Method

Three reviews named the same gap: depth without scaffolding, so the hardest machines read like joining a conversation already in progress. The answer is a crash course on the pages that need one, decided by counting: a page stands on a prerequisite when it uses an idea another machine here is about. The shared half is written once per topic and inherited. A machine gets the grounding of every topic it already declares in the roster, so a new machine tagged errors gets its onramp without anybody writing one, and a grounding that serves a single machine is not a grounding. Each one ends by naming a machine where the idea can be pressed rather than read. All of it is collapsed and sits after the lede, because support that measurably helps a novice hurts a reader who already has the schema. There are three answers and not two: its own paragraph, the topic grounding alone, or no crash course at all, which is what a bit answers structurally by being already that size. Folding the last two into one key stripped the onramp from twenty-one pages while three gates stayed green.

The honesty ledger at the foot of each machine is where the stopping point gets written down. Interleave says the interleaving is real and the Reed-Solomon arithmetic is not performed, because it counts what would be corrected instead of correcting it. Core Memory says the switching rule is modelled and the magnetics are not: no B-H curve, no domain walls, no coercivity. Punch Cards says the card's outside dimensions are the real ones and the hole pitch is approximate. A ledger saying "some details are simplified" would be worth nothing. The test is whether somebody who knows the real machine could read the ledger and agree with it.

A standard first, where a body defines the thing: ECMA-130 for a compact disc frame, the GS1 General Specifications for a barcode, ITU-R M.1677-1 for Morse timing. Then a patent, where an inventor had to describe a mechanism to claim it. Then the paper the result was published in. Then an institution: a museum record, a manufacturer's manual, a scanned document in a library. Then contemporary reporting, for what was known at the time. Everything else comes last, including an encyclopaedia, which is a fine place to start reading and a poor place to stop.

Banning them would produce a worse dishonesty: dropping the citation and keeping the claim. They are registered instead, each one naming what it carries, why nothing better does, and a date to look again. Some will never improve. The origin of the word nibble is folklore, and an encyclopaedia is the only place any of it is written down. One is correct as it stands: PageRank cites an encyclopaedia for the version of the formula that appears everywhere the paper is cited, which is a claim about the secondary literature. A primary source would make that citation wrong.

CRC cites a 1961 paper that would not come out from behind its paywall. Fifteen Planes cites Marsaglia's 1968 note, which PNAS, PubMed Central, NCBI and Europe PMC all refused in different ways. Masking cites two papers by DOI, checked against Crossref for title, authors, volume and pages, and says plainly that neither was read. Interleave carries a pair of published burst figures that appear in none of the five documents it cites, and says so instead of dropping them.

Any figure printed here has to be produced by a test that computes it from something more basic, independently of the code the page runs. That is the rule for a figure with one right answer. Fifteen Planes derives its recurrence from the multiplier and the modulus, then checks every triple against it. Interleave derives the correctable scratch width from a four-frame delay and four parity symbols, and reports 2.6 mm against a published 2.5 mm instead of rounding until they agree. Checkout rebuilds all fifteen barcode symbols against an independent implementation, then reads the drawn symbol back with a real decoder.

A measurement is a different kind of claim and gets a different rule. Spectre times a cache and False Sharing times two threads sharing a line, and the number a reader gets depends on their processor, their browser and whatever else the machine is doing. Nothing can derive those from first principles, so they are instrumented instead: the page says what was measured, what it did about the noise, and where the measurement stops being real. False Sharing rejects overlapping trials, reports a median and a middle half rather than a best time, and states plainly that it cannot see the physical cache line it is arguing about.

Everything above is a promise. This is the check that keeps one of them, running here rather than described: give it a sentence and it will show you the quantities it found, the runs of digits it decided were names, and whether anything in the tests would argue with what is left.

It is a second implementation, deliberately and carefully. The gate is a Python program in the deploy chain and cannot run in a browser, so the algorithm above is written twice. A decision written twice is how a change to the real one leaves every test green, so the split is drawn where it can be defended: the exemption families, the number-word tables, the two run patterns and the thresholds are not duplicated at all, they are generated out of the gate itself, and only the algorithm is rewritten. The two are then compared on two thousand sentences built from fragments with a fixed seed, and on the values, which is the only thing that has to agree.

That comparison earned itself immediately. The thousands separator in a number like 49 500 is a narrow no-break space, which in a terminal is indistinguishable from an ordinary one. Copied across as an ordinary space it welded “49,500 2026” into 49,500,202. All fourteen of the worked examples still passed. The swept corpus found it on its first run.

What it leaves out: the debt list that carries a claim somebody has decided to owe rather than fix; decimals, which the gate holds whole and separately; the extraction of a page’s own words from its shared shell, which is why a machine’s footer and era pill are not read as its claims; and the checks for a page vouching for its own verification. The two test corpora in step four are illustrations, not the real files.

Two things it gets wrong that this page can demonstrate on itself. It cannot tell a quotation from a phrase in quotation marks, so the sentence above about a ledger saying “some details are simplified” is held by a test as though it were a citation. And it cannot tell a live figure from a historical one: the thirty-nine claims a scoping change found is what happened that day, re-running it now gives twenty-seven, and the gate would ask the same question of both.

The exemption list is incomplete by construction. Each of the ten families was added because a real page was blocked by it: a DOI read as 203.0198, a patent as 3,886,525, a bit pattern as 1,010, a hostname as 1,401, a RAND memorandum as 3,420, an IBM 7090 as 7,090. The next machine can carry a kind of name nobody has met, and until somebody teaches it, the gate will stop a deploy over a catalogue number.

Every source that can be reached is fetched, read and archived, and one that cannot is recorded anyway: its metadata, the routes that failed, and an explicit note that nobody here has read it. Every quotation is checked against the archived copy and against the page, so a citation cannot drift on either side without something going red. Where a document is a scan, the quotation matches the scan's own text, spelling and typographic quirks included. Tidying a quotation is how a misattribution starts.

Longest Chain said bitcoin's target was around ninety zeros' worth of difficulty, and that the toy here was a millionth of a millionth of real. The first was wrong by seventy zeros and the second by eight orders of magnitude, and both sat on the page for months, because every check above looks for digits and neither sentence contains one. A quantity written in words is converted and pinned like any other number now.

That check used to read every test file at once, so a number counted as covered if it appeared anywhere in any of them. Putting the wrong sentence back on the page turned nothing red, because a ninety sitting in the Fringes test was pinning a claim in Longest Chain. A pin that any unrelated machine can satisfy is not a pin. Scoped to the file that belongs to the machine, thirty-nine claims turned out never to have been covered at all.

A gate loads every machine and compares the parts they share, because three different renderings of the same header each look fine on their own. Another walks every drawn surface to confirm it is visible somewhere. Another confirms every control does something. None of this replaces looking at the page, and all of it catches things looking does not.

Every machine has two plates: a card on the chronology and a share image, which are the same drawing at two sizes. Both are measured. A plate has to be at least three per cent ink, so that it is a drawing and not a caption, and it has to use at least seventy per cent of its frame in each direction, so that it does not sit in one corner of it. Width and height are measured apart rather than multiplied, because a product hides a failure in one direction behind success in the other. CPU Cache used 58 per cent of the width and 81 per cent of the height, which multiplies out to 0.47 and reads as merely lowish, while what a reader actually sees is the right third of the frame empty. A planned machine's plate is type, and that is correct: the machine does not exist yet and a drawing of it would be a lie.

A rule being present in a stylesheet says nothing about what a browser draws. The shared stylesheet gives links their colour by listing class names, which makes it an allowlist, so any context nobody thought of falls through to the browser default. Machine prose was such a context: cross-links rendered at 2.07:1 against the page where the standard asks for 4.5. Colour is now sampled from a ring of real pixels around the glyphs, on every page, at desktop and phone widths, in both light and dark.

Every other check here asks whether a region has a name. None asked whether the name was true. When the checks table stopped scrolling sideways, fourteen machines went on telling a screen-reader user it was scrollable, and the label was present, unique, and describing something that had stopped being the case. A name that no longer matches the element is worse than a missing one, because nothing looks wrong.

Masking shipped a threshold line drawn behind the bar it crossed. The markup was right, the value was right, and on most columns you could not see it. Every check passed: the element existed, was not hidden, had a size, had a background, and had contrast against whatever was behind it in the document. None of them asked whether anybody could see it, which is the only question that matters. One does now, across every drawn surface.

The column carrying each verdict sat outside its scrolling wrapper on 67 of 71 machines at desktop width, and on every one of them at phone width. Nothing was missing and no rule was broken: the wrapper scrolls, it takes focus, it is labelled, and a wide table is an explicit exception to the reflow rule. A reader on a phone saw a list of claims and no answers. Conformance and usefulness are different questions and only one of them was being asked.

Safari's Reader and Firefox's Reader View are the same library, and it throws the instrument away and keeps the prose. That is a hard case for pages that compute most of what they say after loading. The real library is run against the real parsed document, in Chromium every deploy and in Firefox and WebKit on a schedule, and a page fails if the opening, the crash course, the honesty ledger or the sources do not survive, or if the controls come through as stray text.

Where a table reports a verdict, the verdict is a word. Where a machine marks the row you are looking at, it changes the weight as well as the colour. Accents are chosen against the published set for measured separation and for contrast against the page, and the separation is checked under simulated red-green and blue-yellow colour blindness rather than only under ordinary vision, because a palette that is distinct to one of those and not the others is not a palette.

A real screen reader is driven now, and this paragraph used to say it was not. NVDA reads every machine on a Windows box kept for it, and what it says is written down rather than inferred. On the twenty-ninth of August 2026 that was 4,931 tab stops across 104 pages, and not one of them was announced without a name. The count moves as machines are added; what does not move is that it is measured rather than assumed. That is worth having because it found something no tree-based check could. When focus crosses into a container, NVDA announces the container and drops the control it landed on, so the control is silent. It is the reader's bug and not the page's, which took an eight-line test page to establish rather than an opinion: every way of naming the container does it, a container with no name does not, and a bare list does. The only markup that avoids it is worse for a reader than the bug.

What that still does not establish is the part that matters most. Nobody who uses a screen reader every day has been asked. Capturing what NVDA said tells you what was said, not whether it was useful, correctly ordered, or maddening. Automated rules decide roughly a third of the standard and the rest is judgement nobody has automated. So this is evidence that a large class of failure is absent, and it is not a claim that a person using assistive technology can do everything here comfortably. If you find somewhere you cannot, that is a defect and it is worth telling me about.

A barcode shipped that decoded to nothing, because an encoding table was copied from one language into another and drifted. A card renderer printed a year that contradicted the roster beside it. Interleave took the constants it computes with from an encyclopaedia for weeks while the gate that checks sources reported everything fine, because that gate could not see the shape its sources block was written in.

Five more of the same shape turned up in a single day. Score Reel's carry struck no chime at all, so one beep in five arrived with no pitch and louder than the four around it, on a machine whose sound is the argument. Morse ran ten units between words where the standard it cites clause by clause puts seven, and nothing could see it because those constants appear nowhere except the audio schedule. Twelve machines that are built told a screen reader they were not built yet, in alt text written while they were planned and never read again. Fifty share plates had never been measured by anything. Punch Cards said a column never holds more than three holes, two paragraphs above describing a card with twelve holes in every column. All of it was green everywhere at the time.

The newest one is a different shape: a check that existed and was not run. A machine goes live on the studio, and the other seven sites derive their footer from the same roster, but derivation only reaches a site when that site deploys. So a machine can be live in one place and absent from seven others, with every tree clean and every gate green. A check for exactly that had been written the last time it happened, and it was in no chain and no document, so nothing ran it. Then an outside reader looking at a cached copy reported the studio’s own counts as wrong. They were not; production had been right the whole time. The only thing that could tell those two apart was the check nobody was running. It runs after every deploy now, and the first time it ran across all eight sites it found seven of them a machine behind.

All three failures have the same shape: a check that cannot see something reports the same green as a check that looked and found nothing. So a defect here is not finished when the page is right. It is finished when the thing that would have caught it exists, and has been proven to fail on the original defect before anyone trusts it.

The page as it stands is the current version, and old errors are not memorialised in the running text. Where two sources disagree and neither is clearly better, both are printed and the disagreement stays: Interleave carries two sets of burst figures and picks neither, and Core Memory prints one wording of a famous Forrester line while showing the museum's shorter one beside it. Where a claim was retired, the page says what it used to say and why it stopped. If something here is wrong, say so, and bring the document.

The machines are built by the author. AI tools assist with drafting, review and checking, as they do on RoamingPigs, and the same rule holds: nothing is published that a tool produced end to end without the author's authorship and review. The card and social images are not AI-generated. Each one is drawn by that machine's own renderer from that machine's own data, which is why a card can go stale and be caught by a gate. None of this moves the bar. Every number still has to be derived by a test, every source fetched and read or recorded as unread with what was tried, every quotation matched to the document it came from, and every gate green before a deploy. All of it is under the author's name and the author's accountability.