It feels really good when something we published helped someone advocate for themselves.
Week 36 runs Sunday 2026-08-30 through Saturday 2026-09-05. The day headings below carry their dates.
The companion piece is Stimpunks.org Changelog for Week 36 2026: From the Condition to the Headline, From the Binder to the Public Page, which records what shipped. This one records what it cost.
Best Accomplishments
Publishing our bylaws for the first time — all forty-two sections, marked Proposed, and up before the board votes rather than after. Around them, a conflict of interest form on an open URL over our own recommendation to hide it, what it costs to sit on our board written down including the parts you usually discover after you say yes, and a transparency log that carries the missed board meeting and the vacant Treasurer seat next to the wins.
Two Why Sheets for people who asked for them: Monotropism at Fergus Murray and Helen Edgar’s request, and Masking and Burnout for a family whose son was refused an evaluation because he is doing fine.
Theme
Conditions travel or they fall off. And our own website is not where we learn what we require of ourselves — the governing document is.
Sunday — 2026-08-30
Accomplishments
- The morning went on catching the record up, and the record had fallen two days behind. Friday and Saturday of last week were both missing from the running changelog — twenty-four commits and eleven new pieces between them — along with two empty days in the reflection log. All of it was backfilled, Start Your Own Knowledge Garden had two stale counts corrected, and the Week 35 changelog was drafted in the afternoon. A week that produced eighteen zines takes a while to say out loud.
- Two pieces, and the better decision in each was about what not to do. Already an Entity (No. 81, in More Than Human) is built on a bibliographic fact — a river as first-listed author of a law journal article, and Mother Earth acknowledged as senior author of another, both indexed, both citable, neither a metaphor. The method is inverted on purpose: running our usual register on First Law would make physics or statute the validator of another jurisprudence, so settler law is the object of study instead. Every person is attributed to a nation in a table — three nations, three bodies of law, not a chorus — and Mary Graham is quoted from her own paper rather than through Yunkaporta, because letting one Aboriginal author be the conduit for another’s thought was the specific failure to avoid. It is the one zine in the series that ends without a claim about Neurodivergent belonging, and the last spread is the reason. Then Rainbow Serpents and the Laws of Thermodynamics, an essay in Foundations that checks the physics half of Tyson Yunkaporta’s argument and finds it correct — the traffic running from his criticism to our discipline, and not the other way. It does not explain the Rainbow Serpent and does not claim Aboriginal Law is validated by thermodynamics, and it holds three infinities apart with lineage explicitly refused: they rhyme, and rhyming is not descent.
Insights and Takeaways
- Two whole days of work had gone missing from the record, and the reason is structural rather than forgetful. The Star Stuff site is built in a repository of its own, and nothing in the knowledge system watches it — so a Saturday of seventeen commits and nine new pieces produces no signal here at all: no changed files, no status line, nothing for any of the logging tools to notice. It was recovered only because somebody went and looked. And then the looking went wrong twice in one day, which is the part worth keeping. The first pass logged Friday when Saturday was the day asked for. The second time the same question was asked, the date had rolled over past midnight while the work was still going, so “yesterday” had quietly become a different day again. “Yesterday” is not a stable address when the work runs past midnight and the record lives somewhere the tools cannot see. The fix is not more diligence. It is checking the clock and the other repository’s log before answering, every time.
- An essay about a qualifier that keeps falling off sentences committed exactly that fault in its own first sentence — and no gate could have caught any of it. Rainbow Serpents opened by saying the missing qualifier was “six words long” and never naming the words. It is four: of an isolated system. The count was a flourish written without counting, which is the No. 79 fault — a manufactured number — committed in the opening line of a piece about manufactured precision. Three more came with it on the same read: the essay argued the running-down picture is misapplied and never said what happens instead; a correction that ran the wrong way, saying Yunkaporta shares an imprecision he does not share, when the imprecision was ours alone; and a closing line that, read strictly, denied that the universe’s total entropy increases. All four were found by Ryan reading the shipped page. Eight automated gates were clean throughout. A page can pass every check it has and still argue against itself in four places, and the only instrument that finds that is a person reading it after it is live.
- Fixing a heading exposed the same mistake two inches below it, and the script that fixed it lied about having finished. The artwork’s section had been titled “The dragon that is not a serpent”, sitting directly above the image with a second negation as its first sentence — three refusals before anything positive was said about the work. A settler-run essay has no business annotating an Autistic artist’s work with a disclaimer about somebody else’s cosmology, so it is now “The head does not take the tail” — the actual finding, which sends the reader into the image rather than against it — with the conflation note moved to the end and reframed as ours to manage, not hers to disclaim. And the fix uncovered something worse. The next heading read “a closed-system model of a person” while its own body said isolated — the exact conflation this essay corrects one of our own zines for, committed in a heading on this page. An audit found four loose usages of ours and checked six more that were correct. Then the real lesson: a scripted fix silently failed to save two of the four, because an assertion aborted the run before the write, and only the after-the-fact audit caught it. The script reported success. Audit the file, not the script’s success message.
- A zine was built as No. 76 against a catalogue that had moved eleven commits past it. Already an Entity was written and numbered against a working copy that was already stale — Nos. 76 through 80 had landed in the meantime, and No. 76 is a different piece entirely — so it was renumbered to 81, with membership, the chain slot, and subject overlap against all five intervening zines re-verified before it shipped. Nothing was lost, and the near miss is the finding. The catalogue is now producing pieces faster than a working copy stays true to it, which is the same shape as the prior-art sweep that missed a zine because it was one day old. When the shelf moves this fast, the check that matters is not “is this good” but “is this still where I think it is.”
Gratitude
- Kaya Oldaker, whose The Autism Infinity Dragon is the first artwork by another artist to appear on the site — released to the public domain by the artist and credited anyway. The open item it surfaced is ours to fix: stimpunks.org/together/ attributes the drawing more restrictively than she did.
- Tyson Yunkaporta, for a criticism of our physics from outside our discipline that turned out to be correct on checking — and for the sentence Rainbow Serpents gives the last substantive word to, on recognizing ourselves as processes rather than objects.
Monday — 2026-08-31
Accomplishments
- Three pieces shipped and every one of them corrected its own starting material before it went out. None of Those Miles opened on a headline that had dropped a fourteen-word condition and turned electricity into energy; Came Back Clean had a headline figure printed flat twice that is really 59% ± 15%, with its own authors saying treat it with caution; and The Play You Keep had to be re-scoped off a proposal whose spine was its own author’s earlier zine, with one candidate example dropped because a bower is courtship and therefore has a payoff — the one thing the zine’s play does not.
- Two sheets went out carrying somebody else’s theory, and getting the credit right took more work than the prose did. Monotropism is the tenth Why Sheet and Monotropism the eighth broadside, both adapted at Fergus Murray and Helen Edgar’s request. The six starting points are Fergus Murray’s, from January 2019, and they are named in three places across the two sheets because Helen asked specifically that the source be carried — while the theory itself is Dinah Murray, Mike Lesser and Wenn Lawson’s, from 2005, and the sheets say so too. Alongside them, Campfire Learn Together: The Subtle Art of Normalising Dystopia for 6 September, built on a contradiction we published two days earlier: the film argues everything taken from us was made boring on purpose, and our own page argues boring is our best defence. Same word, opposite directions, and that is the session.
Insights and Takeaways
- Seven diagrams were wrong at render size across two pieces, and the gates passed every wrong state of all of them. Three of four on None of Those Miles: a strait drawn as two stroke-only shorelines that read as an X crossing rather than a narrowing, markers merged into an axis arrowhead, and a return wire running through its own layer labels. Then three passes on Came Back Clean before the drawing was right — briefs that rendered as buckets, then hips flared the wrong way with the waistband detached — plus a figure that drew the underpants floating above the soil line, contradicting the very protocol it illustrates. Contrast, overlap and markup passed all of it, and were right to: the overlap checker measures text against text, and every one of these is a shape, a topology, or a meaning. The nine final instances are generated from one shared path so they cannot drift apart — which is the only durable answer here, because the thing that caught each wrong version was somebody looking at it.
- An open item logged in the morning was closed by the afternoon, and closing it produced the better lesson. None of Those Miles recorded that a stylesheet rule was printing diagram labels lighter than intended — an attribute selector quietly outranking a type selector fourteen lines above it, affecting 127 labels across 18 pages. Nothing was broken; both values clear AA, which is why no gate had ever complained. The fix is one rule. The verification is what took the work. Full contrast sweeps before and after came back byte-identical — and that is the safety result, not the proof, because an identical diff is equally consistent with the new CSS never applying at all. It took a separate probe under print emulation to confirm the labels had actually changed. When a change is invisible to the gate by construction, run the gate for regressions and something else entirely for confirmation.
- The largest line on a sheet about careful credit turned out to be our own paraphrase, and its subject told us so after it shipped. The Monotropism broadside face read “Fewer things at once, more mind on each.” Helen Edgar said plainly that she found it hard to read, and that “more mind doesn’t make sense to me” — the mechanism is attentional resource allocation. She is right, and the fault was ours: “more mind” is not monotropism’s vocabulary, it is a paraphrase we introduced — on a sheet whose entire point is naming whose ideas these are, in the biggest type on the page. It now reads “Fewer things at once, more attention on each.” Two alternatives were considered and refused for reasons worth keeping: “attentional resource allocation” is a label rather than a claim, and the wrong register for a line meant to be read across a room; “attention is energy” is an equation the sheet does not argue, and it would have collided with the subline directly beneath it. And adopting the fix exposed a second fault — the standfirst two centimetres above was the same claim in longer form, so the face would have stated one thing twice. A correction that only fixes the sentence it was aimed at is half a correction.
- An ingest of twenty files found a body of work the index had no shelf for, and the tell was that eleven items which are one conversation had to be split across three sections to be filed at all. Whyte on Indigenous climate change studies and the Dakota Access Pipeline, Redvers et al. on Indigenous natural law, Deborah Bird Rose on totemism, Mary Graham’s Aboriginal worldviews, Zoe Todd on the ontological turn, two Yunkaporta transcripts, a study of how autism is understood among Citizen Potawatomi Nation participants — filed as Environment, Philosophy and autism-neurodiversity respectively, because those were the shelves that existed. That fracture is not a filing error; it is the index reporting a gap in itself. The ingest flagged it rather than fixing it in the same pass, and the shelves came in a separate change: Indigenous Knowledge & Country with eleven items, and a Science shelf holding one — a 1942 paper on how few photons the eye can detect, which had been parked in the explicitly uncategorized drawer. Two decisions in it are worth keeping. The Citizen Potawatomi study moved off a 155-item shelf on findability grounds — a specific twelve-item section surfaces it; a general one buries it. And placement is not cosmetic, because first match wins: the new shelf sits ahead of autism-neurodiversity so that a future paper on Indigenous autism lands on the specific shelf rather than the general one.
Gratitude
- Fergus Murray, whose six starting points from January 2019 are the spine of both the Monotropism Why Sheet and the broadside — named in three places across the two sheets because Helen Edgar asked specifically that the source be carried.
- Dinah Murray, Mike Lesser and Wenn Lawson, whose 2005 theory both sheets exist to hand on, and whose names the sheets carry for the same reason.
Tuesday — 2026-09-01
Accomplishments
- One zine shipped, and the rest of the day went to the things underneath it rather than to the next piece. The Width of Itself went out in the early afternoon; everything after it was repair. A card that had shipped nested inside another card on the live front page, found by Ryan after the push. A landmark migration from that same morning that had pointed twenty-four pages at nothing, the front page among them, while its own new check passed all twenty-four. Five pages whose titles were styled divs, two of them returning an empty headings rotor to a screen reader. Three new gates came out of it in a single day — the seventh and eighth markup checks, and a count that shrank to zero and turned into a rule — against a bar that says a new check has to be invisible to every existing gate and have already recurred often enough to prove that remembering is not a control. All three cleared that bar, which is the uncomfortable part. The visible output of Tuesday is one page; the real output is that three classes of silent decay now fail loudly instead of shipping.
- The garden became something a teammate can open, and not one line of it earned a changelog bullet, because none of it is a thing to publish. Sixteen commits went into making the knowledge system usable by somebody who is not Ryan. The Drive mirror stopped being a copy of one person’s disk and became a team product, with a map and a ten-minute setup written for a reader who cannot run any of our skills. All six corpora now render as a searchable static site of 15,653 pages, faceted by corpus and by concept — the concept index is the reason it is not just another theme, and the build refuses to run rather than warn when that index is missing, because a site missing the most useful thing in it still looks complete. Two static sites got mirrored and a fourth read-only corpus was added, because cavendish.app publishes eleven pages while its repo holds 115 — mirroring only the site would have put the thin surface in the garden and left the substance out. A deploy skill followed whose real check is not the deploy: it fetches the live site afterwards and requires a 401, because the whole basis for shipping fifty-one files of real organizational material is that the site is gated, and that gate is a dashboard setting nothing in the repo can enforce. And the site is now the Stimpunks Knowledge Garden, which is what this project has called itself all along — “Knowledge System” was the outlier, at one mention against twenty-nine in our own rulebook. The day also built the thing that is writing this: the daily log-catcher exists because two days of Star Stuff work vanished from the record last week, and it was shaken out on a Monday that turned out to be hiding five unlogged commits behind the three already recorded.
Insights and Takeaways
- Three faults in one day were invisible to eight gates, and in every case the gate was right not to see them. A measurement unit written with the micro sign, inside a label setting
text-transform: uppercase, uppercased to Greek capital Mu — which in Space Mono is indistinguishable from a Latin M, so both cells read “0–70 MM FROM ROOT”: a factor of a thousand on the one number the zine turns on. The markup was well formed, the class was live, nothing collided, and a wrong character has exactly the same contrast as a right one. A card nested inside another card is valid HTML with balanced tags, sibling anchors and hrefs still ascending. And an empty<main>satisfies “exactly one<main>” perfectly. A check that counts a thing cannot tell you whether the thing does its job. Two of the three had a replacement available and got one — does exactly one card sit directly inside each wrapper, is the page’s own<h1>inside the landmark — and both were regression-tested against the real broken files rather than only against synthetics, which is what proved they actually fire. The unit fault has no such replacement and probably cannot have one. It is a glyph substitution performed by CSS at paint time over correct source text: the source is right, the render is wrong, and nothing that reads files can tell. The rule that came out of it is narrow on purpose — never put a cased-symbol unit inside an uppercasing container — and the broader half is the uncomfortable one. Some faults exist only at render size, and our only instrument for those is still somebody opening the page and looking at it. - The Drive mirror reported success for weeks while doing nothing, in two entirely different ways, and the second one we diagnosed wrong.
rclone syncapplies its filter file to both sides, so tightening a filter hides a file rather than deleting it: Drive still held all 33,179 objects and 1.9 GiB, and every run said it had succeeded. Excluding is not removing, and nothing anywhere said so — the run now lists what is stranded, and the prune is guarded with Drive’s own thirty-day Trash as the recovery path. Then the switch to our own OAuth client reconnected as the wrong Stimpunks account, and that failure is the one worth writing down. The token authenticates perfectly, reports an 18 TiB quota, and lists absolutely nothing — no shared drives, an empty My Drive, no permission error either time — because from that account’s point of view an empty answer is correct. Nothing anywhere says “wrong account,” and the reconnect command does not tell you which account the browser consent used. We named the account as the cause first, then talked ourselves out of it when a reconnect failed to fix it, and pivoted to OAuth scope on evidence that fit both readings equally. The reconnect had simply used the wrong account again. Ryan found it. Abandoning a correct hypothesis because the fix for it failed is how a right answer gets discarded once and then never reconsidered. The same day produced the mirror image on the other side of the repo: a freshness guard that shipped watching the wrong directory and therefore refused every possible run, comparing an output against its own input. A guard that always fires has stopped meaning anything, exactly as surely as one that never does.
Gratitude
- Nick Covington, whose “Every Dandelion is a Triumph” for Human Restoration Project prompted The Width of Itself — and whose argument the zine sharpens rather than disputes. Triumph is the right word; solution is not.
Wednesday — 2026-09-02
Accomplishments
- Five pieces went out, and by the end of the day three of them were auditing the other two. Two zines — There Have Been So Many Earths at half past twelve, The Lizard That Was Never There at half past seven, the collection’s longest at sixteen spreads — plus Glimmer Wire, Until There Were Ten written and published the day the paper appeared, and Six Principles and a Room at a quarter to midnight. The digest shipped with zero items graded VERIFIED, because six publishers refused a direct read on the compile date and every entry had to name what was read and what was not. Within nine hours one item had been verified from the primary, one had been corrected by the zine that followed it twelve hours later, one had been fact-checked into a piece that got parked, and one seedbed card had been audited and found to be missing the study’s central limit. The day’s real output is not five pages. It is a digest that started grading itself the hour it existed — and every one of those four movements was a downgrade or a correction, not a win.
- A refusal letter set the agenda, and answering it meant deciding twice what to leave out. The tenth Why Sheet and ninth broadside exist because a community member’s son was refused a special education evaluation for doing fine — high-masking, at grade level, and already burned out of school for half of fifth grade, which his mother had put in writing before the observations began. Two findings that would have made the sheet hit harder stayed off it on purpose: the camouflaging-and-suicidality research is a small correlational study of undergraduates, and the children’s burnout audit is a twenty-child conference abstract. A sheet has no room for the hedges either one needs, and a page that says “masking is linked to suicide” without them is worse than one that says nothing. Both are on the Why Sheet with their hedges visible. The family also got a response letter to send the school — not published, and not in this voice, which is the part of the day with no artifact to point at. The rest went to infrastructure the sheets depend on rather than to sheets: a print stylesheet where 211KB of CSS had carried exactly one print rule and it belonged to a search plugin, a print-or-save button that keeps the reader’s own font and size, and a footer that stopped reading as one more content section. None of that is a publication. All of it is what a printed sheet needed in order to be true.
Insights and Takeaways
- We published an accessibility bug that does not exist, and the correction cost 347 edits that were already written. An hour after No. 84 shipped we reported that a
<br>with no space beside it makes a heading announce as one run-together word — “BoneSong” — across 163 headings on 48 pages. The textContent half is true. The accessible-name half is not. Chrome inserts the whitespace itself when computing the name, in every shape we could construct: bare<br>, between two<em>s, between two<span>s, after a comma, inside a link, inside a button — measured against copies of the pages from before and after the change, “Bone Song” both times. The error was reading a tool’s output as the accessibility tree when the line it printed was textContent. This repo already has a rule for exactly that shape — aggregators are leads, not citations, go to the primary — and it had not occurred to us that the rule covers our own instruments. The sweep was written, measured, and thrown away rather than shipped: covers pixel-identical at 2x, search index byte-identical at 2,069 records, accessible names unchanged, and three index hits that had looked like proof of harm were false positives — one of them our own changelog quoting the wrong claim back at us. The count was wrong in the other direction too: re-derived by semantics rather than source pattern-matching, it is 348 across 74 pages, for a defect that is not there. What survived is one real case and one accident of the same mechanism — Underground‘s cover announces as “Under ground” precisely because Chrome adds the space. Chrome only; WebKit and Gecko untested and not claimed. - Four rounds went into a footer that was never the bug, and the finding underneath it is site-wide and still open. The band looked like it blended into the content, and it does — one Solarized step off the page, 1.15:1 at a lightness difference of ΔL* 4.8, the tightest pair in the whole palette, designed as a background highlight inside content and never as a region boundary. But the colour was the wrong suspect, and chasing it would have made things worse. What actually said “this is just another section” was a footer heading pixel-identical to every content heading and a separator identical to eight others on the same page, rendering above the band instead of as its edge. The fix that worked was full-bleed against inset, not light grey against slightly-lighter grey — edge-to-edge is a categorical difference; a slightly different grey is not. Then the real finding surfaced, and it was never a footer bug at all: two theme slots resolve to base02, and every panel built on it drops its text below AA. Callout body at 4.11:1, links inside those callouts at 4.06:1, the worst pair on the site, block quotes at 4.11 on base02 while the same quote block passes at 4.75 one step darker. All eight accents checked against both grounds: none clear 4.5:1 on base02, and exactly three clear it on base03 — which are precisely the three we use. The palette was chosen correctly; base02 is what broke it, and the link colour that works is off-palette because Solarized has nothing that fits there. And there is no size-based way out, because every font size we read turned out to be a
clamp()floor: typing “24px” sets the clamp’s maximum, the footer heading only reaches it at 762px wide, and it runs 18.7px on a phone. WCAG’s large-text threshold is 24px, so a fluid heading changes accessibility class with the viewport. The colours have to be fixed, and that is carried forward rather than closed. The contrast gate learned dark mode the same day and its first run reached by measurement what the morning had reached by eye — which is the difference between knowing a thing and being able to catch it again.
Gratitude
- Helen Edgar, twice in one day and in two different repositories: her question after a BACP training session on ethics became Six Principles and a Room, and her Supporting pupils through Autistic Burnout teacher guide — used with her permission, her words attributed and held outside our CC0 grant — is what the Masking and Burnout sheet is built on.
Thursday — 2026-09-03
Accomplishments
- Two pieces went out on Star Stuff and the day’s most consequential work was subtraction. Older Than Either of Us and A Field Guide to Bird Minds shipped inside twenty minutes of each other, and around them sit four things that did not ship. The obvious zine — the one the papers were begging for — was declined as three times already spent, and it took grepping the argument rather than the topic to find that out; the yardstick case belongs to No. 36, the nomenclature to No. 85, and FG 11’s crow card had already said many roads to a mind. Pepperberg’s best-known line stayed in field notes because it only reaches us at one remove. A volume number was refused because the source’s own reference list prints it wrongly — a number we could have copied and been wrong about, with the citation right there to blame. And then the withdrawal: two claims came out of a guide that had been live, both from a reframe table’s “actually” column, one of them a mirror-test result credited to the wrong species of elephant. That correction was pushed as its own commit eighteen minutes after the drop it came out of, rather than folded in, because a retraction hidden inside a publication is not a retraction. The day’s ledger reads two new pieces and four claims removed, and the second number is the one that took the reading.
- The accountability hub grew by four pages and almost every hour of it went on correcting our own published words rather than writing new ones. Twenty-nine commits landed here across the day and the ones that cost the most were repairs. The worst error was ours and it came from reading our own website: reading the 2022 Whistleblower Policy in the morning and finding it routing every concern to “the President,” we concluded we had no such officer and changed it to “Board Chair.” The bylaws make President and Secretary the two required offices; Chairperson is optional and we never filled it. Counsel had drafted that policy against those bylaws and was right, and it took until late evening to put President back into seven routing points across two policies and the board page — logged publicly rather than quietly renamed, and with the roster pages deliberately left unfinished because Betsy’s title is not ours to publish without asking. Frequently Requested Information had been carrying the exact figures we retracted on 9 August — “$120,000 split 65% program, 25% administration, and 10% fundraising” — on the page a funder opens first, three and a half weeks after Where the Money Went was corrected and Impact updated with it. It now runs from the books: $122,626 of 2025 expenses against $32,599 of revenue, a $90,027 deficit covered from reserves, 88% program, and the admission that the accountant is still tracing part of what is booked as investment income. Near midnight the policy that had produced the retracted 65% was finally rewritten — it hadbeen filing staff coordination and technology as administration, which is how a model got to 65% while the books say 88%. And the previous day’s Document Retention page turned out to have been built from the consultant’s sample without anyone checking the formation binder, which held counsel’s own 2022 policy all along. Both misses went into the Transparency Log as entries. Publishing a policy is an afternoon; discovering that our own site was the unreliable source for it is the whole day.
Insights and Takeaways
- Every audit we run asks how strong a result is, and neither of the two claims withdrawn today failed on strength — one failed on species and the other on format. The elephant card in Field Guide No. 11 listed self-recognition among four capacities, and the mirror study is three Asian elephants while every figure on the card is the African elephant. No previous check had thought to look, because “is this evidence strong?” and “is this evidence about the animal on the card?” are different questions and we only had a habit for the first. It is not pedantry, and that is the part worth keeping: the study rules out a general propensity to touch the marked area precisely because the female Asian elephant’s temporal gland is vestigial and nonsecretory, where male African elephants often touch their own heads owing to its secretion. Carrying the finding across species silently reintroduces the confound the species choice was there to remove. The corvid claim failed differently: it was sitting in a reframe table’s “actually” column, which is a format that strips hedges by design. A reframe table promises the reader a corrective, so a five-magpie result with two significant birds and a failed replication arrives there as a flat fact — and the honest version does not fit the column. We hedged it narrowly rather than reversing it, because the replication reproduced the mirror-versus-control difference and failed specifically on the mark test, so “debunked” would have been the same kind of error in the opposite direction. Both corrections came out of the authors’ own caution rather than our scepticism — Prior et al. say their sample is too small for a definitive estimate, and Plotnik et al. pre-empt the one-of-three objection themselves. The primary source was already carrying the qualifier; we had simply not printed it. And the repair introduced a fresh fault the gate was right to miss — a longer label crossing 1.1px into its neighbour, under an overlap checker built generous on purpose so it can only ever under-report — found by rendering the glyph and looking at it, the same instrument as every other week.
- Three of our hypotheses about one search bug died in a row, Ryan asked the right question an hour before we accepted it, and we argued him out of checking it. Search on the garden had two faults — quoted queries returning nothing, and every unquoted query quietly under-returning, monotropism at 160 results when the true count is over 1,700, so every search the team had run was incomplete without saying so. The upgrade that fixes it then killed search entirely on the live site. We blamed the Web Worker; noWorker made no difference. We blamed a CDN decompressing the wasm in transit; Netlify serves it byte-identical. And when Ryan asked whether the auth method itself was what the new version choked on, we said no and argued it from evidence: search worked live under the old version with the same password gate, therefore the gate is not the variable. That reasoning is simply invalid — both the gate and the version were variables, and “the old one tolerates it” says nothing about the new one. He was right, and the mechanism is now recorded rather than the symptom: Netlify’s site password is HTTP Basic Auth, and
fetch()does not reliably replay Basic credentials on subresource requests, so the index chunks come back unauthenticated and search never starts — while team login is a session cookie that same-origin fetches send automatically. Both answer 401 to a stranger, which is exactly why a two-state model could not tell them apart while they behaved differently for the one thing that mattered. Ryan got there by changing one variable and looking, twice. Two smaller admissions belong with it. We verified the index through the API and shipped a dead interface — never once touching the widget the team actually uses — and a silent hang is indistinguishable from a slow search, which is why the page now runs one known query on load and names its own failure after twelve seconds. And the guard we wrote at 15:11 to enforce the lesson was itself too broad by 15:18, because it had encoded “gated” as the axis when the axis is how. A rule written the moment a bug is understood is written at the moment we understand it least.
Gratitude
- Helmut Prior, Ariane Schwarz and Onur Güntürkün, and separately Joshua Plotnik, Frans de Waal and Diana Reiss — whose own printed cautions are what made today’s two withdrawals from Field Guide No. 11 possible. Neither correction required scepticism on our part; the qualifier was in the paper and we had not carried it.
Friday — 2026-09-04
Accomplishments
- Four pieces went up on Star Stuff and roughly half the day’s hours went on unwriting things we had already published. Glimmer Wire shipped at midday unable to open a single source — every publisher, wire and DOI resolver refused at the egress proxy — and the honest move was to publish the edition with the blockage above the fold and nothing in it eligible for promotion to a zine, a Glimmers entry or a card. When the network came back hours later the whole edition was re-run against the primaries, and one item upgraded while all six were corrected: author orderings, a species name, two lag figures printed as provisional, a claim about “>500 species” that appears nowhere in the paper, and a lethal-sulfide-drawdown line that the paper does not contain. The Nature Materials item was left ungraded on purpose even though the authors’ own preprint was read in full, because a preprint is the authors’ text and not the version of record — the most expensive thing in the amendment was declining an upgrade we could have taken. The rest of the day’s cost was documentation debt coming due all at once: the collection split surfaced a stylesheet comment reading “new editions are PREPENDED” that had been copied into the source of the two pages it misdescribes, and three separate counts in our own build notes were stale — one of them stale twice in the same day, because a commit landed a sixteenth collection while the edits fixing “fifteen” were still in flight. Four items were held rather than filed, including the ADHD-as-creative-advantage story, the item most on-frame for the whole site, because its sourcing was a congress presentation and its framing sat an inch from telling people to regulate themselves out of their material conditions.
- The governance push ran from three in the morning to half past nine at night, and the thing it produced that nobody asked for was a hole in our own protection. Eleven commits, five new pages, three signable forms — and writing the staff and volunteer onboarding page surfaced that Article V indemnifies directors and officers and not volunteers, so our unpaid Co-Creative Directors may have no protection at all. §5.7 permits extending it by resolution, so it became an eleventh board decision and was logged publicly the same evening rather than quietly noted; the Volunteer Agreement now says outright that Article V does not cover volunteers and that we will not claim otherwise until the Board acts, with its own footer recording that adopting the resolution reissues the form at 2.0 because a signature against 1.0 should not be read as agreement to different terms. The day’s quieter cost was finding, again, that we had built from a stranger’s template when our own documents already held the answer: the retention policy was written from a consultant’s sample and only afterwards did counsel’s 2022 version turn up in the formation binder, unread, disagreeing about who is responsible — so the 2022 text stands until the board rules, and that went in the Transparency Log beside the wins. Two pieces of after-the-fact tidying closed it out: a document-ID collision where the board and staff onboarding pages shared a base code whose suffix means version everywhere else in the scheme, which would have read as the staff page superseding the board one, and IDs added to the three signable forms — backwards until then, since the forms are the instruments people actually sign and that we retain for seven years. The 51-field conflict-of-interest form also could not be pushed through the API at all — forms are a custom post type the endpoint has no parameter for — so it was generated as an import file and carried across by hand.
Insights and Takeaways
- We spent an edition complaining that coverage garbles findings, and the two worst garbles in it were ours — both produced by reading the coverage carefully instead of reading the paper at all. One entry attacked a “reverses ageing” framing that we had constructed out of misreading the title’s own subject: “restored clearance … limits organ aging” says that restoring the clearance limits ageing, and we had read it as the impairment doing so, then built a complaint on top. The other described the ordinary Hall effect — applied perpendicular field, 1879 — and reported the paper as contradicting it, when the paper is about the anomalous Hall effect, which is magnetisation and symmetry with no applied field. That is the press release’s own conflation, reproduced in the act of objecting to it. A summary is a lossy copy of a paper, and reading it more attentively does not recover the loss — it only makes the copy feel like the original. The rest of the amendment is the same lesson at lower cost: three date windows turned out to be cooling intervals in deep-sea sediment cores rather than dated encounters, a proposed correlation printed as an observation, and a single author reported as “a NASA-funded group.” The structural conclusion is that the grade did its job and the sourcing did not. Publishing NONE VERIFIED above the fold is what made the next day’s amendment ordinary rather than embarrassing — the standing open item is still open, and the edition says so on its own face.
- Three times this week our own published website has been the wrong place to learn what we require of ourselves, and Friday made the pattern impossible to miss. Thursday it was the officer titles — reading our site’s “Board Chair” and concluding we had no President, when the bylaws name President and Secretary as the two required offices. Friday it was twice more: the gift acceptance policy rewritten against our own Certificate of Incorporation rather than the consultant’s sample, because Article NINTH is what actually limits what a private foundation may be given; and the retention policy built from a sample when counsel had already written us one in 2022 and it was sitting in the formation binder, unread. The governing documents are the source; the website is a rendering of them, and a template is somebody else’s rendering of somebody else’s. The corollary showed up the same day in the opposite direction, and it is the better half of the lesson: the Board Member Agreement had no template to copy at all — neither the consultant’s packs nor the BDO set contains one — and building it from the bylaws is what produced the part worth stealing, the list of things we refuse to ask a director for and the Part 4 obligations running back the other way, including advancing the cost of defending a proceeding before it concludes. Reading our own founding documents keeps finding real gaps rather than just correcting words: it is what turned up the volunteers who Article V does not cover.
Gratitude
- The counsel who drafted our 2022 policies against our actual bylaws — right twice this week while we were wrong from our own website, once about who a whistleblower concern routes to and once about document retention. That policy had been in the formation binder the whole time, unread. The document we already had beat the sample we went looking for.
Saturday — 2026-09-05
Accomplishments
- Five new pages went up on starstuff.earth in one Saturday, and most of the day’s actual labour was spent disagreeing with sources we had already decided to trust. The Indigenous punk rack corrected its own spine book twice at primary — a Hall of Fame year off by sixteen and a guitarist’s first name wrong — and declined three further claims rather than repeat them, including leaving the positive account of powwow drumming to the people who play it. No. 88 corrected the New York Times and Rolling Stone against Link Wray’s own account of what he pushed through which speaker, and threw away a shortlisted title because the “factor of 100,000” everyone repeats about Harold Black’s 1927 test is not in the paper it is attributed to and its primary could not be found. Which Distortion declined the Guardian’s paraphrase of a US statement because the word was not in the statement and made a substantive objection sound procedural. Three pieces, and in each the expensive part was the sentence we did not write. Adding the rack also meant recounting the shelf it joined, which is how three of our own published numbers turned out to be wrong — one of them unreproducible and disagreeing with itself in two places — so every figure was re-derived from the markup with its method printed beside it rather than corrected quietly. The day closed on two pieces of infrastructure that exist to make the site legible from outside: a What’s New page and feed generated from the cards and from git rather than maintained by hand, and a Mission page that ends in eight falsifiers a reader could check from outside — four of which have already happened and are named there.
Insights and Takeaways
- Every fault that mattered on Saturday was invisible to eight passing gates, and the two worst were found because Ryan opened the page and looked. Two cards on the new rack said “Video unavailable” — both had passed the collection’s own embed check the day they shipped, and both still pass it. oEmbed answers 200 with the correct title and channel for a video that no longer plays: it confirms identity, not existence. Sweeping all 269 embeds turned up two more that had been dead or age-gated for weeks, including one where an age gate blocks embedded playback as completely as deletion while looking perfectly healthy from outside. The other two faults were the same shape: a dark SVG panel fill sitting outside the print stylesheet’s inversion list, which would have printed every figure as an ink block with the strokes lost inside it, and two leader lines starting inside their own label’s text box — found by measuring, because the overlap checker cannot see that by design. The decision that came out of it is the one worth keeping: the new embed checker was deliberately not made a ninth gate. All eight gates are hermetic and never touch the network, and a gate that can turn red while nobody has edited the repo is one people learn to push past — so the checker is a tool you run, not a wall that cries wolf. Ryan also caught the piece arguing against itself in its own summary: Which Distortion opened by saying the Assembly voted “to correct the world map,” four screens above the sentence explaining that correct is exactly what no flat map can be. The clean version of an argument is the last place we look for the argument’s opposite, and a passing build says nothing at all about whether a page contradicts itself.
Gratitude
- Al Carroll, whose Indigenous Punk was read in full and became the spine of the eighth Sound rack. It is specific enough about dates and names to be checked, which is why we could find two things in it to correct — a vaguer book would have given us nothing to disagree with.
Looking Back
Meaningful Moments
Writing Masking and Burnout for a family whose son was refused an evaluation for doing fine — and sending them a letter for the school that will never be published and has no artifact to point at. The part of the week with nothing to show for it is the part that had somebody waiting on it.
Life is Teaching Me
Our own published website is not always a source. Three times this week it was the wrong place to learn what we require of ourselves — the officer titles, the gift acceptance limits, the retention policy — and every time the answer was already sitting in a governing document nobody had read. Go to the binder, not to the rendering of it.
Habit Insights
The changelog and reflection habits hold. The habit doing the most work, though, is not an automated one: nearly every fault that mattered this week was found by opening the shipped page and reading it. Eight gates stayed clean through four wrong arguments in Rainbow Serpents, a card nested inside another card, two dead video embeds, and a piece contradicting itself in its own summary. The gates catch what they count. They do not read.
Looking Forward
One Change I Can Make Next Week
Close the corrections we have already found. Three were still open when the week closed, and the first of them still is: stimpunks.org/together/ licenses Kaya Oldaker’s dragon CC BY-NC when she released it to the public domain, the base02 contrast failures are carried forward rather than fixed, and the parked fibromyalgia piece has its unpark condition written down and unmet. This was last week’s one change too, which is the finding.
I’m Looking Forward To
The board voting on the bylaws this month — the first vote we will take on a governing document our community could read beforehand.