new Design Field Guide pages in a day, taking it from five to nine: Designing for Motion, Controls That Defer, Rhythm and Pause Points, and Zoom and Reflow. Motion found 80 animations across 41 pages where we believed there was one, none honoring the reader’s setting, every one hiding its own content behind opacity: 0 — required, not recommended, and we’d been failing it while believing otherwise. Zoom measured at 320 CSS pixels (WCAG 1.4.10) and found sixteen of nineteen code blocks failing and three embeds silently clipped — then the browser’s own overflow number turned out to be wrong by a factor of thirteen, and the fix we shipped broke every code block on the site until Ryan caught it by eye. Both failures are on the page.
Published Form Accessibility and Accessible Typography as Design pages, lifted out of the House Style Guide because neither is a writing rule. Typography publishes our eight reading fonts and four text sizes for the first time — Montserrat at an 18px baseline for hyperlexic reading, Atkinson Hyperlegible for character disambiguation, and Playpen Sans and OpenDyslexic because preference is legitimate — with the mixed evidence on dyslexia fonts stated plainly, and sizes that scale up from your browser’s default rather than replacing it.
Published Lily Pads and Designing for Memory. Lily Pads had been documented in two places and split in two halves 730 lines apart; the new page holds the practice, including five rules written down for the first time — the six tints are decorative, not semantic. Memory Craft names the contradiction every previous version quoted past: Lynne Kelly argues for loud pages, our palette is deliberately quiet, and she’s writing about private memory work while we’re building a public site where a loud page isn’t memorable, it’s unusable. We take the structure and decline the volume.
Split “How We Tell Our Stories” into three pages that each say what they are — the short practical version, Page Conventions as the manual it always was, and Rolling Our Own as the essay underneath. 1,459 lines holding a manual, four essays, and a design system under a title that promised none of them. Turned up three byte-identical duplicates against the House Style Guide, none flagged by any tool, all found by a person reading — and each pair had drifted, so in one case the better wording was in the copy about to be deleted.
Gave every Field Guide section a real explainer, replacing eleven stubs — “Fundraising stuff.”, “Belong.” — with prose naming fifty-eight pages inside it, surfacing things buried two levels down. Each section’s own index is now the authoritative list for its tree, so the front door never has to be kept in sync by hand again. Orienting and Welcome Aboard got intros that say two things at the door instead of hundreds of lines down: nobody will ask you to explain, justify, or disclose anything to be here, and we won’t hand you every key on day one.
Put the whole website into the garden. All of stimpunks.org now lives locally as Markdown — every post, page, and glossary term, from the public REST API, with no login, no vendor key, and no AI in the loop. And it’s in git, so the website has a version history: a sync diff shows exactly what changed and when. Get your shit in git, including your own website. Building it turned up a pile of things worth writing down, each now a comment explaining why the workaround is there, because the dead ends are the expensive part.
Mirrored Autistic Realms and More Realms — 300 items of Helen Edgar’s writing, with her enthusiastic permission. Nothing is auto-detected, because WordPress’s own directory for Autistic Realms doesn’t mention that the site has posts. It has 167. A tool built the obvious way would have copied 70 pages, skipped every post, and printed a cheerful summary. The mirror is organised by domain now, and the domain is the attribution — attribute by name, never fold it into we say. If she asks us to stop, we delete the folder.
Taught the library to index ideas, not just files. A concept layer keyed to our published glossary — 445 terms — gathering everything we hold on each idea from every corpus at once. Matching weighs a title far above a passing mention, because 335 of our site pages “mention” monotropism, nearly all of it navigation furniture. We now extract text from most of the library, so a paper filed as 40614_2022_Article_338.pdf is findable by what it argues. Its first health check reported that we are standing on sources we do not have.
Bought the books the index asked for. Anna Stenning’s Narrating the Many Autisms (Routledge, 2023), which cites our Cavendish Space framing and attaches it to a universal claim about intermittent collaboration — universal design, not accommodation. Then the Damian Milton arc end to end: the 2011 Autscape talk, the 2012 paper where the double empathy problem was named, the 2022 ten-years-on editorial, and Emma Marsh’s monotropism explainer. Milton was cited inside 79 of our files and we’d owned nothing he wrote. An unfiled source is not an unread one — the real gap was never comprehension but durability: no page number, no exact quotation, nothing to hand a colleague when the URL rots.
Made the research library searchable with qmd — keyword search, vector search, and a reranker, running entirely on this machine with no vendor, no network, and no API key. “Why does being interrupted mid-task hurt so much” finds nothing by grep and now returns our own Tendril Theory. A person in distress does not arrive knowing our vocabulary. Also credited the source of the architecture: the raw-plus-index split is Andrej Karpathy’s llm-wiki, and the page now says so.
Opened seven of nine documents that were never files — a .gdoc on a Mac is a 173-byte stub holding a Drive id, so nine research assets sat indistinguishable from corruption. Two we could not open and left alone, relabelled: their Drive targets are gone, so those dangling links are the only surviving record that we held them. The fix for a dead pointer is a legible label, never a deletion. And we retracted a bad inference — we had described those documents as privately shared, based on nothing but a Drive timestamp, and stamped five public community resources with do not quote without permission. A timestamp records that a file arrived. It says nothing about why.
Indexed 4,105 pages of bookmark text as a sixth searchable collection, after a sweep reported the garden held nothing on curiosity while Susan Engel’s paper sat in our bookmarks the whole time. Coverage is 38.5%, reported as 38.5% — 4,685 bookmarks sit on hosts we skip deliberately (twitter.com alone is 4,156), 1,163 come back blocked by publishers, and 210 are too thin to keep. Every refusal is recorded with its reason. A coverage number you cannot inspect is indistinguishable from a coverage number somebody invented.
Found 29 glossary links, cited across 41 published pages, that return 404. Our own link checker had /glossary/ on an exemption list from back when that part of the site had no local copy, so the one prefix holding the links most worth checking was the one it skipped — and it exited 0 every time. All 29 are drafted; the recommendation is publish six and redirect two, because a second thinner entry beside Deficit Ideology or Flow makes the glossary worse. A missing page is not automatically a page that should be written. In the opposite direction, an inbox pass cleared 37 sources — 20 new, 17 already held and deduped rather than duplicated. Not one landed in Unidentified.
Fixed the sync that deleted 1,014 of 1,466 mirrored files on a variable named complete that meant “I reached the last page of this query” rather than “I saw the whole library.” Nothing was lost, and that is not luck — the mirror lives in version control precisely so a bad delete is recoverable and visible. The repaired check was proved three ways, including a planted decoy to confirm the genuine cleanup still catches real orphans. Separately, a new prune catches the corpse a page leaves behind when it moves — same content, same id, and the dead address still sitting in the field our own rules cite from. It was built at breakfast and caught one in production by dinner, unattended.
Fixed a staleness check that was structurally incapable of firing, twice in two days, while two pages asserted we didn’t own books we’d just filed. First a bad threshold — gaining one book moves a link count by less than one percent, but a synthesis isn’t refuted by volume. Then the real cause: every ingest regenerated the wiki, which reset the baseline it measured against. It pins to the writing itself now. A check nobody has seen fire should be assumed broken.
Scoped the nightly backup to notes only, after it fired mid-edit and committed a half-finished change under “Daily backup snapshot.” An automated snapshot is a safety net, not an author. Scope had to cover what the job destroys, not just what it saves, and the run now ends by naming what it deliberately left behind.
Added a drift check for the Knowledge System page, which had been wrong three times in one day. It counts eleven things the page asserts and refuses to pass when any no longer match — and found three more within a minute of existing. Hedged claims get a tolerance and bare numbers get none; Helen’s page count is measured as every mirrored domain that isn’t ours rather than by name. It says in its own opening lines that it checks counts and not arguments, because the worst error was prose no counter would catch.
Stopped the search rebuild reporting “all stages clean” over a half-embedded index. The embedding step exits successfully with work remaining — twice in one day it stopped at an internal batch limit and left 1,836 then 2,001 documents unembedded. Keyword search stays complete and correct, so nothing looks broken, while vector search silently cannot return what was never embedded. It reads the pending count now and refuses to call the run clean while any remain.
★ stuff is now a Stimpunks Foundation × More Realms collaboration, with both organizations on all 49 pages at equal weight — and per-piece credit unchanged, because a shared mark is not a shared byline. Published Cosmic Connections with receipts in both directions, crediting James Burke and Carl Sagan for two histories already inside the phrase, and Kirby Ferguson by name for Everything Is a Remix. An uncredited phrase is not a phrase you coined, it is one you borrowed and forgot.
Two new ★ stuff pieces: Zine No. 30, Always Migrating, a guest zine by Helen Edgar, and Field Guide No. 8, The Ways of Migrating — fifteen entries breaking the word open, centred on partial migration, because variation within a population is not the anomaly; it is the mechanism. The care note is said twice: this is not a frame for human displacement.
★ stuff started showing its work — a public changelog backfilled across all 121 commits with the errors named, a written design system that publishes the contrast value that fails alongside the ones that pass, and browser-side search with no server and no analytics. Then five failures that produced no visible symptom: 44 of 46 pages printing blank, a manifesto publishing with its own title invisible, a font request returning HTTP 400 so every visitor silently fell back to system fonts, 14,040 words of field notes that had never once reached a sheet of paper, and a search index quietly covering 12% of a page while reporting a healthy record count.
Started a daily and weekly reflection practice alongside this changelog, kept in Ryan’s notes rather than published — because a changelog records what a week produced and a reflection records what it took. Reconstructing the first one from the commit history found that every hour of work was recoverable and almost none of the experience was. Deliberately no tool for it yet: automating a practice before you know its shape is how you get a beautifully engineered container for a habit that never happened.

Week 32. We built the thing that remembers for us, then published how it works, why it’s a learning practice, and how to start your own. We taught our library to index ideas instead of filenames, and it immediately told us which books we’d been arguing from without owning — so we bought them. We gave the whole thing away as a template, because a resilience practice you keep to yourself isn’t much of one. And every guard we built to watch our work caught us: a sync that deleted a thousand files, an alarm that could never sound, eighty animations where we believed there was one, and a search index that reported success over work it hadn’t done. Broken systems, not broken people. Bus factor 1 is a design failure; a pile of files is not a memory; a check nobody has watched fire is not a check; and a thing that does not appear looks exactly like a thing that was never there.