The accession register

A trail and a control are different things, and both stay. .qe-crumbs says where you are; it sits inside the masthead, below the tray, and on a sheet it is the third thing down. A reader in the middle of nine thousand words wants the other one. Ryan kept looking in the top bar and finding nothing there, which is the whole of the case.

It carries a word, and the unlabelled licence next door does not reach it. The fox and the hedgehog are allowed to be a picture alone because nothing is withheld — that page is in every footer, in the drawers menu and in the finding aid. The home page is the destination a reader is most likely to want and least likely to hunt for, so hiding it behind a drawing would have been the original complaint arriving again wearing a mark.

At 320px the tray is three lines now rather than two, and that was measured rather than discovered later: 133px of tray against 91px, no target under 44×44, no two hit areas touching, and scrollWidth equal to the viewport at 320, 375, 414, 768 and 1280. Both overlay panels hang off the tray rather than off their own summaries, so neither moved — which is the fault the drawers control found the hard way, and the reason that rule exists.

Four marks were drawn, rasterised at 16, 20 and 27px, and then magnified off a canvas rather than redrawn large. A vector redrawn large is a lie about what a reader gets — the lesson favicon.svg paid for twice and the pocket lens a third time. What the pixels said was that a pitched roof is the only silhouette that reads as home at 16px, and that everything else needed the word to carry the entire meaning.

The site’s own mark was tried first and lost. Home as the logo is the oldest convention on the web, and the pressed leaf does not survive the trip: it has a paper ground and a 64px canvas on the icon, and at 20px in moss beside the lens it is a dark blob with less character than the instrument next to it. Its midrib in paper is a white gash that splits the leaf into two lobes and reads as a pea pod. And --qe-rust goes pale pink in the drawer, so the pin that makes the favicon legible becomes a pale bulb at the base.

A chest of drawers was refused without drawing it. The cabinet itself is a drawer in the menu two controls along. One picture, two destinations.

So was a mended cup, and that one is a house rule rather than a matter of taste. Kintsugi is already spent here: --qe-marigold marks a re-determination in this register and the join in a restored attribution, and a seam requires an actual repair. A gold-mended mark in the tray would have asserted a correction that never happened, an arm’s length from the real mends on every provenance line on the site.

The wabi-sabi that survived is the ridge, set off centre. The apex sits at x 11.2 against a body centred on 12, so the two slopes differ and the bays run 4.8, 5.4 and 4.6. It is visible at 20px, it is the same wonk axis already running on this site’s type and on the rail’s ticks, and it claims nothing — which is the only licence ornament gets here.

Two versions were drawn wrong first. A full glazing grid with a transom and one missing pane goes muddy by 20px, and the missing pane does not read as a void at all: it reads as a dark window, and sitting off centre it drags the whole mark sideways. Then a shoot drawn up through the open door — the botanical cue this mark otherwise lacks — read as a small person standing in the doorway at the size it ships at. That is the lilac panicle’s lesson in a fourth costume: do not put one small mark where the eye is expecting a figure.

Installed, this site would have had a generic browser icon and its page title. A manifest is the declarative half of the thing the acronym covers — it runs nothing, caches nothing and makes nothing work offline; it says what the mark is, where the window opens, and what colour the chrome around it should be. The other half is a service worker, and there is still none here.

display is minimal-ui on purpose, and that is the one field in it that is an argument rather than a fact. standalone hides the address bar, and this is a site whose whole practice is that /on-being-ill is a real address a reader can copy, share and check us against. An installed copy that hides the bar the address sits in would be the cabinet quietly withdrawing the thing the ledger exists to make possible. minimal-ui is still installable in Chromium and keeps the address in view.

start_url is a bare /, against the spec’s own example. That example appends a tracking parameter so installed launches can be counted separately; there is nothing here that could read it, /privacy says so, and it would mint a second address for the home page on a site whose redirects exist to ensure every page has exactly one.

short_name is the site’s name and not a nickname. A home screen may truncate Queering Earth, and being truncated by somebody’s launcher is not the same as us keeping a second, shorter name for the site — which is the drift this register refuses everywhere else.

Android applies its own mask and guarantees only the central 80% of the square. The mark is a pressed leaf pinned with the wordmark’s rust dot, and at the scale every other icon here is drawn at, the drawing reaches 0.484 of the square from its centre against a safe radius of 0.400 — measured out of the finished PNG, not calculated. The pin is the part that goes. A leaf with no pin is a leaf; the pin is what says specimen, and it would have been cropped off on every Android home screen with nothing anywhere looking wrong.

So the maskable one is a separate render at its own scale rather than a purpose on an existing icon: an icon declared any maskable is used at both jobs, and padded enough to survive the mask it is a small leaf adrift in a field of vellum everywhere the mask is not applied. The ground was already full-bleed paper with no transparency, which is the other half of a maskable icon done right, so only the scale had to move.

0.33 was the arithmetic and 0.32 is what shipped. The first re-render measured back at 0.400 of the square exactly — on the line, with no margin for a platform that rounds its mask the other way.

images/icon-512.png had been written on every run since the day the mark was drawn — 7 September, before this register existed to record it — and was referenced by nothing — not a page, not _headers, not the sitemap. Somebody laid this groundwork in the first week and stopped.

There is no symptom. A manifest served as the wrong type, or naming an icon that is not there, is simply ignored: no console error, nothing wrong on the page, and an install affordance that never appears. So _headers types it explicitly — the .avif lesson, where production sends nosniff and a type the local server guesses is one the live site cannot recover from — and tools/serve.mjs learned it too, so local and live agree.

Two gates had allowlists that would not have seen it. check-cache.mjs and check-metadata.mjs each decide whether a link is a fetch by matching its rel against a list, and neither list held manifest. The one asset every page on the site fetches would have gone through the undeclared-asset check unseen, and a manifest pointed one day at somebody else’s CDN would have gone through the third-party check unseen. An allowlist is only as good as its newest entry, which is the argument for adding one in the same edit rather than the next.

Check 11 is the new guard, and it checks the copies rather than the judgements: the two colours against the --qe-paper in the stylesheet’s :root block, the name and the description against the home page’s own, every icon against the file on disk and against the size it claims to be, and every page for the link. It also guards the one number the pixels cannot give back — Node has no image decoder, so the maskable scale is measured where the drawing happens and the gate only keeps the constant from being tidied back up. Made to fail in five shapes before being believed, each reverted after: a page that does not link it, a colour that is not the paper, a sizes that lies beside an icon that is not there, the maskable scale raised, and a description typed rather than copied.

The rendering gates were ported from our sibling site and the port numbers came with them, comments and all. 9412 is the contrast gate in both repositories; 9413 is overlap here and check-sheets there; 9414 is width here and two gates there. Two checkouts on one laptop could not run their rendering gates at the same time, and the symptom read as a broken gate rather than a busy port — which is also the shape of the five-day orphan this file already records.

The refusal was right about the danger and wrong about the remedy. cdp.mjs refused to run if anything answered on its port, which protected against measuring in a stranger’s browser by declining to measure at all — and fired on a neighbour doing nothing wrong. The number a gate asks for is now a preference: a busy port is stepped over in hundreds, 9412 to 9512 to 9612, on ladders that are disjoint so the three gates never chase each other onto one another’s ports.

And the guarantee got stronger rather than weaker. Once the endpoint answers, the listening process is walked up its parent chain to the child we spawned — Chrome’s helper holds the socket, so the listening pid is routinely not the one we started, which is what makes this an ownership proof instead of a guess. The old refusal never proved anything; it only declined to be in the room.

A busy port now says which kind of busy it is, because that is the judgement a person gets wrong under pressure: a headless Chrome whose parent is still running is somebody else’s sweep in progress and must be left alone, and one whose parent is gone is an orphan and safe to kill. That distinction was got wrong in this very session — a Chrome was killed on a heuristic that could not tell the two apart, and the parent was checked afterwards rather than before.

Star Stuff is not edited from this repository, so the whole fix is on our side. Proved in six shapes and every decoy cleaned up after: a quiet ordinary run on the preferred port; a live holder, correctly named as live with the advice to leave it; an orphaned holder, correctly named with the pid to kill; a full parallel sweep with all three preferred ports blocked, which is the exact case that refused this morning and now passes in 305s; the ladder exhausted, which still refuses and exits 1, because a gate that cannot own its browser must not measure in anybody else’s; and the precedent this follows — .claude/launch.json stopped hardcoding 8766 for the same reason, a second session being unable to preview the site at all.

Measured minutes after this sheet was published, when the branch had been deleted from both ends: draft-known-and-felt-and-seen--queering-earth.netlify.app answered 200, still serving the pre-publication draft with its review layer attached, and a full copy of the site beside it. Netlify keeps the last branch deploy at its subdomain after the branch is gone.

The same wrong sentence was in three places — this site’s own build notes, start-draft and publish-draft — written once and copied twice, and never checked in any of them. That is the drift this register exists to catch, in the files that describe how the register is kept.

The noindex is what makes it survivable, and it is doing more work than we knew. The orphaned deploy inherited X-Robots-Tag: noindex from the drafts base, so it cannot be crawled. The live risk is a reader rather than a crawler: anybody holding the old review link keeps reading a superseded draft and may annotate a version that no longer exists. Removing the deploy is now a step in publish-draft, and it is a dashboard action — nothing in this repository can do it and no gate can see it, which puts it in the same small class as the bookmarklet: a check a person has to run.

Check 7c. Section 7b asserts that every third-party origin our scripts can reach is named, which is the check that matters for what a browser may talk to and is structurally blind to how many recordings there are — one origin covers any number of videos. So it counts the data-embed-id facades on every page, reads the list under #the-embed, and compares them per page. It also reads the number the page says out loud in its own prose, because a list that grew while the sentence above it did not is the same failure one layer up.

Made to fail in five shapes before being believed, each reverted after: the historical bug put back — the two Sheet No. 12 recordings unlisted with the count sentence agreeing with the short list, which is exactly how it shipped; a count word wrong on its own; the list naming a recording no longer on the page; a new facade added to a sheet that had none, which is the case that will actually arrive next; and the two ways the check could be quietly defeated — rewording the sentence it reads, and renaming the section’s id. Both of those fail loudly rather than going silent, because a gate that stops applying when somebody edits around it is worse than no gate.

A named target that could not be found printed SKIP not found and the run went on to report PASS, 0 problem(s) and exit 0. It happened here: a page was checked from a branch it did not exist on, and the gate certified it — 0 tags, 0 ids, 0 elements balanced, and a clean bill of health.

This is the empty-pass fault the rendering gates already guard, arriving in the one gate that had no guard for it. The wording is taken from theirs, because the reasoning is identical: a page that measures nothing reports zero problems, which is indistinguishable from a clean page in every other line of the output. Unread targets are reported in their own block, never mixed into the problem count — an unmeasured page is a broken run, not a bad page — and under --check the run fails.

Proved in four shapes: the missing file alone; a missing file beside a real one, where the real page must still be measured and the run must still fail; without --check, where it reports and does not gate, which is this tool’s stated contract; and a real page, which still passes.

The cross-page half is the half that matters, and it is why this could not be a per-page check. Ids here are authored topical anchors precisely so a reworded heading keeps its address; seven anchors point at the plate alone, and nothing verified that one of them landed. So the gate builds its map of ids from every page in the cabinet even when the run is scoped to a single one.

It reads anchor elements and never the string href=. A code span documenting a link writes the tag with only its brackets escaped, so the href survives verbatim and a text search would match this site writing about its own links — the same fault the review-layer guard already records, arriving a third time. Proved by injecting one beside three real faults: the three were reported by name, the documentation was not, and the count of links resolved moved by four rather than five.

The contents list, the rail and the entry index never appear in it: they build their links at runtime from ids they have just read, so they cannot be wrong. What the gate sees is what a person typed. 864 of those resolve today, and the number prints on every run, because a check that measured nothing looks exactly like a clean one.

_headers had carried frame-ancestors ‘none’ and nothing else since the header shipped, with a comment saying a script-src would kill the snippet every page runs before first paint. That was a plausible reason nobody had checked. Counted instead: one executable inline script on the site, byte-identical on all twenty-nine pages, so a single hash covers every page. The twenty-eight JSON-LD blocks need none — a data block is never executed, and hashing them would have buried the one hash that matters among twenty-eight that do nothing.

What is locked now: scripts to this origin and that one hash, no eval, no inline handlers, no plugins, no base tag, no form submission anywhere, and frames only from the one embed the privacy policy already names. style-src keeps ‘unsafe-inline’ and that is not laziness — 1,299 inline style attributes carry each sheet’s own foxing seed, stamp rotation and card colours, and a CSP hash does not cover a style attribute, only a <style> element. The alternative was not a stricter policy; it was deleting the per-page palette.

The snippet it hashes lists the nine typefaces the picker offers, so narrowing the picker changes the hash in the same commit — and a stale hash does not warn. The browser simply refuses the snippet on every page, and every reader who asked for the cabinet or for plain view is shown the other thing before the stylesheet catches up: the exact failure the before-first-paint rule exists to prevent, arriving silently and visible only to somebody who had a preference stored and was looking for it. So the policy is written between markers by a tool, check-cache.mjs fails when the header and the pages disagree, and the generator runs last of the five, after everything that can rewrite a page.

Proved in a browser before it shipped, both ways. The policy went into a copy of the home page as a meta tag with a preference stored: the snippet ran and the cabinet came up. The same copy with one character changed in the hash refused it — and Chrome’s own message named the hash the generator had computed, which is the browser confirming the arithmetic rather than us confirming it. The gate was made to fail and then to pass the same way.

The site asked the reader’s device for nothing and said so nowhere. It says so now: thirty-five capabilities denied outright — camera, microphone, geolocation, payment, USB, serial, Bluetooth, MIDI, the motion sensors, the credential and attribution APIs. None of it costs a reader anything, because this site calls no gated feature of its own, which was checked rather than assumed.

A deny-everything list would have broken the one control a reader actually presses. An iframe’s allow attribute can only narrow what the page already has — it cannot grant what the header denies — so a blanket denial would have left whoever pressed the Eliot facade with a player that would not start and a fullscreen button that did nothing. Five features are delegated to that one origin, never to everything, and only fullscreen also keeps this site itself.

clipboard-write is not named here, and looking it up is what saved it. Both injected layers copy out through the clipboard, so it seemed the thing most in need of protecting — and it is not a directive in the registry at all. Writing it would have been the sitemap relation a second time: a token that reads as standard, is not, and rides along on every response saying nothing.

What is left unnamed is the other half of the care. Screen-reader announcements, and a reader’s own translation and speech tools, are not capabilities to take away from somebody trying to get at our words. Denying them would have looked more thorough and been less careful.

Proved behind a throwaway proxy putting the real header in front of the real site, because a permissions policy cannot be set by a meta tag the way the content policy was tested: camera, microphone, geolocation, payment and USB all refused for our own pages; all five player features allowed for that one origin and refused for an unrelated one; and the facade pressed, the player built, and Eliot read.

All three are git-backed editors over collections of files with named fields. A sheet here carries eighty to a hundred and twenty hand-drawn SVG elements, eight to thirteen authored section ids, forty to seventy inline style attributes seeding its own foxing and stamp rotation, and JSON-LD whose about.author is a judgement about who made the thing being read. There are no fields to put in a schema. That is the standing reason — the layout is the argument — stated as a measurement rather than a conviction.

The deferral now rests on one leg. Its third reason was struck yesterday; its second said Netlify Identity’s git-gateway was being sunset, and Netlify reversed that on 19 February 2026. The remaining reason is the one that was always carrying it.

Nothing on any page loads it — a bookmarklet injects it, so a reader’s visit fetches nothing extra and no page markup changed to allow it. It cannot save, which is the design: saving from a browser needs a token in the page, and that is the account problem the CMS question was trying to avoid.

What it refuses to edit is the important part. Quotations, captions and citations are not editable, and the page says why where you try. This site’s characteristic failure is a tightened source rather than an invented one, and a tool that let anybody reword a mounted quotation in two clicks would be a machine for producing exactly that. The accession stamp and the provenance line are refused for the same reason in another key: a record you can edit in a browser is not a record.

The patch is decoded text and the first version claimed it was the file’s bytes. innerHTML decodes entities, so a before-string matched nothing for any paragraph holding an apostrophe. Re-encoding cannot rescue it either: the sources are mixed — 1,300 raw em dashes beside 1,512 — — and no encoding reproduces a file that is not consistent with itself. So it says what it is, and tells whoever applies it to match on decoded text and stop if nothing matches exactly. Proved both ways before it was believed.

That tool fences its code-span placeholders with literal NUL bytes, so file calls it binary data and grep prints nothing rather than saying why. A search for a string that is plainly in the file exits quietly with no output, which looks exactly like a file that does not contain it. Three searches and one wrong theory about corruption before file was run on it.

The fence is right and is not the thing to change. A placeholder needs a delimiter no source text can contain, and this is a repository whose prose is about its own markup — every printable candidate is one an entry may legitimately write inside a code span, and the day it does, a placeholder is restored into the middle of somebody’s sentence. The cure is knowing: grep -a, and run file when a search comes back empty on a file you can read. It bites an edit script too — the apparent spaces around the placeholder index are NULs, so a pattern typed from what the terminal renders will never match, which is how the first attempt to teach that tool strikethrough failed on an anchor that looked correct character for character.

The three Chrome gates here bind 9412, 9413 and 9414. So do Star Stuff’s, because these were ported from there and the numbers came with them, and the two repositories sit side by side. A session working in each at once is two suites contending for three ports: today a sweep hung for twenty minutes, and a later one had check-width.mjs refuse in 0.2s while a neighbouring check-overlap.mjs held 9414.

The refusal is cdp.mjs working exactly as it was written to, after a gate spent five days measuring in an orphaned browser it did not launch. That guard cannot tell an orphan from a live neighbour, and should not: driving either is the same mistake. Renumbering would end the collision and end the protection with it, in the one direction it has already paid off — so it was weighed and refused. The collision is loud and costs a wait; the alternative is quiet.

Checking the port first was the obvious advice and it is wrong, which was found by writing it down and then trying it in the same hour. The ports were free, the sweep started, contrast and overlap ran their 55 and 42 seconds, and width refused anyway — the neighbour had taken 9414 in the seconds before it spawned. The window is every moment until each gate binds, and nothing outside the tool can close it. What works is re-running the one refused gate: the refusal costs 0.2s, the other eight have already reported, and width came through on the twelfth attempt. Wait rather than kill — a neighbouring gate is somebody’s real run.

It is del and not a class. A line-through drawn on a span is invisible to a screen reader, so the one mark on these pages that says this is no longer true would be missing for the readers least able to infer it. The element carries the meaning and the colour only shows it: --qe-coral on the rule, --qe-moss on the letters, the colour on the strike and never on the word — the idiom the restored attribution on Miranda already uses, reused rather than restated.

The struck sentence drops out of the finding aid, and that is correct. The index skips del subtrees, because plain text has no vocabulary for this is the reading we got wrong and a snippet would hand a searcher a retracted claim as a current one. Checked rather than assumed: the page still indexes that passage once, through the 12 September entry that quotes it as history and says why it was struck. The retraction is findable; the retracted claim is not.

What has arrived counts a page’s corrections out of this register and prints the total under the word since. A sheet is routinely corrected while it is being mounted — a citation checked against the primary, a story everybody tells found to be wrong — and those entries are filed to the sheet like any other. So The Swell and the Dwell read Corrected four times since for four corrections dated the day it arrived, and five sheets read as corrected while their own provenance lines said Not yet corrected. Nothing was miscounted. The listing was asserting a chronology it had never checked, which is this house’s own recurring fault wearing a number: a claim that looks like a measurement and is not one.

The line now splits the two — Corrected twice on mounting and once since — in the vocabulary The Die Uncast already uses on its own provenance line.

The obvious key is the accession that mounted the page. Measured, several pages have more than one: this register files three Mounted accessions to the ledger on the day it arrived and two to My Gloomy Sunflowers, because Mounted is this register’s word for anything added to a page rather than for a page coming into being. The arrival day is a thing, because it is a measurement of the repository and the listing already had it.

It is also the only rule a generator is entitled to. Within one day the order of accessions here is editorial and nothing can verify it, so a tool that placed one same-day entry after another would be claiming a fact it does not have. A provenance line may draw the finer distinction, because it is authored and knows which accession mounted its sheet — The Die Uncast separates four corrections on mounting from three more when the plates changed, all seven on 8 September. The listing says the coarser true thing and the sheet says the finer one, and neither contradicts the other now.

Thirteen provenance lines were read against this register and twelve were exactly right. The Army of the Upright was the one: it said re-determined once, and the register holds two. The missing one is 9 September, when a line of Woolf’s turned out to be Woolf guessing at what Coleridge meant and the quotation was re-cut to start where the guess does. The line names it now.

The expected answer was the other one. Five sheets say Not yet corrected over register entries filed to them, and the first reading was that all five had drifted and wanted mending. They had not: an accession history reports what has happened to a sheet since it was accessioned, and five other lines go further and say on mounting or before it shipped in their own words because they chose to. Forcing a gold seam onto four clean sheets would have been decoration asserting a fact, in the component built to make real mends findable. Measured before mended.

The rule had one user and the user is gone. A rule with no element is not the fault this house watches for — that one is a class with no rule, which renders black and no gate here catches it — but it is the same drift from the other side: a stylesheet describing a component the site no longer has, with a comment above it explaining why the component is not optional. The comment was true when it was written and had to go with the rule.

The ledger and what is settled both carried the old state as a resolved question. Both now carry the re-determination beneath it rather than instead of it, because the 7 September call was half right and stays half right: crediting him was the right one of the three options, and the masthead was the wrong place to do it.

Only half the problem was ever hard. Reading a draft never needed an account — a Netlify branch deploy is a permanent URL. Giving feedback did, because a pull-request comment thread was the whole of the answer.

Drafts now sit at the root of a long-lived drafts branch, so a draft is byte-identical to the page that will ship and promotion is a merge. A drafts/ folder on main was measured and refused: it works mechanically, because every generator and gate here reads the root non-recursively, but it would put unreviewed attributions on queering.earth itself — unlisted, public, unauthenticated — and a page one directory down needs ../queering.css, so it renders something other than what ships. The gates that fail on that branch are the accession checklist reported by name: the missing 301!, the missing card, the missing <loc>.

The review layer is three lines in one head block, and publishing means deleting them. Armed, every block grows a visible focusable Note control; notes live in localStorage and leave only through Copy my notes, which builds plain text carrying the section, a deep link to its authored anchor, the quoted line and the reviewer’s own words. No endpoint, no account, no third party — this site makes no third-party request and a review tool is not the place to start one.

It writes words, which every other script here is forbidden to do, and that carve-out is stated rather than taken. These are instrument labels on a page that will never be published, and deleting the include line removes every one of them; authored templates instead would mean a block of markup pasted into every draft, which is the friction this exists to spare.

The guard keys on sitemap.xml, because that is this site’s own manifest of published addresses: a draft is not in it, so the branch stays green, and the moment a page is accessioned the gate demands the block be gone. It fails on the noindex half too — a page the manifest advertises and the markup hides is two statements that cannot both be true. Made to fail in both shapes on a real published page before being believed, and reverted.

Its first version was defeated by this register. The check grepped for the bare string drafts/review.js, and the sweep failed on this page and on what is settled — the two pages that had just been given an entry describing the feature. A gate that matches a string matches the documentation of that string, which on a site that documents its own build is not an edge case. It requires an unescaped script tag now. Check 7 in the same file already knew this, exempting a link because a link the reader chooses is not a request the page made; that distinction is called the whole difference between a usable gate and a useless one in its own comment. Same lesson, second gate.

Nothing gates runtime furniture, so it was measured by hand, and the first two measurements found real faults. --qe-paper-deep was the obvious ground and is wrong: moss and rust measure 6.90:1 and 6.21:1 on it in daylight, reproducing to two decimals the numbers that withdrew the recessed panel once already. Everything sits on --qe-card instead — ink 15.26:1 and 13.39:1, moss 8.35:1 and 7.53:1, rust 7.52:1 and 7.57:1 across the two grounds. And the 44px floor missed every Note button, because a selector for a button inside the review element does not match a button that is one; they were 40.3px until they were measured. All 54 controls clear 44×44 at 375px, none overlap, nothing scrolls sideways.

Two smaller things, each paid for by watching it go wrong. A definition list is one target, like a quotation — the specimen block fragmented into twelve note buttons first, which is the cropping fault from the other direction. And a note whose paragraph is rewritten under it is marked, never dropped: it re-attaches by position, says so on its face, and still exports. Silently losing a reviewer’s note is the worse failure.

start-draft puts a page on the branch and prints the URL to send Helen. save-draft commits, pushes and hands the URL back. publish-draft takes the file onto main — never the branch — and walks the accession checklist this file already states: mount it, card it, log it, file it, stamp it, route it, crumb it, group it, rail it.

tools/draft.mjs holds the one thing all three need to agree about, which is which file is the draft. Written out three times it would be three answers the day one was edited, in the one place where being wrong means publishing the wrong page. A draft is identified by its marker rather than by a name kept somewhere, so a page stops being a draft at exactly the moment it is published, with no state file to fall out of step.

It matched the bare filename first, which is the third time in one day. Check 10 shipped that way and fired on this page and on what is settled; this tool did the same an hour later and reported both of them as drafts in progress. On a site that documents its own build, anything matching a filename matches the prose about that filename — and the discriminator is always the same, because prose escapes the tag or wraps the name in a code span.

Each Chrome gate spawned Chrome and then polled /json/version until something answered. On 7 September an orphaned headless Chrome from a neighbouring project’s probe script took port 9414 and never let go. Every check-width.mjs run from that day to this one bound nothing, attached to that five-day-old browser, measured all 27 pages in Chrome 152.0.7977.77 while .84 was installed, killed its own portless process on the way out, and reported PASS with no sign whatsoever. Five such orphans were up, one per interrupted probe, holding about 400MB, with 303 temp profiles behind them.

This is the house’s own recurring fault in a new place — a run that measures a fraction of what it claims looks exactly like a clean one. It was found by timing the gates, not by any check, and no check here could have found it: every gate reports on the site, and nothing reports on the instruments.

cdp.mjs now owns the spawn, which reverses that file’s own stated position that the duplication “carries no footgun”. The port must be free before the spawn, and the process we started must still be alive when the endpoint answers — the old poll swallowed every failure and fell out of its loop silently, so a Chrome that died on launch produced a sweep of nothing rather than an error. The temp profile is removed on the way out, which no copy did. Made to fail in three shapes before being believed: a decoy on the port refuses and exits 1, a browser that cannot start reports Nothing was measured, and the port free again passes.

Measured before anything was written. The four generators are 4.8s and the six offline gates 5.6s together; check-overlap is 40s, check-contrast 53s, and check-width is 4:22 — three quarters of the whole bill, because it renders 27 pages × 10 typefaces × 4 widths × 2 papers.

Three things made twenty minutes out of six, and none of them needed new capability. The three Chrome gates run at once, which they were always built for — they take one port each and cdp.mjs has said why since it was written. The sweep is scoped to the pages that moved, through the positional arguments all three already accepted. And nothing is run twice to read an exit code a pipeline discarded. Full sweep: 4:47. A one-page change: 24s, generators and all nine gates included.

The scope is taken from git, and only after the generators have run. That order is the reason this is a tool and not an alias: make-records and make-whats-new write into pages nobody touched, so asking git first scopes to the edit and misses the pages the edit caused. Untracked files count, because a new page is the case most in need of measuring and a plain diff cannot see one.

Which assets break scoping is declared, never inferred — the two-list shape check-cache.mjs already uses, because nothing in a file’s bytes says whether it can move a page it is not named in. The stylesheet, the script, the fonts and the gates themselves force a full sweep; queering-search.js adds the finding aid and nothing else. Everything else is page-local by declaration, --all is always available and always right, and a new shared asset needs a line in one of the two lists.

Both embedded recordings are the owners’ own deliveries and were verified as such before they were mounted, not assumed from the channel name: Mad Rush is Sony Classical’s delivery of the 1989 Solo Piano recording with Glass himself at the piano, and the Op. 47 first movement is Warner Classics’ delivery of Renaud Capuçon and Frank Braley. The first draft of this sheet carried a listener’s upload of the Glass; the official one exists and was found by looking, which is the outcome the policy is for.

The Immortal Beloved scene has no such copy and is linked rather than embedded. No official clip or licensed excerpt of the film could be found. The policy asks for a copy the rightsholder is paid for; this is not one, and the relaxation is Ryan’s call, recorded here and stated on the sheet’s own face in what it does not settle. A link is a request the reader makes; an embed is one the page makes for them. That distinction is the whole of what the relaxation rests on, and it is why the film gets a link and not a facade.

Eliot’s four lines, the Syringa vulgaris drawn from Ryan’s yard and the click-to-load recording were mounted above the footer and last on the page, at 16:17 on 9 September. They were the last thing on it for two and a half hours. The cabinet’s own pages came out of the footer and onto the home page at 18:54 that evening, the thirteen aims followed at 19:47, and the two lists became the founding papers and the cabinet itself on 11 September. Every one of those edits appended, which is the ordinary thing to do, and the sum of them left a coda in the middle of the page reading as a digression between the plate and the furniture.

It now sits after both lists, which is the foot of main again — and last for a structural reason rather than a chronological accident, since nothing on the home page follows the drawers.

The section had no h2 and had never needed one: a passage at the end of a page, after the argument has finished, announces its own genre by position. Moved into an outline it has to be named in that outline, so it is Coda: the lilacs — the word says what the section is, which is the job the position used to do.

The address is #lilacs and not a slug of the heading. Ids here are topical and authored precisely so a reworded heading does not take every shared link with it; Coda is a word we might argue about, and the lilacs are not. The rail picks the section up with nothing else added, because it derives from the page’s own headings — and the home page has no contents list to keep in step, its plate being the index it already has.

check-markup.mjs’s header had said for its whole life that unclosed tags were out of scope by design and that guessing at intent would make the gate noisy. Both halves were wrong, and the header now says so. The cost of the first was the accession this register lost from the feed for a day. The second was tested rather than argued: every non-void element on every page was counted before a line was written, and the whole repo balanced but for two — that section, and an unclosed strong on Two Cohabitating Modes that had been bolding the rest of a list item since the sheet was mounted.

So it checks every element rather than a declared list of containers: a list would be a second thing to keep in step and the measurement says it would buy nothing. It is stricter than HTML, which lets p, li, dd and td close implicitly. That is deliberate — three generators here parse this markup with regexes, so balance is a requirement of the repository and not a matter of taste.

A stack, not a counter. Counting opens against closes returns zero for <div><section></div></section>, which is mis-nested and breaks those generators identically; a stack also names the element and the line where a counter can only say a total is off by one. That distinction paid for itself on the first run.

The house rule is that a guard nobody has watched fail is a guard nobody should trust. Three faults were reintroduced and each reverted after: the original missing </section>, reported against the </main> a browser would close it at; the unclosed strong, reported against its </li>; and a deliberate mis-nesting that balances numerically, which is the case a counter structurally cannot see. Each exits 1; the clean tree exits 0.

22,709 elements balanced across 27 pages, and the count prints on every run — the same reasoning as the contrast gate’s element counts, because a gate that checked nothing looks exactly like a gate that passed.

/feed.xml has shipped since the agent-readiness pass. It was in every page’s head as rel=alternate and on every response as a Link header — so a client that had already been handed the URL could find it, and a person could not. There was no visible link to it anywhere on the site. A feed with no human twin is a feed for machines only, which is not what a feed is for. The new page is that twin, and carries the first pressable route to either feed.

The feed carried this register’s accessions — new sheets, corrections and housekeeping notes in one stream. On Ryan’s call it is now split. /feed.xml is the pages as they arrive, the same list as the new page and out of the same pass, so the two cannot be made to say different things. /register.xml is this register, under the name the site already uses for this page.

The pages took the old name rather than the other way round, and that is the part worth writing down: anyone already subscribed stays subscribed to a working feed and simply begins receiving pages, which is the list most people mean when they say feed. Every page advertises both, the Link header carries both, and check-metadata.mjs now fails on a page that names one and not the other.

The obvious source for an arrival date was this page, and it does not work. Mounted here means anything added to a page, not a page coming into being: the colophon carries five such entries, Two Cohabitating Modes six, the front of the cabinet ten — and the privacy page carries none at all, because nothing was ever mounted on it. Taking the earliest would have dated several pages by accident and one not at all.

So the date is the commit that first added the file, which is a measurement of the repository rather than a second hand-kept copy of anything. And it is gated against what we wrote down. Eleven sheets carry an accession stamp with a date in it and twelve pages carry a provenance line with another; the generator compares all three for every page and refuses to run on a disagreement, because the day a stamp and a commit stop agreeing this site is about to say two different things about one day. All twelve agree today — checked before a line of the tool was written, not after.

Two things were found by running it rather than by reading it. A page being written has no commit that added it, so an uncommitted page is dated today, loudly, and corrects itself on the next run — self-healing rather than an override table, which would have had to be right forever. And the first version of that fallback used UTC and dated this page 12 September while it was still the 11th at the desk, a day ahead of the accession you are reading.

The request included lyrics. For Waitin’ Around to Die the answer is the ordinary rule — we quote what we have read, and no licensed text was in front of the session; the performance is embedded instead.

For Polly the same applied, and a second reason is stated on the sheet itself. Every word of that song is spoken by the man who committed the crime, which is what Cobain built it to do. Setting those words in a ruled box under a herbarium label would give a rapist’s speech the mounting this cabinet gives to Woolf and to Plath. The recording is linked whole instead. Whether the wording belongs on the page is Ryan’s call, and it is recorded here as a decision rather than left as a gap.

The gate shipped this morning and was already blind in the way this house had written down three days earlier: a page that checks the fraction of itself that happens to be visible reports zero failures and looks exactly like a clean one. It rendered the default typeface on every page while nine others sat unmeasured.

The face list is read from tools/font-files.json and never typed into the tool. That table is where make-fonts.mjs writes the faces and where check-metadata.mjs already reads picker_id to prove the markup’s options match it. A hand-kept copy would be a third list — and narrowing the nine is an anticipated operation, which would leave the gate sweeping a face nobody can pick while missing one they can.

The extra passes are cheap because they ask a cheaper question. The default face gets the full walk — every text node, every client rect — which is what names a culprit and what proves the page was measured at all. The other nine ask only whether the document is wider than its box, and escalate to the full walk only when it is. A gate ten times slower is a gate people stop running, which is its own failure mode. 1,500 measurements in 4m11s.

document.fonts.ready after every switch, because a face is fetched when a rule using it first meets rendered text — measure before that and you have measured the fallback. And the faces go to paper, which earned its keep on the first run by finding the two printed conditions the by-hand check had missed: a reader’s typeface is an html class and the print sheet does not reset it, so what they picked is what they print.

The failure advice branches, because two faults wear the same number and take opposite cures. An unbreakable string wants a break opportunity and must not be fixed by shrinking the type; a display line too wide in one face has no string to break, and there the size is the fix. Printing the first advice under a typeface finding is how somebody ends up putting overflow-wrap on a masthead — which happened here, earlier the same day.

Made to fail before being believed: withdrawing the Sporting size reports five failing passes on the home page, names the typeface in each, points at <h1.qe-wordmark> and the word Earth, and exits 1.

tools/check-width.mjs: no page scrolls sideways, at 320, 375, 768 and 1280, and nothing runs past a Letter or A4 sheet. Native, not ported, and it clears the bar check-cache.mjs set — the failure had already reached production. This house runs eight gates rather than eleven because a check that cannot fail is a check nobody reads; this one had three live faults waiting for it.

No existing gate could see it, and check-overlap.mjs is the near miss worth naming. It is the only other tool here that knows where anything is, and it asks a different question: is this text on top of that text. Text past the right margin is on top of nothing — there is nothing out there. Its clip walk does not catch it either, because the document is not a clipping container: it grows. And its screen pass is 1280px, where none of the three faults exists.

Folding it into that gate was considered and refused. Widening the overlap sweep to 320 and 375 would surface a flood of narrow-width collisions nobody has ever triaged, which is a separate project; and a gate named for text landing on text should not also be the one that measures margins. 320 is not arbitrary either — WCAG 1.4.10 asks that content reflow to 320 CSS px without two-dimensional scrolling, which is 1280 at 400% zoom.

It has no --check flag, deliberately. Two gates here shipped as a report that printed its findings and exited 0, and the contrast gate stayed advice for as long as CLAUDE.md forgot the flag. A gate whose default is to pass has not shipped.

Made to fail before being believed, per the house rule, and both halves separately: withdrawing the wrap rule reports nine failing passes across the three pages, withdrawing the print margin reports two on the home page, and the restored stylesheet reports PASS across twenty-five pages and six passes each. Its first draft named the wrong words and was corrected — it reported the run ending furthest right, which for an unbreakable string is the innocent tail after it, so it pointed at , not out of a reading of the source. rather than at the code span before it. It reports the widest overflowing run now. A gate that names the wrong words is a gate somebody edits the wrong words to satisfy.

The cabinet drawer said six pages in three places and now says seven. The heading’s id did not move: #the-six is an authored topical anchor and not a slug of its own wording, and outliving a reworded heading is the entire reason ids here are authored. Renaming it to match the new count would break exactly what the rule protects.

The footer row gains Quoting, which is a nav label and not the page’s name — allowed for a link, refused for a group heading, and the distinction is already in CLAUDE.md. The row now wraps to fourteen targets and was re-measured after, because nothing in the build measures a hit area.

Ryan’s, after the drawers shipped. They sit directly after Queering Earth on all twenty-three rows, which is the level they are: the home page, then the drawers, then the pages they collect.

The labels are not shortened, and that is this file’s own licence running out. A row of links is allowed a short label — Colophon, Register, Decisions — because those name pages. The founding papers and The cabinet itself name sets, in the same words in tools/pages.mjs, on the home page, in /llms.txt, in the drawers menu and on each collection page’s own heading. A short label for one of those is the second name for a set this register keeps recording as drift.

The row wraps now, and a single line was never the invariant. The stylesheet’s note recorded eight links at 575px on one line and had gone stale twice over — the row has been ten links and two lines since the founding papers shipped. What actually has to hold is that no two padded targets on adjacent wrapped lines can touch, which is what line-height: 2.9 is for. Measured at thirteen: two lines at 1280 and 768, four at 375, six at 320; no overlaps at any of them, every target past 44px in both directions, and scrollWidth equal to the viewport at all four. The note says so now, and says to remeasure when a link is added.

The reading panel had been right: 0 on its own summary since it shipped, which quietly made its position a function of how many controls sit to its right. Measured at 1280px it cleared the measure by 31px. Adding one control moved the control 81px left, took the panel with it, and landed 50px of it on the lede on eight pages. check-overlap.mjs reported it; nothing else would have, and no amount of care about the new control would have predicted it. Both panels now hang off the tray, whose right edge is the gutter at every width, and a seventh control cannot move either.

The tray’s own note said a new control is a measurement, and it was right about the row and silent about the overlay. Measured at 320, 375, 768 and 1280px after: every hit area 44×44 or better, no two overlapping, scrollWidth equal to the viewport at all four, and both panels inside the gutters at 320. The cost is paid explicitly — the 44px box .qe-controls > *::after grows only works because each child is positioned, so for these two it moves onto the summary and the tray’s rule is switched off for them rather than left to misfire against the tray.

Two overlay panels in one tray can both be open with JavaScript off, because a disclosure toggles natively and the close-the-other convenience was script. They now share a name, which is HTML’s exclusive accordion: opening one closes the other with no script at all. Verified by setting open directly, with no click and therefore no listener involved.

That fix made the two rendering gates disagree, and both were taught. reveal.mjs dissolves the group before opening, or one panel would be measured shut — the eight-percent-of-itself failure its own header is about, arriving through a fix for something else — and puts it back afterwards. check-overlap.mjs then knows that two members of one exclusive accordion can never be on screen together and refuses to call them a collision; the exemption is declared by the markup rather than listed in the tool, which is the bar the invisible-text list already set. Proved both ways: the drawers panel’s hints in a 2.15:1 colour report three failures under the contrast gate and none once restored, so the panel is genuinely being measured.

A hanging bell reads as a bud at the 21px it shipped at, and lobing its rim to say mouth is invisible at that size; the flower turned to face the reader instead, because what reads at sixty pixels a plant is silhouette and count. Florets on an arc are a ring, and a ring reads as a wreath — this repository’s own words about the lilac panicle, reproduced exactly. They sit on a golden-angle spiral over the head’s area now, the same walk the drift rail’s lean uses. And thirteen florets in a head this broad filled a tenth of it and read as beads on a wire: with the area fixed, a wider head needs more and smaller florets rather than the same ones spread further.

Three dead classes, caught by hand because the gate for them is not ported. The stems carried .stem-main, .stem-branch and .stem-twig beside .stem, with no rule anywhere — the stroke hierarchy was inline all along, as Sheet No. 1 does it. They did not render black only because every one of those elements also carried .stem. check-classes is still the next port from Star Stuff and this is the second drawing to want it; until it lands the check is a dozen lines, and it is worth running on any new drawing.

Drawn twice, rendered at 368, 544 and 900px in both grounds, refused twice. Pass one gave the mark colour fills and features — rust pelt, marigold dome, an eye, an inner ear, a white bib, blocked feet — and read as children’s-book clip-art, fighting the house register of ink line on vellum and sitting four inches above two wood engravings whose hatching makes a flat rust mass look like a sticker. Pass two stripped it to one ink colour and no features; the cabinet inversion came out right, and at 900px the construction showed: the neck a facet, the two ears horns with a notch, the nose a ball, the legs and feet six rectangles, and thirty spines that are a coat at 40px turning into a cog.

The rule is the favicon rule inverted. That one says a mark is judged at 16px because two earlier marks looked good large and died small. The converse is less obvious and cost two passes here: a mark drawn to survive 20px carries 20px tolerances, and enlarging it enlarges them. A masthead is a different drawing, not the same drawing bigger.

And the position has a weight. Every other masthead here is thin moss stems, small lichen leaves, a small moth — sparse line. Two filled quadrupeds are the wrong weight for that slot however well drawn, which is why another iteration on the same approach was not the answer. The sheet carries no drawn masthead; the Bewick plates carry the animals, which is what a mounted specimen is for. Whether an outline masthead — single closed contours, thin ink stroke, animals as line rather than as mass — would sit in the register is left open in the decision log. The tray mark is unchanged.

Checking two-cohabitating-modes.md turned up a fault older than this sheet. make-markdown.mjs treats span as an inline wrapper — correct nearly everywhere, and wrong for .qe-poem .l and .qe-verse .l, the one element here whose whole job is to be a line. So every stanza on the site arrives in its .md as one reflowed paragraph: Promise me no promises, So will I not promise you: Keep we both our liberties…

It contradicts the rule the component was built on — a poem is a specimen whose line breaks are the object; reflowed into prose it is a different poem — and the reason the generator exists, which is that the Markdown an agent fetches must not disagree with the page. Affected: Rossetti on Sheet No. 3, Dickinson and Ryan’s reply verse on Sheet No. 6, Sappho on Sheet No. 7, and now 53 lines here, which is what made it visible. search-index.json is already right, because make-search-index.mjs keeps a blockquote and a div.qe-verse as whole records — only the Markdown is wrong.

Not fixed here, on purpose. The fix is two trailing spaces after each .l inside a verse container, and it edits a shared generator and rewrites eight .md files plus llms-full.txt. That is its own accession with its own diff to read, not a rider on a new sheet. Recorded as open in the decision log, with the trap noted: the break belongs to .l inside a verse container and never to span generally, or every span on the site puts a line break through the middle of prose.

Measured before the mark was added rather than after: at 375px the four existing controls filled the tray exactly — search link at x 20, the Reading disclosure ending at x 355, inside a 20px gutter, with nothing to spare. The row is justify-content: flex-end, so a fifth item overflows off the LEFT edge and takes the document’s scrollWidth with it — the horizontal body scroll the layout rule forbids outright.

flex-wrap: wrap is the whole fix. On a narrow screen the tray becomes two right-aligned rows; at any width that fits, it is a no-op. The alternative was to drop the mark below some breakpoint, which would have withheld a feature from phones — and the mark is the only route to the page that a reader discovers by looking, so hiding it on the devices most people read on is where it would matter most.

Robert Garrison read the whole poem for LibriVox — 29 minutes 24 seconds — and released it into the public domain. It is in the sheet’s label block as a link.

A player was considered and declined on capability grounds, not taste. It would need a second encoding, because Ogg Vorbis alone is unreliable in Safari and nothing here transcodes audio; a MIME type taught to serve.mjs, which has no audio entry and would send application/octet-stream under nosniff — the .avif trap for the third time; a rule in _headers; a line in check-cache.mjs; and 18 MB in a repository smaller than that. /privacy exempts a href because a link the reader chooses to follow is not a request the page made, so the link costs nothing and adds no capability at all.

He is credited as the performer rather than as a licence. Somebody read a poem aloud for half an hour; filing that under a public-domain recording on a sheet about coinages losing their coiners would have been the sheet doing the thing it condemns, one section below the place it condemns it.

A sheet arguing that coinages get separated from their coiners has to show its own working, so one section is about this site: Post-normal possibilities is Helen’s phrase from the founding brainstorm, one hyphen from the end of Nick Walker’s subtitle, with a third lineage in post-normal science nobody here has read in the primary. The masthead says after Nick Walker — not his phrase, which would take the coinage off Helen, and not silence, which is how it shipped at first.

Nothing about the credit changed. What changed is that the reasoning is now on a sheet a reader will actually meet, instead of only in the ledger where it was logged on 7 September.

Six of the nine are under CUTE — Conditions d’Utilisation Typographiques Engageantes. It permits use, copying, redistribution, modification and commercial use, and states outright that it is not technically compatible with any FLOSS license to date, taking a position radically and deliberately inconsistent with them. It is explicit that it governs fonts and not the documents they set — its own example is that a book typeset in a CUTE font is not governed by CUTE — so nothing on this site is relicensed by using them.

Three of its conditions bind us. Credit the designers and link the original publication source — the first specimen named the collective in plain text with no link, which is half of it. Share complete files rather than lone extracted faces. And do not delete the post-binary characters or the OpenType features that activate them, which forbids subsetting: the six CUTE faces ship whole while the two house faces are cut to the latin subset. Its donation is a condition and not a suggestion; Stimpunks met it at the for an association tier.

Reading a licence properly is what found the faults at home. Sourcing these candidates meant reading nine licences, and that turned up two of our own: the house faces had been redistributed for a day with no licence file beside them, and the colophon credited both of them to a search engine rather than to Undercase Type, Phaedra Charles and Flavia Zimbardi, and Production Type. Every credit on this site is now fetched from upstream metadata rather than typed, which is the same standard the quotations are held to — and the reason it was worth reading the licence rather than assuming a font is a font.

The first option is not a no-op, and that is the whole reason it is worth showing. html is set to font-size: 100%, which inherits whatever size the reader’s browser is already set to, and every one of the 129 type sizes on this sheet is in rem — not one in px. So are the spacing scale and the measure. Raising the root therefore moves the line height, the gaps and the 66-character measure together, rather than swelling the words inside a box that stays where it was.

Larger only, and that is a house rule rather than a shortcut. The style guide puts body text at 18–20px and this sheet sits at 1.19rem, which is 19px on a default browser. A smaller option would take it under that floor, so the panel does not offer one. A reader who wants less turns their own browser down and the sheet follows — which is precisely what the first option names.

It is a <select> because two steps is not a checkbox, and native for the same reason the panel is a native <details>. It also makes something visible that was implicit: for the three checkboxes, follow your device is the absence of a stored value and a reader has no way to see it. Here it is an option they can choose by name.

The gate caught this key one setting after it was written, which is the point of it. The storage check found settings by their html.qe-*-on and -off classes, and text size stores larger and largest — so the pattern would have let qe-textsize go undocumented on /privacy while reporting green. Widened to any state, it failed immediately and by name. A check written for the cases that existed is a check with a blind spot for the next one.

Verified the way the last one should have been: a real click opened the panel, the control was confirmed to be the top hit target at the point a reader would press — the check that would have caught last night’s bug — and a real keypress moved it through both steps, 16px to 18 to 20, with the class, the storage and the reset button following. A real click on Follow my device again put all four back.

There is no toggle for forced colours and there cannot be. It reports an operating-system state — Windows Contrast Themes and its kin substituting the reader’s own palette for everything we declared — and a page can only respond to it. That is the point of the feature: the reader’s choice outranks the site’s. It sits beside More contrast in the panel as the thing the panel deliberately does not contain.

Most of this sheet survived the substitution already, on decisions taken for other reasons: every control is a real <button> with a real border, and the register’s four kinds carry words as well as an accent, so a palette that flattens their colours does not flatten their meaning. The rule that put the colour on the rule and never on the glyph paid a debt it was not taken out for. What needed repair was the card hover, which was a background change and nothing else and therefore vanished outright; the edges that leaned on a box-shadow; and focus, which now tracks Highlight rather than our rust. The botanical art is allowed to be forced — colour is not information in a drawing of a leaf, and the spec reserves opting out for the case where the colour is the information.

Three new keys meant /privacy was one edit away from being wrong about its own site, which is the failure this register exists for. So the policy names them, and a gate now enforces that it always will: every qe-* key the site reads or writes must appear on that page. It finds them two ways, because one is not enough — literal getItem calls, and the html.qe-*-on classes in the stylesheet, since the reading settings build their keys from a table and appear as no string literal anywhere. Proved by adding an undocumented setting and watching it be named in the failure. It is the sibling of the third-party-origin check mounted yesterday: not forbidden, named.

It would be easy to treat a document Ryan and Helen initiated as theirs to carry across without ceremony. Nine people wrote it. The foot of the sheet names every contributor and the licence, in the page rather than only in the ledger, and says plainly that what happened here is an adaptation: no sentence carried over verbatim, the argument taken off learning spaces and pointed at reading.

Share-alike is the reason this exists and the reason it is not the end of anything. CC BY-SA 4.0 is why the adaptation is permitted, and the same clause is why anybody may do this again to our page.

It does not restate the home page, which was the risk the mission page was shaped to avoid. The home page explains what queering is; the aims list what we are for and where each is kept; this argues why any of it follows. Three genres, one subject, no duplicated paragraphs.

Anchor on <div class="qe-register" id="latest">, never on the last accession you wrote. The register runs newest first, so the top is a fixed address and the previous entry is not. Inserting above your own last entry is right for exactly as long as nobody else adds one — which was true every day this register has existed until this evening, when two sessions appended to it for three hours. Ryan asked for the rule; it is in CLAUDE.md beside the checklist a session reads before writing an entry, rather than only here where it would be read afterwards.

The marker moves in the same edit. · latest comes off whatever holds it and goes on the new entry, and check-markup.mjs refuses a page where it is not on the first accession. That gate plus this rule cover the two halves: the gate catches a marker in the wrong place, the anchor stops an entry landing in one.

This entry was inserted by the rule it describes, anchored on the register div with the marker moved off the section below. Which is the smallest possible test, and better than none.

check-markup.mjs now checks that accession dates do not increase down the page, and that exactly one accession is marked latest and it is the first. The date check is worth having and was blind to this: every accession that day carries the same date. The marker check is the one that bites. Within a single day the order is editorial and nothing can verify it — but which section is newest is a claim the page makes out loud, and a claim can be checked. Both were made to fail before being believed, the first by putting this evening’s actual fault back.

It also turned up a third entity table. This gate carried its own six-entity decoder, which did not know · and handed the date check back a raw entity. tools/html.mjs was written to stop precisely that — an entity table is a lookup that one tool teaches and the other does not — and the gate simply was not using it. It imports the shared one now, which throws on anything it does not know rather than passing it through, and all seventeen pages still pass.

They went in this order, over about two hours. More than two headings — retired, because two entries that save twenty screens beat seven that save two. The block must not be too tall — retired, because 2,322px against a 120-screen page is a bargain and the comparison that decides it is the block against its own page. The page must be long enough — retired now, with the finding aid at 1,307 words.

Each of the three measured the list. None of them measured the reader. What is left is structural and has one instance: does this page already have a better index of itself? The home page answers yes — its plate of eight numbered cards is a contents list, and a second one above it would put two competing lists on one page. The error page is an error page. Everything else has one.

Ryan’s call, and it is the right one. A contents list needs more than two headings was the stated reason both record pages shipped without one, and it counts the wrong quantity: the reader’s problem on a twenty-thousand-word page is length, not heading count. Two entries that save twenty screens of scrolling are worth more than seven that save two. The question is only ever whether the page is long enough to want one.

Retiring a rule is a good moment to check what it was hiding. The privacy page is 2,058 words with eight headings and had no list — longer than Sheet No. 1 at 2,045, which has one. It was never a decision, just a page nobody counted. It has one now.

Four pages still decline and not one of them declines on heading count. The error page; the home page, whose plate is already a contents list; this register, whose accession headlines average 148 characters and measured 1,293px against 319 to 510 elsewhere — it gets its index as back matter instead; and the finding aid at 1,307 words, the shortest page here that is not an error page and well under the 2,045 of the shortest page that has one.

reveal.mjs says in its own comment that it is for a fetch or a hidden state, not for anything merely built at runtime, and this component ships hidden — which reads like a case for it. It is not: the drift rail’s if (r.hidden) continue is only sensible if queering.js has already run under the gate, so anything that unhides itself is measured without help. Confirmed by counting rather than by reading the source: /ledger went from 1,782 measured elements to 1,845 with the index in, and its print count did not move, because the index is display: none on paper.

Two columns above 34rem, one below. Fifty-five short names in a single column is a screen and a half of nothing but links with the measure wasted on them; two columns of wrapped link text on a phone is a column of three-word fragments. The divider is a column-rule rather than a gap, because a bare gap between two lists of links reads as one ragged list. check-overlap.mjs was the thing to watch here and reports zero collisions across all three passes.

One spacing fault came out of looking: .qe-index-link is shaped for the register, where the pointer is the last thing before the entries and needs no bottom margin. Here the body follows it immediately, so the line read as the opening of the paragraph beneath it.

Each cabinet page was introduced in the footer by an authored one-line gloss. Five sentences restated across sixteen pages is eighty copies of editorial prose, every one of them free to drift, and drift is what happened — which is how two pages could go missing from all of them at once without anybody noticing. The glosses still exist, once, in the furniture list on the home page, which is where this cabinet already said those sentences live.

The row uses short names, and that is a second name for those pages. It says Colophon, Register, Decisions; the pages call themselves How this site is made, The accession register and What is settled, and what is open. A row of links and a page title are different registers, and the rule against a second name is about the group headings, where a name identifies a set. The addresses are unchanged.

The home page keeps its bare footer: its furniture list sits two paragraphs above, with the same seven pages and their glosses, and a row of the same links directly beneath would be two lists of one thing.

Six sheets carried a paragraph of sourcing in the footer — which printing, how many scans, which page. The question was whether it belonged in the provenance line instead. It does not, and the reason is worth writing down: provenance here is accession history — when the sheet was mounted, how often its label has been corrected, each clause linked to the entry that did it. It is identical in shape on all eight sheets, and that sameness is what lets a reader scan it.

Which copy we read is a different fact with a different life. It is about the works the sheet reads rather than the sheet’s life in this collection, and putting it in the provenance line would bury the mounted, corrected twice pattern under prose that changes from sheet to sheet. It also already has two homes: the specimen block at the head of every sheet, and the ledger.

So the footer keeps one sentence — every quotation here was read in the work it comes from, and is logged in the ledger — and the trail stays where it was. What is lost is detail like read in three separately scanned copies, which is in the ledger, and which the sheets saying it were saying for the third time.

The ask was grasses whose taller stems form an arrow pointing up. A chevron drawn over grass is an interface arrow wearing a costume, and this cabinet already has its own word for up: growth, stems drawing themselves on from the ground. So two tall blades arc up and converge at the middle stem’s seed head, and the silhouette is an arrow without one being drawn.

The first attempt put two short awns on the head and read as a small figure with its arms raised. It was looked at, and redrawn — which is what this cabinet’s rule about judging a drawing at the size it ships at is for. The words carry the meaning either way: the drawing is hidden from the accessibility tree, the link is named by real text, and a reader who sees only grass has lost an ornament rather than a control.

This file already refuses a contents list for the home page because two competing navigational lists is a defect, and the rail plainly puts a second list beside the first on seven pages. The refusal stands and does not reach this: the contents list is front matter — read once, in sequence, in the measure, before the sheet begins. The rail is an instrument, glanced at from the middle of a sheet nine thousand words in. The home-page defect was two lists competing to be read; only one of these is for reading. Ryan’s call, against the recommendation to hide one, and the design work went into making them unmistakably different objects rather than into hiding either: the rail is smaller, mossy rather than rust, unnumbered, unlabelled, and out past the measure in space no line of text occupies.

It is unlabelled because a visible heading would be a third copy of the words On this sheet and the one free to drift. The aria-label is authored in the markup, once per page, and queering.js still writes no words — one builder now fills either container, because a second loop would be a second place for the three heading exclusions to be got right.

The register is the page that does not get one, for the reason it has no contents list either: twenty-eight headings averaging 147 characters, longest 226. In a 12rem column with no truncation allowed that is a four-thousand-pixel rail. The same measurement that sent this page’s index to the foot keeps the rail off it.

Two mistakes in the one rule this house wrote down because no gate enforces it. The instrument first: the view controls grow their targets with an absolutely positioned 44px box behind each control, and reaching for it here would have been wrong — rail links are stacked, so the boxes would overlap, the topmost would take every tap, and the rule asking for 44px would have produced a rail where most entries could not be pressed at all. Stacked targets take real padding.

Then the arithmetic. The padding was set from the measurements of a 1rem line and shipped 44px in a comment above targets that measured 31.1px — a control whose own note explains the failure, committing it. Found by measuring the rendered box rather than by reading the rule back, which is the whole reason this site measures. A 0.82rem line at 1.35 is 17.71px, so the padding is 0.83rem either side and the box is 44.3px. Verified after: no two hit areas overlap, and the tallest rail on the site — the colophon at fourteen entries, 690px — still clears a 900px viewport before the clamp scrolls it.

The rail lives outside the landmark, so no generator changed a single byte when all three were re-run — checked rather than assumed.

The rail shipped showing every label all the time, and needed nothing from tools/reveal.mjs — both Chrome gates load over file:// and queering.js is fetched by relative URL, so it runs and the labels were measured like any other text. That was proved rather than assumed, by putting a 2.9:1 colour on the contents list and watching the contrast gate report exactly one failure per heading on seven pages.

Then it was collapsed to hover, and all 109 labels went into the blind spot at once. A gate skips what it cannot see, and opacity: 0 is invisible. Proved both ways with a probe colour on the rail: with the new reveal line, 109 failures caught; without it, a clean PASS and 218 elements silently dropped from the count. The distinction is the thing worth keeping — REVEAL is for a fetch or a hidden state, and never for anything merely built at runtime.

The labels are left in flow at zero opacity rather than floated in pills, and that is what makes the reveal safe for the other gate. Pills would let the ticks sit at an even pitch, as they do on the page this was modelled on — and the moment the reveal turned them on for measurement they would lie over one another, so check-overlap.mjs would report every label sitting on the next. In flow, what the gates measure is the expanded rail that was already clean at 1280px. The pitch is uneven instead, because a row is as tall as its own label: one row on the home page, three on the waste garden.

Collapsed, the rail was a column of identical rules at an uneven pitch — uneven because a row is as tall as its own label. Ryan’s read was that this was an imperfection worth leaning into. The trouble with leaning into it as it stood is that asymmetry only reads as asymmetry against a norm, and with nothing else varying the pitch read as an accident rather than a hand.

So the ticks were given the norm to deviate from, and the deviation is inherited from the sheet rather than invented: each tick is now as long as its section is tall, normalised across the page. The rail is a small scale drawing of the document, and you can see which section is the long one before reading a word. On the Dickinson sheet the longest section draws 32px and the shortest 9px, monotonic across all eight.

That it means something is the licence for it, not a bonus. A tick whose length varies looks like it encodes something, and this cabinet refuses decoration asserting a fact it does not have — a gold join on a sheet nobody corrected. Four treatments were drawn against the real stylesheet before choosing, and the one that varied length without meaning was refused for precisely that reason. The lean beside it claims nothing, which is what leaves it free to be ornament: up to 3.2 degrees, walked by the golden angle so the same sheet draws the same rail every time rather than rolling dice on each visit.

The two part company in plain view: the lean goes, the lengths stay. Plain view removes decoration and has never removed information, and splitting one effect into an honest half and an ornamental half is what makes that line drawable at all.

The question was whether small doodles above, below or around the rail would add to the handwritten feel — asked with the risk already named: that it might spend the negative space. Both arrangements were drawn and looked at, and the scatter was the wrong one. A cluster at the head of the rail lands in the top-left corner of the page, which is where a reader’s eye starts and where the skip link appears on focus, and it makes the rail’s decoration the first thing anyone sees. This component is unlabelled precisely so that margin stays empty; filling it spends the thing that made the rail acceptable.

What survives is the version that is not an addition at all. The rail’s rule already reads as a stem, and a stem that simply stops is the only unbotanical thing about it. It now ends in a terminus — two leaves and a bud, continuing the line rather than sitting beside it, pulled back past the nav’s padding to meet the border. Twenty-eight pixels, stroke only, in moss. It occupies margin nothing else wanted and puts no second object on the page.

It wears .qe-botanical, so the growth is inherited rather than restated: the stem draws itself on and the leaves unfurl, both already gated on reduced motion and already kept-but-stilled in plain view. Only the size is new, exactly as with the moth. Because it takes a palette token it follows the ground with no cabinet rule of its own, which was checked in both.

That no gate here measures a drawing against text is the reason it is small. check-overlap.mjs catches text sitting on text, so an ornament in the margin is guarded by nothing at all and could drift onto the measure at some width without a word being said. Keeping the mark inside the rail’s own box means the clearance already measured for the rail covers it too. Anything free-floating would need measuring by hand at every breakpoint, which is a standing cost this cabinet has not agreed to pay for decoration.

No. 1, No. 2 and No. 7 were missing their On this sheet list. They run 2,045 to 3,098 words against 2,495 to 5,018 on the five that had one, so the bar the existing sheets set was already cleared and the gap was an oversight rather than a decision. All eight sheets carry one now. The opt-in stays in the markup: nothing derives a contents list from a word count, because which pages want one is editorial, and the pages that still decline all have a reason on the record.

“After the lede” turned out not to mean “before the first heading”, which was checked before anything was placed. Five of the existing lists carry opening prose after them — the Dickinson sheet nineteen blocks, the Quetelet sheet thirteen. A contents list is front matter for the sections, and the lede is not a section.

No. 2 puts its list somewhere no other sheet does, and that is authored. It has no lede, and all six of its headings are in the 2026 commentary — the 2022 essay above them is unheaded and stays that way, because the note directly above it says Ryan’s words are unchanged and a heading dropped into them would not be. A list at the top would name six sections that begin a fifth of the way down and say nothing at all about the essay, which is the reason the sheet exists. It sits immediately after that note instead, where the sheet says out loud that it is changing hands, and maps exactly what follows it.

macOS is case-insensitive by default. A page at /decisions is decisions.html, and make-markdown.mjs writes a .md beside every page — so it would have written decisions.md, which is the same file as DECISIONS.md. The generator would have silently overwritten its own source with a round-tripped copy of itself, on the first run, with no error and nothing to restore from but git. attributions.md collides with ATTRIBUTIONS.md the same way.

Checked before a line was written, by creating AAA.md in a temp directory and asking for aaa.md. The addresses are /ledger — the site’s own name for that file in every footer it appears in — and /what-is-settled, which is the other file’s own opening line.

The decision log carries one shell block, so make-markdown.mjs and make-search-index.mjs both threw on pre — the throw-rather-than-drop rule working exactly as intended, in two tools, within a minute of each other. The Markdown converter needed more than the tag. Its text handler collapses whitespace and escapes punctuation, which is right everywhere except inside a fence: the first output put a shell command and its continuation on one line with backslashes through the quoting. A fenced block is the one place in Markdown where neither pass is wanted.

The search walker indexes the block rather than skipping it. Skipping was the quieter choice and the wrong one — a command in the decision log is content, and a searcher looking for the deploy hook should find the page that runs it.

The aims run as one sequence across four group headings, which is four ol elements with start on three of them. The converter ignored the attribute, so the page said thirteen aims numbered 1 to 13 and /mission.md said 1–3, 1–5, 1–3, 1–2. A converter that silently drops an attribute is the same drift as one that silently drops a tag, which is the rule that already makes it throw on an unknown element — it arrived by the other door. It now reads start and throws if it is not a number.

Two components were added for the page and both are the bare tier: the one-sentence statement, and the aims list. The statement is not a blockquote — nothing is being quoted, it is the site speaking, and marking our own words as a quotation is the mistake most of these rules exist to avoid. The aims use native ol numbering rather than a counter on a pseudo-element, so a screen reader announces the position, and no colour goes on the marker: a ::marker is invisible to the contrast gate, which is exactly how this register’s own entry bullets once sat at 2.15:1 unreported.

Star Stuff checks ascending order within each series, because one of its collections interleaves Zine and Field Guide numbers on purpose and a single merged sequence would fail a page for doing what it means to do. This site has no series. There is one accession run and the kind chip is a label on it — No. 1 a Reading, No. 2 an Essay, No. 8 a Wall — and they ascend together.

Ported unchanged, that split would have been worse than no gate. No. 8 is a Wall and No. 7 is a Reading, so 1, 2, 3, 4, 5, 6, 8, 7 is two perfectly ascending series. Proved rather than reasoned about: the broken plate was put back, the series split was patched into the tool, and it reported ok · 0 problem(s) on it. Then the split came out and the same plate reported No. 7 follows No. 8. A port is a rewrite when the thing being ported encodes an assumption the new site does not share.

Every rule was made to fail before it was believed. Rule 1 on the restored plate; rule 2 by dropping a card outside every grid; the empty-authority guard by breaking the card selector on the plate and watching it report NOT MEASURED rather than clean — the overlap gate’s empty-pass lesson arriving in a second tool. And --check both ways, because that flag has now been load-bearing and missing twice in this repo: exit 1 with it, exit 0 without.

It does not check the plate against tools/pages.mjs. They agree today and the temptation is obvious, but they are two different facts: the number records when a sheet was accessioned and pages.mjs records the order a reader should meet the pages. Mount Sheet No. 9 and decide it reads best third, and the two diverge legitimately. Coupling them would turn an editorial decision into a build failure.

It also declines to order the unnumbered cards — there is no mechanical answer to where the colophon belongs relative to the register — and it says so in the summary rather than quietly leaving six cards out of the denominator.

A dt that is entirely a link is not an inline link inside a sentence, so WCAG 2.5.8’s inline exception does not cover it — and the ink alone is 23px tall. Padding takes each target to 45px, which is 2.5.5’s figure rather than the 24px floor. It grows into the gap above the description instead of pushing the entries apart, so the list did not get taller to get reachable, and the next target starts a whole description later, so no two can touch. Measured afterwards, because nothing in the seven gates measures a hit area and a number nobody checked is a number.

The padding is also spacing, so the scale is taken net of it. 0.7rem inside an entry and a full --qe-space-3 between two — two to one, which groups the pairs by eye with no rule drawn anywhere. Left uncompensated the list would have started 11px below where every other block on the page starts, and an unauthored deviation makes the authored ones mean less.

The home page’s footer lost the four links and no other page did. On this page they are two paragraphs above; everywhere else the footer is still the way there. One page, one list of the same four things.

check-metadata.mjs has enforced no page may fetch from another origin since the fonts came home, and it did it by scanning HTML attributes. That check would have reported this page perfectly clean. The request lives in a script, and a gate that cannot see the place the failure would occur is a gate that certifies the failure. Three assertions were added, and each was made to fail before being believed:

Every third-party origin our own scripts mention must be named on /privacy. Not forbidden — named. Proved by renaming the host on the privacy page and watching the run go red. It found a real fault on its first honest run, too: the script contacts www.youtube-nocookie.com and the policy named the bare domain, so the policy was one subdomain away from being wrong about its own site.

A facade must stay a facade. An element carrying data-embed-id must have a plain link to the same video beside it — that link is the no-script path and the reason the request counts as chosen — and the served page must contain no iframe at all. Proved by deleting the link, and again by adding an iframe. The second turns out to fail twice: the Markdown converter throws on the unknown tag before the new check is even reached, which is the converter’s teach me or I will drop content rule catching a privacy regression by accident.

The contrast gate caught the other one. The player’s subtitle was --qe-moss on --qe-paper-deep: 6.9:1, clearing WCAG AA and missing the house 7:1 by a tenth. That is the same surface this site already removed a recessed panel over, and it is the third time --qe-paper-deep has done this. The line is --qe-ink now, and the hierarchy is carried by size and weight, which is what should have carried it in the first place.

The drawing is about seven hundred circles. Raw, it is 85KB; compressed, which is how every response here is actually served, it is nine, because markup this repetitive is very nearly free to a compressor. The home page went from 43KB to 143KB raw and from about 7KB to 22KB over the wire. Measured before it shipped rather than assumed either way.

The obvious optimisation was tried and reverted. Defining the leaf and the floret once and referencing them cut 20KB raw — and rendered every cloned leaf and floret in black. Outer stylesheet selectors do not cross into a use shadow tree; only inherited properties do, so .leaf and .floret simply stopped matching while the definitions themselves still looked perfectly correct. That is a fraction of a kilobyte compressed in exchange for a styling path that fails silently, and the comment in queering.css now says so, because the next person to look at seven hundred circles will have the same idea.

A class with no rule renders black, and nothing here catches that. An earlier pass had the generator emitting puff-pale while the stylesheet only knew puff and puff-deep: a third of every flower cluster came out in solid black, and it was found by looking. Star Stuff’s check-classes is the gate for exactly this, and it is the strongest candidate for the next port — the failure it catches is now known to be possible here, which is this repository’s stated bar.

In print the mass becomes an engraving. Every floret prints as an open circle in line, and the tone comes from how densely they overlap rather than from ink coverage — which keeps the design sheet’s claim that these drawings print as line work true, and looks better than the three flat greys the screen fills would have made. Checked under print emulation, where a second fault turned up: the print block sat above the base rule at equal specificity, so the caption printed moss-green. Source order decided it, and the print block now comes last.

--qe-lilac takes the forewing, --qe-lilac-pale the hindwing and the sea of the globe, and both have --qe-cab-* counterparts aliased through both ground switches. The decorative five is a decorative seven now, and the sentence on the palette sheet that counted them has been recounted — a number in prose that nothing checks is a number that goes stale silently, and this one had been true since the site shipped.

The pale one goes darker in the drawer, not lighter. That is the gating-surface inversion arriving in a component with no text in it: in daylight the wash is the recessed thing under the wing, and lifting it in the cabinet would have it out-shouting the ink line. Neither token is for text in either ground, so 7:1 does not apply — there is not a letter anywhere in the drawing.

Black is --qe-ink, so the black inverts. In daylight the outline and the continents are near-black on vellum; in the drawer they are cream. The song says lilac and black and the drawer says otherwise, and the drawer wins, because the alternative is a literal black line on dark brown that nobody can see. The one version that would have kept the black in both grounds was the black card, and it lost on the surface rule above.

The print sheet flattens the wings to line work like every other drawing here, with one exception written into the rule: the continents keep their fill. Flatten those along with the wings and the globe stops being the Earth and becomes a ruled circle.

This is the funniest available failure and the least acceptable one. The drawing quotes a book cover that has a brain in the middle of it. Hand-drawn continents — three separate attempts at them — came out as a pair of symmetrical dark blobs either side of a central meridian, which is a Rorschach card, which is to say: a brain. A picture whose whole point is that the Earth is where the brain was cannot have a middle that reads as the thing it replaced.

So the coastlines are coarse lat/lon outlines put through an orthographic projection centred on 25° west and 12° north — the Atlantic view, the only one where the Americas and Africa are both legible. Points on the far side are pushed out onto the limb rather than folded back across the disc, which is what a continent going round the edge actually looks like. The foreshortening is Earth’s and not ours, and Africa is unmistakable at a forty-pixel radius, which is the size that matters.

It has no tails, after four attempts at them. Hanging straight down they read as legs; swept outward, as a bracket; curled inward, as a moustache; integrated into the hindwing path, as a boot. What actually says moth is a broad fanned hindwing and a feathered antenna, both of which it has. The cover’s tails say swallowtail, which we are not, and the caption carries that reference instead of the drawing pretending to.

One thing the review process itself taught: the first four screenshots showed no wings at all. Every wing is inside a .sprout, which starts at opacity: 0 and unfurls on a delay, so a screenshot taken half a second after load is a photograph of an animation that has not happened. The growth gating is correct and the reviewer was wrong; noting it because the next person to check a drawing here will hit the same thing.

mix-blend-mode: multiply is what makes the mark look stamped rather than printed: the sheet’s own foxing comes up through the ink the way it does through a real impression. On the dark ground it took a mark measuring 7.2:1 down to very nearly nothing — multiply darkens towards the ground, and in the cabinet the ground is the dark thing. Caught by looking at it in the drawer.

This is a blind spot in check-contrast.mjs and not an oversight in it. The tool composites computed colour pairs, so it saw --qe-moss on --qe-cab-paper and reported the figure that pair measures. It has no concept of a blend mode, and it never will without becoming a screenshot differ. It joins the two exemptions already written down: a ::marker, which is not an element with text of its own, and a texture, which varies luminance per pixel. A pass is not permission, and this is the third way to earn a false one.

The fix is a token rather than a second selector, because this is the fourth time that lesson has arrived. --qe-stamp-blend is multiply in :root and normal through both ground switches — the same shape the foxed-paper wash, the shadow a lifted sheet casts and the pale flesh of a specimen all ended up in, each of them a property that differs between the grounds and is not a colour. Two selectors carrying one decision is two selectors that drift.

The mark is in flow, at the head of the provenance line, and the corner version was refused. A stamp belongs in a corner of a real sheet, and an absolutely positioned one lands on the prose — at a narrow width, at 400% zoom, or under a long heading — which nothing in this repository can currently see happen. Star Stuff has check-overlap for exactly that, and the house rule is to port a check when the failure it catches becomes possible here. Not floating the stamp is cheaper than porting the gate, and the accession block reads better as one object regardless.

The stamp belongs in the corner of a real sheet and it is not in the corner of these. An absolutely positioned mark lands on the prose at a narrow width, at 400% zoom, or under a long heading, and nothing here could see that happen. Star Stuff has check-overlap.mjs for exactly it, and the house rule is to port a check when its failure becomes possible. Refusing a design for want of a gate is a reason to build the gate, so it is ported — and the first thing it did was catch the refused design: seven collisions on Sheet No. 2, with the stamp’s own Sheet No. 2 sitting on the masthead’s Queering Earth.

It is not a copy. The reveal step and the browser client were lifted out of check-contrast.mjs into two shared files rather than duplicated. The reveal step is the one that mattered: it is the list of things a page builds at runtime, it is editorial, it grows, and its own comment tells the next author to add to it — so two copies would be a component one gate learns and the other does not, which is the failure the shared entity table already exists to prevent. Both extractions were proved to change nothing: the contrast gate printed a byte-identical report before and after.

Two faults in the imported tool turned up on the way, and both are now fixed here. Its clip check walked upwards from an element’s parent, so an element clipping its own overflowing text was never treated as the thing doing the clipping. And the function that says where a collision is reported page for absolutely positioned blocks — the precise case the gate exists to catch, given no address. Both are recorded in DECISIONS.md, because Star Stuff is not edited from a session in this repository.

Two of its four findings cannot fire here, and that is a house rule showing up somewhere nobody expected. There is not one rendered text element in any page on this site: the botanical art is 535 paths, 381 groups and 160 ellipses and circles, named to a screen reader by aria-label. That is never an image of text, and it means the two faults that motivated this tool upstream — a drawing’s label on another label, and a label on the prose — have no instance to find. They are kept anyway: dropping them would stop the sweep measuring SVG text, and the day a sheet carries a label the gate would report a clean page it never looked at. They were proved by injecting labels, and the file says so rather than letting them pass for tested.

An exemption was removed for exempting nothing. The skip link went on the ignore list on the strength of the upstream note, where its equivalent fires on five of seven pages. Tested here, the list is empty and the count is still zero, because this site’s skip link has no clipping ancestor to be outside of. An exemption that exempts nothing is a named hole a real fault can fall into, and worse, it reads as evidence that somebody checked.

It measures paper too, which the tool it came from does not. Star Stuff hands print collisions to a paper gate; this repository has none, and paper is the medium this house has already been burned by — forty-four of forty-six pages once printed blank. Three passes now: the screen, then US Letter and A4 at the width a sheet actually leaves for content.

The viewport moves, and that is the whole reason the pass is worth having. The contrast gate emulates print media at desktop width and is right to, because it measures colour and colour does not reflow. A collision is a position. In print this stylesheet lets main run the full width of the sheet, so measuring paper at desktop width measures a line length no printer produces. The counts say it plainly: this page is 4,884 pieces of text on screen, 3,567 on Letter, 3,604 on A4. Both papers, because those last two numbers differ — twenty-two pixels is enough to rewrap a line and move a hand-placed mark, so picking one and calling it print would be this pass’s own mistake at a smaller scale.

And the guard against measuring nothing is per pass, which turns out to catch the original disaster. A stylesheet that renders nothing on paper measures zero there while measuring hundreds on screen — and the run reports the page as not measured, naming which passes came back blank, rather than reporting it clean. Forty-four pages printing blank would not survive this gate. That was not why the guard was written, which is the pleasant kind of accident.

Pagination is the honest limit. Print emulation reflows to the width but does not break a document into sheets, so a collision that exists only because two blocks land either side of a page break is still invisible. break-inside: avoid on the accession block is there because pagination is real. Reaching it means pulling text positions out of a generated PDF, which is a different tool and not a flag on this one.

The five existing gates were all green while the stamp rendered as three unstyled spans. Correctly green: check-markup reads tags, check-sitemap reads the manifest, check-contrast composites colours, check-metadata regenerates the derived layer, and check-addresses — the one that already probes the edge — asks about redirects. A cache policy is invisible to every one of them. So is a blend mode, which is the other thing that went wrong today; that one has no mechanical answer and is written down instead.

tools/check-cache.mjs asks one question: can a reader be served new markup with an old asset? Four checks offline — an asset outliving the pages that fetch it, an asset nobody has ruled on, must-revalidate paired with a stale- directive that RFC 9111 makes it cancel, and a rule left behind by a rename — and with --live it asks the site, which is the only thing that actually knows.

It does not flatten everything to zero, and a version that did would have been switched off within a week. The rule is the one this register already stated: staleness is safe only where the stale copy cannot disagree with the fresh HTML. A stylesheet knows the class names and can disagree. A derived index can disagree with the page it cites. A font cannot — it has no knowledge of structure, and a year of caching on the faces is correct. So the plates keep their thirty days and the faces their year.

Which assets are coupled is declared and never inferred, for the reason an entry’s data-sheet is declared: the inference is available, plausible and wrong. Nothing in a file’s bytes or its extension says whether its content knows about markup — queering.js and fonts/newsreader-latin.woff2 are both static files fetched by every page, and only one of them can contradict a class that shipped this morning. Two lists, each entry carrying its reason, and the gate fails on an asset in neither.

It reads our own scripts, because an HTML scan would have missed the file that prompted it. search-index.json is reached by a fetch inside queering-search.js and appears in no attribute on any page. Proved rather than assumed: with the index put back to ten minutes the gate names it, and that finding is reachable only through the script.

All five detectors were made to fail before the gate was believed, and the first attempt at proving the main one was a false pass — the edit meant to loosen the stylesheet had loosened the catch-all rule with it, so the pages moved too and there was genuinely no skew. The test was wrong, not the gate, and a gate whose green run has never been earned is worth nothing. Two findings were then regrouped: one wrong line covering sixty-five plates reported sixty-five times, and one wrong stylesheet policy reported once per page. A defect described fourteen times hides the next one.

Indexed, eight stamps put eight near-identical records carrying the site’s own name into the aid — Queering Earth Sheet No. 4 Accessioned 8 Sep 2026 — which is most of what a search for Queering Earth would then return. That is precisely why .qe-masthead is already skipped, and the reasoning transfers without amendment.

Nothing becomes unfindable. The provenance line beside each stamp says Mounted 8 September 2026 in ordinary prose and is indexed; every sheet number is written out across this register’s own entries. A stamp is the sheet restating its accession in the object’s own form, not a new fact to find. It stays in the Markdown sibling, where there is no ranking for it to crowd — an agent reading /on-being-ill.md gets the transcription the way a herbarium record would carry it.

Search shipped reachable only from the footer, which on Sheet No. 6 is nine thousand words below the fold — and most arrivals here are not by way of the plate. A shared link or a search result drops a reader straight onto a sheet. So it sits in the control cluster at the top of all fourteen pages now, which is also where it belongs by argument rather than by convenience: Search, Cabinet and Plain view are three ways of looking, and that cluster is the instrument tray. It is outside the landmark, so nothing about it reaches the Markdown, the search index or the mirror — verified by regenerating and getting zero changed files.

The first proposal was to hide it in the header illustration, as a magnifier resting on a flower, and it was measured rather than argued about. An a inside SVG is a real link and is keyboard-focusable — that part works. On a 375px phone the masthead art renders at 335px for a 470-unit viewBox, so one viewBox unit is 0.713 css pixels: a flower is drawn at 29 units and lands at 21px, and a link wrapped round one measured 12×12. Reaching 44px would need a 62-unit invisible disc — larger than the biggest element in the whole drawing and 37% of its height, swallowing the blooms either side. Two further faults would have followed: the stems are fill: none and so are not hit-tested at all, and a keyboard user would land on an unlabelled shape in the middle of a picture, which is the same objection this cabinet already makes to hover-gated section marks.

The loupe survived the argument, at a size where it can be seen. A botanist’s folding pocket lens, drawn at 22px beside the word. Sixteen candidates were rendered between 16 and 90px and looked at, which is the lesson favicon.svg already paid for twice, and the drawing was wrong twice before it was right — both times in a way only looking could catch.

Below about 20px two overlapping outlines go muddy, so each part is filled with the paper and occludes what is behind it rather than crossing it. That fixed the mud and left a worse fault, which Ryan spotted and the person drawing it had not: the two parts don’t look hinged together. They were a teardrop case with a plain ring beside it and nothing reaching up to the rivet — two shapes that happened to touch. Both parts are teardrops with their apex at the pin now, rotated apart about that single point, which is what the object actually does. The last correction was that two equal lobes read as a wishbone rather than an instrument: the case is the fatter of the two and the lens ring smaller, as they are in the hand. The word carries the meaning regardless, so the drawing only has to be handsome and unmistakably a thing a botanist would have in a pocket.

And the ground control was 53×25. Found while measuring the other thing. WCAG 2.5.5 asks 44×44 and 2.5.8 asks 24×24, both of the interactive area rather than the visible mark — so it was one pixel over the floor and far under the number this site holds itself to on contrast everywhere else. Enlarging the ink would have made the quieter control loud, which its own note in the stylesheet asks not to happen, so the hit area is grown behind all three instead. Measured after: Search 72×44, Cabinet 53×44, Plain view 102×44, no two overlapping, each receiving its own tap, and the skip link still winning on top where it crosses them.

The first draft highlighted a match the way everything highlights a match: a wash of marigold behind the word. It failed twice, and each failure is one of this cabinet’s own rules arriving in a new component.

It failed the drawer, by measurement. Marigold at 0.30 over --qe-paper composites to 11.2:1 against the ink and passes comfortably. The same wash over --qe-cab-paper composites to 6.94:1, and every alpha above it is worse — because lightening a dark ground moves it towards light ink. That is the gating-surface inversion the stylesheet already warns about for cards, met again from the other side. No alpha of marigold clears 7:1 in the cabinet.

And it failed the seam rule, which is the one that mattered. --qe-marigold is spent on mends: a re-determination, a restored attribution, the provenance line of a corrected sheet. A search hit is not a repair, and a gold mark on a page nobody corrected is decoration asserting a fact — the failure that makes the real mends unfindable. So the mark is a rule under the word and nothing behind it, in verdigris, which nothing else here has claimed. The colour goes on the rule and never on the glyph, and the ink keeps 14.3:1 and 12.8:1 untouched in both grounds with no compositing left to measure.

The new capability is not in queering.js. That file may derive navigation from the DOM and may never create words, and a result is a sentence from another page put onto this one, however carefully. Widening the boundary to fit would have cost the boundary — so the search code lives in its own file loaded by the one page that needs it, and a reader of a sheet downloads neither it nor the 357 KB index. It still writes no label of its own: every word a searcher sees is authored in search.html and cloned, the way the index by sheet clones an entry’s existing chip.

Two entity tables became one, and one page order became one. The new generator needed the same tokenizer, the same comment stripping and the same case-sensitive entity table the Markdown generator has — and that table is the reason: it knows Þ from þ because the sheets quote Old English, and a second copy would be a lookup one tool learns and the other does not. Both now read from tools/html.mjs, and the editorial page order that /llms.txt and the new manifest both need lives once in tools/pages.mjs. The refactor was proved by the freshness gate: it regenerates and compares, so a green run is a byte-identical run.

Thirteen scans went out as JPEG only, with no picture element and no modern format — 2.4 MB of lithographs, every one of them larger than it needed to be. They now ship AVIF and WebP through a fallback chain, and the numbers are not marginal: Dickinson’s manuscript leaf drops from 312 KB to 34 at twice the display size and 12 at once, the Hughes from 330 to 100, the Waterhouse from 341 to 150. A reader on a retina screen now fetches AVIF and never touches a JPEG at all.

The widths are measured rather than guessed, and that is what keeps the ladder short. A plate is displayed at a fixed css width — .qe-plate is 23rem and never reflows wider, .qe-plate-wide is the measure at 34rem — so each scan needs exactly that width and its 2x, and sizes needs no guesswork about viewports. Nothing is ever upscaled past the scan: inventing pixels an 1863 lithograph never had would make the plate look worse and the file bigger, which is both halves of wrong.

Quality was looked at, not inferred from a number. The hard cases here are not the paintings — they are Dickinson’s pen and the engraved caption on Cooke’s fungi plate, because fine line work is what a lossy encoder loses first. Both were cropped at 1:1 against the resized source and compared side by side: indistinguishable, with the blue rules of her writing paper and the words AGARICUS (LEPIOTA.) SISTRATUS. Fries. intact. The measurements agreed — 36.9 to 40.7 — but the crops are the reason we believe it.

The near miss is the part worth keeping. tools/serve.mjs had no .avif entry, so the local server was sending the plates as application/octet-stream and the browser was quietly sniffing them — which works locally, because there is no nosniff there. Production sends nosniff on everything, and a picture element does not fall through to its next source when one fails to decode: the choice is made on the declared type, before the fetch. A wrong content type would have broken every plate on the site with no fallback to catch it. So the variants shipped one commit ahead of the markup, purely to confirm that Netlify types them image/avif before anything depended on it. It does.

tools/make-plates.py is a second Python tool, and CLAUDE.md stopped counting. Its rule was the one Python tool here, and the reason behind that sentence was never the number — it was that something has to rasterise and nothing may need npm install. It is a boundary now: Python touches pixels, Node does everything else. make-images.py draws from the palette; this one re-encodes somebody else’s scan, and folding them together would have muddled both. That is the same correction the queering.js rule already took, for the same reason.

Five new detectors in check-metadata.mjs, all made to fail before being trusted, and one of them is the quiet kind: a scan whose bytes no longer match the hash it was encoded from. Replace a plate with a better scan and forget to re-run the tool, and nearly every reader gets AVIF of the old scan while the JPEG nobody fetches shows the new one — two versions of a specimen on one sheet, and nothing looking broken. Mount it, card it, log it, file it, route it, and encode it.

Every response carried the same header, so 2.4 MB of scanned lithographs was revalidated on every visit — thirteen plates that have not changed since somebody photographed them and never will. They now cache for thirty days. The half of this that was already right is worth naming: the spec’s recommendation for HTML is public, max-age=0, must-revalidate, which is exactly what Netlify was sending. It is written into _headers now so it is a decision we own rather than a default we inherited.

The plates get thirty days and not a year, because immutable is a promise this site cannot keep. That directive tells a browser the body can never change, and the spec reserves it for URLs carrying a content hash. Ours carry descriptive names: annales-lugduno-batavi-1863-tab4-gonystylus-miquelianus.jpg names one plate in one volume, and tab 4 of the 1863 Annales is not going to be re-cut. But almost certainly never is not what the word means, and a better scan could overwrite that filename tomorrow. Asserting a fact we cannot enforce is the same failure as a gold seam on a sheet nobody corrected. Thirty days is a claim we can keep.

queering.css and queering.js get five minutes, and the number is chosen for the shape of a visit rather than the size of the file. A reader moving through several sheets in one sitting revalidates once instead of once per sheet; a bad stylesheet reaches everyone within five minutes instead of sitting in caches with no way to bust it. That bound is the point. This site’s characteristic failure is a palette or print change that breaks pages silently — Star Stuff shipped forty-four of forty-six pages that printed blank — so an unbustable stylesheet is a worse exposure here than a slow one.

stale-while-revalidate is on the plates and deliberately not on those two. The HTML is always fresh, so a stylesheet one visit behind it would render a change that adds a class and its rule together as an unstyled page for that visit. Staleness is only safe where the stale copy cannot disagree with the fresh HTML. And stale-if-error now covers everything that is not must-revalidate, so a Netlify incident leaves the plates showing rather than broken — the two directives are never paired, because RFC 9111 cancels them and the spec calls that out by name.

Not fingerprinting the two shared assets is a decision and not an omission, and it is written up in DECISIONS.md. The clean answer the spec offers — a year and immutable — is available only to a content-hashed URL, which would mean a generator that renames both files and rewrites the reference in twelve pages. Every palette fix would then churn twelve HTML files and make its own diff unreadable, and this register depends on those diffs being legible. The prize is thirty-one kilobytes over the wire. It can be reconsidered later without undoing any of this.

robots.txt has always opened with one sentence: Everything here is meant to be read, quoted, and reused (CC BY-SA 4.0). It was a sentence addressed to humans, in a file read by machines. Every training and retrieval crawler is now named and allowed explicitly, with Content-Signal: search=yes, ai-input=yes, ai-train=yes, and every response carries tdm-reservation: 0. Nothing about the policy changed; it stopped being an inference. A reservation of 1 beside a CC BY-SA licence would have been a contradiction somebody eventually has to resolve, and probably not in our favour.

The answer is yes, and the ask is attribution — so the same pass built the trail. Granting permission to mine a site whose one correctness requirement is credit, without making the credit reachable, would be the generous half of a decision and not the whole of it. So: a Markdown source beside every page, an index at /llms.txt, the whole site at /llms-full.txt, a feed of this register, JSON-LD on every page, and an Agent Skill at /.well-known/agent-skills/ whose main subject is how to cite this place and how to tell a checkable fact from a reading.

Nothing in the Markdown is typed. tools/make-markdown.mjs derives each .md from that page’s own main landmark — the same landmark the mirror reads, so the copy an agent fetches cannot disagree with the page — and derives the feed from this register’s own accessions. The rule it is built around is the bluntest one in CLAUDE.md: two copies of the same words drift, and the accessible one is always the copy that rots. The converter throws on a tag it has not been taught rather than dropping it, because a converter that quietly discards an element is that same drift arriving by another door.

One thing it refuses to flatten. <del> and <ins> stay as HTML in the Markdown instead of becoming ~~strikethrough~~ and nothing, because the one pair on this site is PROSPERO struck through and Miranda, First Folio, 1623 written in — a restored attribution, the kintsugi seam. Markdown has no counterpart for <ins>, so the pair would have read as old new with nothing saying which was which. Losing the difference between what we got wrong and what is true is not a formatting compromise here.

The JSON-LD is authored and not generated, and this is the reason. Every sheet’s label already distinguishes the Maker — who made the thing being read — from Read by, who wrote our reading of it. In the structured data that is about.author against author, and the two are now stated separately on every sheet: The Army of the Upright is by Helen Edgar and Ryan Boren, about a work by Virginia Woolf. Collapsing them would tell every agent on the web that we wrote Woolf. check-metadata.mjs refuses the conflation outright, which makes it the first gate here that checks an attribution rather than a fact about a file. Where a page states no reader, no author is asserted at all.

Two traps paid for, and they were the same trap twice in one afternoon. A comment that names a tag in angle brackets is indistinguishable from that tag to a regex. The note in flower-codes.html explaining that positioning goes on the outer g and the animation on the inner one made the converter refuse the page; then a comment added to every head, explaining the new discovery links, said main in brackets and made a main-landmark extractor match inside the comment and take the head as the page body. The mirror reads that landmark too. Fixed on both sides — comments stripped before parsing, and the brackets taken out of the comments — and written into CLAUDE.md so it is not paid for a third time.

The status code was always right. This was never a soft 404 — a wrong address answered 404 from the day the domain was attached, which is the half of the spec’s error-page requirement that actually matters to a machine. The half it failed was the half a person sees: the body was Netlify’s default, a teal system-font page with no plate, no nav, no ground toggle, and nothing to say the reader had ever been here. 404.html now carries the masthead, both view controls, the before-first-paint snippet, the footer, and all eight cards, so a character out of place lands somewhere you can carry on from.

The drawing is an empty mount: a stem mounted diagonally across the sheet, six leaves thinning as they climb, and nothing at the tip. The herbarium version of a 404 is the sheet whose specimen is not on it. The first attempt was a tall narrow sprig that filled a sixth of its viewBox, which at width: 100% scaled into one enormous vertical stroke — a reminder that these drawings are sized by their box and not by their content. --len is the measured path length, 496, read off getTotalLength() rather than guessed.

Fixing the page introduced the exact fault the spec item leads with. 404.html is a real file, so /404 and /404.html both answered 200 with the error page — a soft 404, at two addresses, created by closing a finding about error pages. check-addresses.mjs --live reported both within a minute of the deploy, which is the first time a gate here has caught a regression the same pass that shipped it. Two rewrites in _redirects serve the page with a 404 status instead; the self-reference does not loop, and that was verified rather than assumed.

Two gates objected to the page for reasons that are wrong for an error page, and both now carry a named exception rather than a silent skip. check-sitemap has NOT_CONTENT, because an error page has no address to list and a listed URL that errors is a coverage fault. check-addresses has NOT_ADDRESSED, and is stricter about the page than the rule it excuses it from: it asserts the file exists at all, that a nonexistent address serves our page and not Netlify’s, and that neither /404 nor /404.html answers anything but 404. Both are allow-lists, so a third entry is a decision somebody makes and reviews.

Three omissions, each argued beside itself in the file rather than here. No rel="canonical", because the page is served at every unmatched address and a canonical would tell a crawler that every wrong URL on the site is one real page. noindex, because the page is thin by design and the spec asks for an explicit policy rather than an implicit one. No social card, because a new sheet needs one and this is furniture that should never be shared on purpose — a broken link that unfurls as a bare link is telling the truth. No provenance line and no contents list either, which is how / and /design already behave: furniture is not a sheet.

Netlify shadows a redirect with a real file, and on-being-ill.html is a real file. That is the whole mechanism, and it is why writing the rule down four times over did not enforce it: the house style says addresses are extensionless, check-sitemap.mjs reported locs still carrying .html — none, every rel="canonical" pointed the right way, and check-markup.mjs had nothing to say. All four were correct. None of them can see what the edge answers. _redirects now carries a forced 301! for each of the ten sheets and for /index.html, enumerated rather than globbed so that a sheet without a rule is reportable by name.

The site was already redirecting the misspelled extension and serving the correct one. Netlify got every near miss on its own — /on-being-ill/ with the trailing slash and /on-being-ill.HTML in the wrong case both 301 here properly, because neither has a file to hide behind. Only the exact-case extension did. That is how this sat unnoticed from the day the domain was attached: the defect was invisible precisely where the file system agreed with the request.

tools/check-addresses.mjs is the fourth gate and the only one that knows what the edge answers. Offline by default, like the other three: it reads _redirects against the pages on disk and reports a missing rule, an unforced one, a wrong target, a stale one, an internal address carrying .html, and a root-relative asset. With --live it goes and asks the site, because a redirect loop is how this file fails and a loop cannot be seen offline. All six detectors were made to fail before the gate was believed — an unforced rule reported as needs 301! — on-being-ill.html is a real file and shadows an unforced rule, which is the failure that would otherwise look green.

Two smaller inconsistencies went with it, both found by Ryan browsing the files locally rather than by any check. The two favicons and the touch icon were root-relative while queering.css and queering.js were not — one rule applied two ways, and now guarded. And the index at the foot of this page pointed at /index, which took a needless hop and was a third address for the plate. The local-browsing question has no fix and that is the trade: nothing on disk is named on-being-ill, so relative links would not help either. Extensionless addresses need a server, which is what tools/serve.mjs is for, and CLAUDE.md now says so instead of leaving it to be rediscovered.

Nothing here said who was allowed to frame a sheet. No content policy, no X-Frame-Options, not from this repository and not from Netlify’s defaults. The usual argument for that header is session theft, and it does not apply to this site: there is no login, no form, and no authenticated action to hijack. The reason it matters here is attribution. Anybody could load a sheet inside their own chrome and present it as theirs — on a site whose one stated correctness requirement is that credit reaches the person it belongs to. frame-ancestors 'none' and DENY now, together, because the older header is what very old browsers still read.

The transport header was Netlify’s default and a year short: max-age=31536000 with no includeSubDomains. It is two years and the subdomain flag now. The spec asks for the subdomain audit before that flag ships, because one HTTP-only subdomain breaks the moment it does — www answers 301 over HTTPS and there is nothing else under the name, so the audit came out clean. No preload, and that is the current advice rather than an omission: the preload list’s own operator now discourages it, browsers auto-upgrade plain HTTP regardless, and getting a domain off the list takes months.

The content policy carries frame-ancestors and nothing else, and that is the trap worth writing down. Every page has an inline <script> in its head — the snippet that applies both stored view preferences before first paint — and inline style throughout. A default-src or script-src added here without 'unsafe-inline' would kill that snippet, and the page would flash the decorated sheet at the reader who turned decoration off, or the daylight sheet at the reader who asked for the cabinet. A hardening pass that looks like an improvement and breaks an accessibility feature is the failure this note exists to prevent. The reasoning is in _headers beside the rule, not only here.

background: var(--qe-card) appeared fifteen times in the stylesheet, thirteen of them with the same 1px rule and the same 2px radius. So the specimen, the mounting label, our own closing commentary and a housekeeping note were the same object to the eye. That is a defect in visual hierarchy and not a matter of taste: a reader cannot tell somebody else’s words from ours by looking, only by reading the label.

Three surfaces now, and a new panel takes the one that says what it is. Bare — space only — for our own commentary, which is the default. Ruled — --qe-card in a hairline box — for an object mounted here. Ruled off — hairlines above and below, no fill — for the ledger’s own housekeeping. .qe-restaged went bare, .qe-untidy and .qe-register-note went ruled off, and ruled now means something.

The third tier is rules and not a tint, and that is a measurement. A recessed --qe-paper-deep panel was tried first. In daylight it drops --qe-moss to 6.90:1 and --qe-rust to 6.21:1 — both under the house 7:1 and both over WCAG AA, so check-contrast.mjs, which gates at AA, would have passed it silently. That is the gap this file has had logged as open since yesterday, arriving as a real near-miss rather than a hypothetical.

The sheet used twenty-five distinct margin values: 2, 2.5, 1.5, 0.7, 2.75, 3, 1.75, 2.25, 2.4, 2.2rem and more. Asymmetry only reads as asymmetry against a norm. This site’s whole typographic argument is deviation — the wonk axis, the leaning slips, the four corner sprigs each drawn for its own corner — and twenty-five arbitrary gaps flatten the deviations that were actually authored.

One line of body text is 1.19rem × 1.65 = 1.96rem, and --qe-space-* is quarter-lines of it. The page was already nearly on that scale by ear: the three commonest values were 2rem, 2.5rem and 1.5rem, and quantizing moved most gaps by one to three pixels. The scale governs the space between blocks; inside a block the hand is still allowed. Paragraph spacing went from 1.35rem to 1.47rem, which is the one deliberate loosening.

Three corner profiles as well, spread so that no two panels a reader sees together are cornered alike: a sheet handled a hundred times does not have four identical corners, and thirteen components carried the same radius on all four of theirs. It is a radius, so it costs nothing in contrast and nothing on paper. That is koko for the price of a token.

The paper wash was position: fixed, so the stains held still as the sheet scrolled past them, which is the one thing a stain does not do. It is absolute now, against a position: relative body, so the foxing belongs to the sheet and travels with it.

The four stain positions were also hardcoded, so every sheet on the site was foxed in exactly the same four places. Nature does not repeat its foxing. They are --qe-fox-* tokens now, stated per page in the markup — the same idiom as --rot on a slip, because a stain is a fact about one piece of paper and not about the site. The front page keeps the original four as the fallback.

Still not texture, and that is now a written rule. The research this came from recommends a noise tile and a blur(0.5px) contrast(1.1) filter. Both are refused. A noise tile varies effective background luminance per pixel, and check-contrast.mjs composites against computed colour pairs — it cannot see a texture at all, so a grain under body text would pass every guard here and fail real readers. Foxing goes on the ground, the tape, the sprigs and the rules, and never under a letter.

Asked directly: should the front page have a contents list? No, and the reason is that it already has one. The plate of eight numbered cards is this site’s table of contents, so a list of the page’s own prose sections stacked above it would be two navigational lists competing on one page — and the one a reader arriving at the front door wants is the plate. It is also the shortest page here, at 1,362 words.

But the two things shipped together and they are separate decisions. Addresses are cheap: an id and a small section mark, no threshold, no furniture. A contents list is furniture and has to earn its place. Conflating them left four pages — the plate, and Sheets No. 1, 2 and 7 — with no way to link to any section on them.

The evidence was sitting on this page. Three entries above italicise the names of front-page sections and could not link to one of them — What grows here, The type plays, and the words do not move, Queering is a verb — while linking happily to a section of the colophon two entries away. The attribution ledger named three more. All six are links now.

Twenty-five addresses added across the four pages. Nothing gained a contents list, and /#what-grows-here is now a direct address for the plate of sheets, which is a useful thing to be able to hand somebody.

The social card for Sheet No. 8 generated correctly and did not appear in the run’s summary, because that summary iterated a hardcoded tuple of filenames nobody had added a line to. Noticed by eye.

Harmless in that direction and not in the other: the same list would print a complete-looking report while a card had silently failed to generate. The summary now enumerates what was actually written and prints a count, so a new card reports itself.

The card also needed its own drawing. The stock sprig puts leaves in lichen and three coloured heads on a stem, which is right for every other card here and exactly wrong for the one plant defined by having neither. A card showing green leaves on the ghost pipe is the card arguing with the page.

The failure this needed protecting from is the silent kind. A typo in an entry’s declared sheet does not throw and does not show: the entry matches no group, vanishes out of the index, and the register above it still reads exactly right. On a page whose subject is corrections, an index that quietly loses one is the fault least likely to be noticed and worst to have.

check-markup.mjs now validates every declared sheet against the pages that actually exist, reports an empty declaration, and reports a group that nothing files into — which catches the same typo from the other end, since a misspelling breaks the match in both directions at once.

Made to fail before it was believed, which is now four for four on this site’s guards: a typo’d entry, an emptied declaration and a stale group were each introduced on purpose and each reported with a line number and a non-zero exit, then reverted and the clean tree re-run.

One layout attempt was thrown away on measurement, too. The index lines were first a wrapping flex row with the kind chip beside the name. A chip plus an entry name is wider than the measure almost every time, so every single line put its chip alone on a row above itself and the index came to 6,726px for no extra information. Ordinary inline flow — the chip opening the line, the name wrapping beneath it — is 5,502px, and is what an index entry has looked like since long before any of this.

This is the longest page on the site — sixteen thousand words — so it got the contents list first. Then it was measured. An accession headline here is a sentence, which is this register’s own design decision written down where it was made, so the labels average 148 characters against 22 to 34 on a sheet. The block came out 1,293px against 319 to 510 everywhere else, and it pushed the first entry three and a half screens down the page.

Both ways to shrink it were closed. Truncating a label is this site’s characteristic failure applied to itself. Writing a short label beside each headline is the second copy that drifts. So the register keeps its section marks and its sixteen addresses, and has no contents list. Removed after measuring, rather than kept because it had been built.

What this page actually wants is an index by sheet, and that is left open rather than approximated. A reader here is usually asking what happened to one sheet, and the accessions are grouped by date. Gathering a sheet’s entries across accessions is a real editorial decision, not a layout one, and the measurement above is precisely the argument against reaching for the nearest component instead.

The style guide asks for 7:1. For this site’s whole life check-contrast.mjs gated at 4.5:1 — WCAG AA — with an entry in the decision log saying so and a line in the working notes warning, in as many words, not to cite the script as evidence of the house standard. A stated principle nobody measures is a wish.

The wish got called in this morning. The recessed panel described two entries above drops moss to 6.90:1 and rust to 6.21:1 in daylight, and passed the guard clean, because 6.21 clears AA. It was built, measured by hand, and thrown away — but it would have shipped on a day when nobody thought to measure by hand.

So the gate moved, in two tiers reported apart. Under AA reads FAIL and means illegible. Between AA and 7:1 reads UNDER and means readable, and beneath the number this site publishes. Both exit non-zero; one count would have hidden which kind had just landed, and they are different bugs. --aa drops to AA and prints a line saying the house target is not being measured — so that an argument for a particular colour is an argument somebody makes on the record, rather than a two-line edit nobody reviews.

It cost nothing, which is the argument for doing it today. All 9,142 text elements on the ten pages clear 7:1 in both grounds and under print emulation, so the gate could be raised without a single colour changing. Raised later it would have arrived tangled with whatever it first caught.

Made to fail before it was believed, the same way the cabinet pass was: the recessed panel was put back and reported 6.90 and 6.21 as UNDER; a 2.15:1 colour was injected and reported as FAIL on the other tier; --aa passed the same tree and said why. Exit codes checked on all three. Then reverted.

Two blind spots are now written into the tool’s own header rather than left to be rediscovered. A ::marker is not an element with text, so nothing measures it — that is how the bullets in the entry above sat at 2.15:1. And a texture is invisible to it entirely, because it composites computed colour pairs. A pass here is not permission for either.

A dark green ground takes the separation out of the colours this site is built from. Moss, lichen, and verdigris share their hue with it, and every accent measures worse against it than against brown for no gain: moss text at 7.1:1 rather than 7.2, rust at 6.2:1 rather than 7.3. Push the green to a real forest tone and nothing but cream clears 7:1 at all. A green ground makes green unusable, which is a plain result and not a matter of taste.

Brown was already the palette’s ground and nobody had noticed. The dark-ground botanical print that prompted this is drawn in cream, sage, marigold, and coral on near-black — which is --qe-ink, --qe-lichen, --qe-marigold, and --qe-coral as they already stood. The five decorative tokens are therefore carried across unchanged and deliberately still saturated. Only ink, moss, and rust are lifted, to 12.8:1, 7.2:1, and 7.3:1. Lifting the decoration to the same target was tried first and turns every flower chalky: 7:1 is a rule for letters, and the brightness belongs to the text.

It earns the mounted plates as well. The scans on these sheets are cream vellum, and cream vellum on dark brown reads as a plate lying lit on a shelf. On green it reads as a plate pasted onto green.

The contrast target has to be measured against the lightest surface any text sits on, and on a dark ground that is not the page. In daylight --qe-card is lighter than --qe-paper, so text on a card gains contrast and the page is the worst case. Invert the palette and keep the instinct — a card lifted above the ground — and text on a card loses contrast instead. Solving 7:1 against a lifted panel drags rust to #eab1a3, a pale pink with no rust left in it.

So the cabinet’s cards are recessed, the ground is the lightest surface again, rust stays an ember, and the lift is carried by the hairline the cards already had. It reads better than the lifted version anyway: compartments in a drawer rather than paper stacked on wood.

Three things were not token flips and are now tokens, which is the same lesson the print sheet taught in 2026: the foxed-paper wash (--qe-wash-a through -d), the shadow a lifted sheet casts (--qe-shade, a near-black that did nothing on a near-black ground), and the pale flesh of a specimen (--qe-flesh). The last was the visible bug: a mushroom’s stipe was filled with --qe-paper-deep, which it only ever matched by coincidence, so on the dark ground the stems of the four mushrooms on Sheet No. 4 became holes and the caps floated. They are cream now, which is what the dark-ground plate does with them.

The ground is a class on <html>, the palette is entirely in queering.css, and the words do not know what they are lying on — the same shape as plain view, and for the same reason. Every hex has one copy: the cabinet values are defined once as --qe-cab-* and the two switch rules carry nothing but aliases of them, so a drifted switch is a missing ground and never a wrong colour.

Three states, one word. With no stored choice the machine decides, from CSS alone and with no script involved, so a reader who has never touched the control still gets the ground their machine asked for. The control then names the ground they can see rather than a preference they have not expressed, and clicking it writes the other one down; the stored choice wins over the machine in both directions. It is the quieter of the two controls on purpose — plain view changes what the page is and keeps the pill, this changes what the page is lying on and gets a word and a hairline. Its label is restored before first paint by the inline snippet in every <head>, which now carries both preferences: deferred to queering.js it would flash the daylight sheet at somebody who asked for the cabinet.

Paper is daylight, always. Both switches are @media screen, so the print sheet cannot inherit a ground and a reader in the cabinet who hits Print gets ink on white.

tools/check-contrast.mjs measured one palette on screen and one on paper. Two grounds are two palettes, so it now runs the screen measurement again under prefers-color-scheme: dark — 2,840 more text elements, and the pass that would notice a cabinet card drifting back up above the ground. The print pass now runs with the dark preference deliberately left set, so a regression that let the cabinet reach paper lands as black-on-brown in the pass that already gates.

A new pass that reports without gating is decoration, and this one nearly was: the failure total it feeds counted screen and print only, so the cabinet column would have printed its numbers and never stopped a ship. Both are in the gate now, and the pass was verified by lifting --qe-cab-ink to a value near the ground and watching 1,675 cabinet failures appear against zero on screen and zero in print.

tools/make-images.py reads the palette out of queering.css with a search for each token name, which returns whichever copy appears first in the file. Two palettes in one stylesheet make that a loaded gun: reordering the sheet would silently turn every social card brown. The lookup is anchored to the :root block that holds the daylight values — a social card is a herbarium sheet, not the drawer it is filed in — and every generated file is byte-for-byte what it was.

The material was gathered from three Stimpunks pages, which is where it lives and is not where it can be cited from. Every quotation was then read in the work it came from — Wise’s post, Ashley’s published PDF, the Internet Archive’s capture of Socol’s 2009 guest post at an address that no longer resolves, Walker’s own page, Helen’s post, and the shedding post. Three originals keep spellings ours would not: Each of us are, pathologizing with a z, and reveling with one l.

One line was cut and then put back, and the rule that cut it was the thing at fault. Language is not a set menu, it’s a buffet is a transcription of speech from a video by The Leftist Cooks, taken down by Ryan and published on the glossary. It came out of the first draft because the caption track returns nothing and the wording could not be re-checked. That is a reason to label a quotation, not to bin it — a spoken line has no printing to go back to, only somebody’s ear, and the citation names the video, the channel and the date. It is on the sheet with the transcription attributed in the caption.

So the ledger’s rule about our own library has been narrowed to what it was actually for. It exists to stop our own prose — a gloss, a summary, a paraphrase — being read back later as somebody else’s wording. A direct quotation Ryan transcribed from a paper he read and published with a citation to that paper is not that; it is a first-hand reading with the source named. Papers get rented and access lapses. Re-reading is an upgrade, not a precondition, and the citation is the thing that has to be right.

The gap was named on the sheet, and then it closed. Jack’s Gender Copia is closed access and is held nowhere we could reach; every hit in our own library was one of our own pages quoting it, which is five witnesses that are one witness. The section shipped its first draft describing her argument and saying in the same breath that it was describing rather than quoting. Then Ryan put a scan of the printed article in the library, and the section was rewritten to quote her: the copia passage from page 3 as a specimen, the array of terms, and the four verbs from the Conclusions.

And the print pages punctuate her differently from the digital text. Our glossary carries masculine=feminine, sex=gender, male=female, nature=culture, symbolic=embodied. The printed pages carry a solidus in every one of those places. Neither is a mistake and this is not a correction of anybody — a slash rendering as an equals sign is ordinary publisher typesetting, the title of the same article is Sex/Gender in both, and the words are identical throughout. It is the Woolf trap in miniature: two real editions, and a quotation that is only right once it is attached to the copy it came from. This sheet quotes the print pages because the print pages are what was read, and the caption says so. Neither text is to be tidied against the other.

The violet mounted is the wrong species. Sappho’s violet is the Mediterranean Viola odorata; the plate is Viola pedata, the American birdfoot violet, photographed for a New York state museum survey in 1918, because that is the public-domain violet plate we had. The caption carries the correction on the page — The plant is the genus, not the poem — and a note at the foot of the sheet asks readers for a better plate. It is logged as open in the ledger, and the fix is one line of markup the day a Viola odorata plate turns up.

The carnations are the wrong colour and there is no right one. A green carnation is a white carnation stood in dye, so no botanical plate of the flower Wilde’s friends wore exists or could — which turns out to be a fair description of the code itself, and the caption says that too. The other two plates, the 1896 pansy and the Hulme lavender, are simply the right flowers.

The sheet’s header drawing is the four plants of the front page’s new border, drawn at size and growing out of one ground line, in the order the sheet reads them. All nine stem --len values are the measured getTotalLength(): three were declared short on the first pass, which is invisible in the finished drawing and leaves the stroke permanently unfinished in the animation.

Violet, pansy, green carnation and lavender were all worn or sent as signals, at times and places where saying it in words was dangerous. Drawing them onto the corners of the plate puts the site’s subject into the site’s own furniture rather than into another paragraph about it. The four are gathered as a set by Sarah Prager in Four Flowering Plants That Have Been Decidedly Queered (JSTOR Daily, 29 January 2020), and the line under the frame credits her for the gathering.

It credits her for the gathering and nothing else. Her specific claims — Wilde’s buttonhole in 1892, Sandburg’s streak of lavender, the violets in The Captive, the Lavender Menace action — are hers and her sources’, and none of them has been read back to a primary here. So the page asserts none of them. If these four ever become a sheet, that reading comes first.

Two things a sheet would have to say out loud, found while looking for plates. The obvious public-domain violet plate is Viola pedata, an American species photographed for a New York state museum survey — not Sappho’s violet. And no plate anywhere shows the flower Wilde’s friends wore: a green carnation is a white one stood in dye, so every carnation plate in existence is the wrong colour. Mounting one under that name would be exactly the quiet generalization this ledger exists to catch. The plates were previewed and held back for that reason, among others.

A four-sided vine drawn as one stretched SVG shears at its corners at every window size. A border on the wrapper never distorts, and four fixed-ratio sprigs pinned to it ride along at any width. Each sprig is drawn natively for its own corner rather than mirrored into place, because a mirrored bottom corner hangs its flowers upside down — fine as wallpaper, wrong on a herbarium sheet. The vine arms end in a tendril curl; a blunt round cap at full stroke width reads as a cut twig.

The frame bleeds out into the gutter by exactly its own padding. Padding it inward instead left the grid 387px, which is under the 417px that two 12.5rem tracks and a gap need, and the plate silently collapsed to one column — a stack of cards, which is a list, not a plate. That is the identical failure .qe-plate-grid’s own comment already records from the day the grid was written, and it still caught this change. Under 30rem the sprigs come in flush with the frame: hung outside it, the violets were clipped by the viewport at 390px.

Every new fill — four flowers and a narrow leaf — went into the print sheet’s line-art rule the day it was written rather than the day it failed, which is the standing debt from the 44 pages that printed blank on Star Stuff. The twelve stem --len values are measured getTotalLength(), not declared, which is the correction the Dickinson entry below already had to make once.

The type plays, and the words do not move is now set per word. The softness rises across the first half and two words sit a degree off the line — and do not move is the one phrase set flat and straight, at SOFT 0, WONK 0. The heading demonstrates both halves of what it says rather than asserting them in an even line. Per word, never per letter: a heading cut into letter spans is one line to the eye and a stutter to a screen reader.

The verb gets the axis at its full extent. Queering in Queering is a verb and queer in To queer is to adapt now run at SOFT 100 against the heading’s own 40. Every h2 here already sets WONK 1, so the deviation a reader can actually see is carried by the softness — in roman, WONK alone moves too few glyphs to register at heading size, which is the same reason the type specimen on the colophon is set in italic.

Two rules in queering.css, reusing the .w idiom the cut-out line already had, and both switched off under html.plain along with everything else decorative. The rotation is dropped below 30rem, where a leaning word collides with its neighbour instead of leaning away from it — the guard the leaning title on Sheet No. 5 already needed.

• The facsimile is CC BY-NC-ND 3.0, not public domain. The first thing mounted here that is licensed rather than free. The credit is the archive’s required form, the site is non-commercial, and the file is unaltered because it was requested from Harvard’s own IIIF service at the width we wanted — so the only derivative was made by the rightsholder’s server. Mounting the 1891 printed page instead was considered and rejected: it is unambiguously public domain and it is the wrong object, because you cannot show what the printers did to the leaf with a picture of the printing.
• The title is Ryan’s phrase, not Dickinson’s — the first exception to the rule that a reading is titled from its specimen. Three titles out of the poem were on the table. None of them holds both halves of the sheet, because her poem only has the swell in it.
• The reply got punctuation and nothing else. Five full stops and five commas; no word, line break or capital touched, including the lower-case continuation lines, which turn out to be consistent across all eight stanzas and are therefore prosody rather than typing. Disclosed in the sheet’s label, itemised in the ledger, and named here — because an unmarked edit of our own text is the same failure as an unmarked edit of somebody else’s, and the site just happens to be the injured party.
• The header drawing is water. One line rises out of the lower left and breaks in a curl at the upper right; one lies almost flat across the foot and curls too, because it is water and not a ground line. The two moths are deliberately together rather than in opposite corners — the poem under them is about reciprocation. Three of the twenty --len values were declared short on the first pass, which is invisible in the finished drawing and obvious in the animation; all twenty are now the measured getTotalLength().

In 2024 Demory quotes her own 2019 sentence and cuts it: to evolve, to repeat, imitate, parody, make new gone, but also to turn gone, and to adapt is to queer — the half that ran the other way — gone. Every fragment that survives is verbatim, and every cut is marked with an ellipsis.

This is the failure mode of this site, performed correctly. ATTRIBUTIONS.md exists because the characteristic error here is not an invented source but a tightened one: a sentence trimmed to fit, with somebody else’s name left attached. What makes Demory’s trim an adaptation rather than a misquotation is that you can see where the scissors went. The front page now says that in as many words, which is a better argument for the ledger than the ledger can make for itself.

The draft arrived as a self-contained page with its own palette written in hex. The typography survived and the hexes did not: queering.css gained a restaging section in which every colour aliases a --qe-* token. The leaning title, the wobbling section heads, the chain of stagers and the staging labels are all hers.

• The chain is an <ol>. It arrived as a row of spans with → between them — five decorative arrows read aloud as words. The order is the content, so it is a list, and the arrows are drawn by CSS with an empty alt string.
• The correction uses <del> and <ins>, which is what Theobald actually did. It arrived as a line-through on a <span> — invisible to a screen reader, which would have erased the one device on the sheet that is an argument about erasure.
• The struck name is moss with a coral rule through it, not coral letters at 0.65 opacity, which is how the draft had it and which fails 7:1 twice over. The colour goes on the strike, never on the letters.
• The plate is the scan already on disk. The draft hot-linked Wikimedia with a JavaScript onerror fallback reading Plate to be inserted — source a high-resolution scan before publishing. Waterhouse’s Miranda was mounted on Sheet No. 4 the same day, alt text and provenance already verified, so the sheet mounts the local file and the fallback is gone.

Both chains carry you, reading this on the end, which is the sheet’s own argument about itself and not a flourish.

Backfilled from the full commit history, cross-read against ATTRIBUTIONS.md and DECISIONS.md so that every correction named here is the one those files record. Linked from the footer of every page, next to the colophon.

It is drawn as a register rather than as a list of version numbers because that is what this site already is: things collected, mounted, labelled, and sometimes relabelled. The four kinds of entry are the four things that happen to a herbarium sheet.

Ryan’s brief was take inspiration from zine walls, and the obvious risk in it is that a photocopied flyposted wall and a Victorian specimen sheet are not the same object. What made it work was noticing that they are: both are things cut out, laid down at an angle, and fixed where they overlap — one with masking tape, one with gummed linen strips. So .qe-wall, .qe-drift, and .qe-cutup are built from the card, the rule, and the accent colour the site already owned, rather than importing a second visual language. No new motion: the drawing at the top grows, the wall does not.

• The wall shipped as one column at every window size. Two 13.5rem tracks plus the gap need 452px and the measure gives 448, so it silently collapsed into a stack of slips — a list, not a wall — and it looked deliberate. Found by reading the computed grid-template-columns in the browser, not by looking at the page. Now 12.5rem.
• The word-drift rail became a column rather than a third exception to the measure. Three states across 448px give 136px each, which will not carry the gloss under each word. The rail runs down the sheet instead. One fewer exception is worth more than one better timeline.
• .qe-cutup uses inline-block, not flex. With flex items the scissored words look identical and copy out as oddifficultimmature.
• .qe-plate-wide came back, exactly as the note about deleting it said it would — Waterhouse’s Miranda is 138cm across, and inside the portrait cap the foundering ship was a smudge in the corner.

A new section, Who reads, and who writes: these sheets are drafted in Claude Code sessions and edited by whoever is on the Read by row. Asked and answered as one honest section in the colophon and nothing in the labels — a label is about the reading, not the tooling, and a tool cannot carry responsibility, which is what a byline is. A site that keeps a ledger of everyone else’s words owes an account of where its own came from.

.qe-poem lives in the shared stylesheet rather than in the page, because poems will recur here and the second one would otherwise copy the first one’s CSS and then drift from it. A poem’s line breaks are the object being mounted: reflowed into prose it is a different poem.

• Every line is its own block element, not a <br>-separated run. A <br> among the lines breaks child counting, so an indent driven by an even/odd rule silently re-indents the whole poem the moment a line is added — the indent is a class the markup states outright instead. And the 1896 alternating indent is a fact about that setting, not a decorative rhythm to apply to any poem.
• A hanging indent on every line, so a line too long for the measure is visibly a wrap and can never be mistaken for a line the poet wrote.
• The type steps down under 26rem, and this is the one place on the site it may: seven of the twenty-four lines wrapped on a 375px viewport at the body size, and a poem read at that ratio is being read as prose. Verified at 24 lines, 0 wraps.
• On paper, a stanza never splits across a page break.

on-being-ill is The Army of the Upright; promises-like-pie-crust is The Die Uncast. The split is deliberate: the filename is the address a reader searching for the source will guess, and the title is what the sheet found in it. An essay written by us is titled by us, so the convention applies only to a reading.

Netlify’s Pretty URLs post-processing is on by default and rewrites internal links to strip .html. So the served page said /coming-to-terms while the repository said coming-to-terms.html, and every canonical tag and sitemap entry said .html. Found by reading the served page rather than the repository, while checking that something else had deployed.

Two ways out were on the table — turn the setting off, or adopt what it produces. Ryan chose to adopt it: canonical tags, og:url, the sitemap, and every internal link now say the extensionless form, and nothing is rewritten on the way out any more because there is nothing left to strip. What it costs is that the filename and the address no longer match, which is now said out loud in the repository’s own guidance instead of left for the next session to discover. The sitemap check was changed to fail on a .html address, and the failure was verified by putting one back.

What the herbarium model means here, what biophilic design survives translation to a website, the five words this site uses for its own parts — specimen, sheet, label, plate, card — the palette with its measured contrast ratios, the type and its deviation axis, plain view, the drawings, and the borrowed plates with their sources.

Until then a sheet’s only internal link was back to the plate, so a reader arriving from a search result had exactly one way onward. Each sheet now ends with Elsewhere on the plate. It sits after the <main> element on purpose: the knowledge-system mirror takes page content from that landmark and nowhere else, so navigation placed inside it would be indexed as that sheet’s own content. No guard can catch that — the markup is valid and the landmark is correct.

The front page became the plate: the sheet the others are mounted on, as numbered cards with a coloured tape strip across the top of each. Card class names were taken from Star Stuff on purpose rather than invented again.

Icons and cards are generated from the shared stylesheet’s own token values, so they cannot drift from the palette. Two traps paid for in the favicon: a doubled hyphen inside an XML comment made the whole SVG fail to parse — it looked fine, served with the right content type, and decoded as nothing — and a favicon is judged at 16px, not at 64, where a tilted leaf reads as a diagonal stroke and anything filled in lichen has too little contrast on paper to be a shape at all.

The home page, the palette as tokens in one stylesheet, plain view as a class over one document, three guard scripts, and the bridge to the knowledge system for finding — never for citing.

• The ground is paper, and it is the other end of the day from Star Stuff. The strongest differentiation from a night-sky site is not the same site in green.
• Fraunces as the display face, because it carries a WONK axis — letterforms that deviate from the norm, exposed as a setting. A site about queering normativity setting its masthead with the deviation axis turned up is the argument made in the type.
• Plain view is a class, never a second page. Helen asked for a plain access version; the alternative — a decorated page and a separate accessible page — drifts, and the accessible copy is always the one that rots.
• Three guard scripts came over from Star Stuff and five did not: markup, sitemap, and contrast pass here; the other five catch failures this site cannot yet have. A check that cannot fail is a check nobody reads.
• The local server roots itself at the repository rather than at whatever directory it was started from. Star Stuff’s copy does the latter, where a wrong-directory run looks exactly like a site with every page missing.

Static files, no build command, no publish subdirectory — the repository root is the site. Deploy previews on pull requests are left on deliberately: a pull request produces the real page at a real address before anything reaches the live site, which is the review path for anyone editing through the GitHub web editor rather than a Claude session.

Creating the project through the API does not wire continuous deployment, and nothing you can do from the API fixes it. The first deploy works, the dashboard shows a green light and Auto publishing is on, and there is no webhook, no deploy key, and no status-check hooks. It looks exactly like success. Three answers were recorded and two of them were wrong before the third stuck:

• Re-sending the repository configuration leaves the deploy key null and kicks a one-off build of current HEAD — a trap of its own, because that stray deploy looks like the link repairing itself.
• A hand-made GitHub webhook pointed at Netlify’s hook endpoint, matching Star Stuff’s exactly. GitHub delivers, Netlify answers 204 OK, and builds nothing. Verified twice. Netlify only honours a hook it created and recorded on its own side. This entry was written as a fix, corrected, and then corrected again — the first correction was also wrong.
• Concluding anything from a short poll of the deploys API. It lags, and it lags by minutes. Give it three before calling a push ignored.

The only fix is netlify init --force in the repository, which provisions the webhook, the deploy key, and the status-check hooks together and needs a browser grant, so a human runs it once. The two commands that tell a wired project from a deaf one are written down, because the dashboard will not tell you.

The site was built for a week against a domain that did not resolve, with two things held back until it did: the host redirect, which would otherwise have sent the only working address to one that does not answer, and the entry in the knowledge-system mirror’s site map, which fetches over the network and would have failed the sync. Both switched on when the domain answered.

Settled the same day the first sheet was drafted, and it is the site’s attribution discipline drawn in the layout instead of left to a reader’s good faith. The specimen is somebody else’s — the essay, the painting, the word. The label is where we speak. Their words are in the specimen; ours are on the label; the line between the two is a fact about the page and not a promise about our behaviour.