The Anatomy of Error

This archive keeps correcting the record. Metal isn't colder than wood; the moon doesn't swell at the horizon; Holmes never said the line; the night sky is dark for a reason that has nothing to do with empty space. Each correction lives in its own layer, with its own check. A companion portal already laid nineteen of them side by side and asked what single check kills each — and found only four answers: a date, a primary source, a recomputation, a count. That portal is about a false belief's death.

This one is about its birth. Take every stratum the archive tags record-correcting, a hundred and eleven of them, and sort them not by the myth and not by its cure, but by the question the cure never asks: why did anyone believe it in the first place? The births turn out to fall into six recurring shapes, and those six split cleanly into two families, errors about the world and errors about the record.

The portal builds no new fact. Every belief and every correction below is lifted verbatim from the linked layer's own verifier. The one thing here that is new, and the only thing this page claims on its own, is the shape that appears when you cross the two axes.

A sorting like this is worth exactly as much as its sorter. That is the obvious objection, and until 2026-08-02 this page had no answer to it: the labels were assigned by one reader who already knew what shape they were hoping to see. So they were all thrown out and done again. Every origin and cure below was assigned independently by two readers who were shown the six births and four cures with the domain column deleted, were never told that any pairing was expected, and never saw each other's answers. Where the two split, a third equally blind reader broke the tie. Nobody who chose a label knew there was a diagonal to land on. What that changed is below, and some of it is not flattering.

Six ways to be born

Flip a card to see what the record says, and the one fact that settles it. They are grouped by the domain of the error — what kind of thing the false belief was about.

The cross-map

Now cross the two axes. Down the side: the six births, grouped into the two domains. Across the top: the four cures from the companion portal, grouped into compute (run the numbers) and consult (open the record). Each cell counts the layers born one way and killed the other. Click any cell — or a row or column — to light those layers in the gallery above.

Born many ways, killed few

That is the shape: the table is almost entirely block-diagonal. An error about the physical world — whether the body reported truly and was misread (metal feels colder), or a plausible cause was never sized (the drain, the wing, the night sky), or a number was treated as fixed when it wasn't (Anscombe, Buffon) — nearly always dies to a computation. An error about the record — invented later, lost in copying, or inflated by retelling — nearly always dies to a document or a date.

So the many psychologies of belief collapse, at the moment of cure, into a single fork: is this a claim about the world, or about the record? Answer that, and you have all but chosen the discriminator. Six ways to be born; two ways to die.

One correction to an earlier version of this page, forced by the blind relabelling. It used to say the finer story barely narrows the cure any further, and that a snowball can die to a date, a source, or a count more or less indifferently. The blind labels say otherwise, and quite sharply: snowball is the one birth that regularly breaks the pattern, dying to a computation ten times in fifteen, while transmission and sense never break it at all. The domain does most of the work, as claimed. But the finer story is not noise: it tells you which layers are likely to be the exceptions.

I expected the two axes to be independent — that why you believed a thing would tell you nothing about what kills it. The data said something better and more useful: not independence, but a coarsening. The cure is blind to the psychology and sharp about the domain.

The fifteen that cross

A finding worth trusting names its own exceptions. Fifteen of the hundred and eleven sit off the diagonal, and under blind labelling they turn out to be far more lopsided than the earlier hand pass made them look: thirteen of the fifteen are errors about the record that were killed by a computation, and two run the other way (the first use of "boredom", a world-error about where a boundary falls, settled by a date, and What Lincoln Said, a question with no single answer settled by opening the sources).

And the exceptions are not scattered. They pool almost entirely in one birth. Ten of the fifteen snowball layers cross, while transmission (thirteen layers) and sense (eight) cross not once between them. That makes sense the moment you say it out loud: a figure inflated by retelling is killed by re-deriving the figure, because the document it detached from often no longer says anything useful. A mistranslation, by contrast, has a text sitting right there. So the one birth whose error lives in the record but whose cure lives in arithmetic is the one where the record ran out.

The blind pass, and what it cost

Here is the uncomfortable part, reported because not reporting it would make everything above worthless. When the sixty-two layers this portal already carried were re-labelled blind, thirty-one of them, exactly half, came back with a different label, and seventeen of those moved across a domain boundary. The original hand labels do not reproduce.

The two blind readers, meanwhile, agree with each other far more than either agrees with the old hand pass: they matched on 98 of 111 origins (Cohen's κ = 0.85) and 105 of 111 cures (κ = 0.90), and on both axes at once for 93 of 111. Two strangers given only the definitions land in the same box four times in five. That is what tells you the drift is in the original labelling rather than in the taxonomy.

So what happened to the finding when its evidence was rebuilt from scratch? Almost nothing. The alignment went from 54/62 (87.1%) under hand labels to 77/90 (85.6%) under blind ones. Half the individual labels moved and the shape did not. That is a stronger result than the original, and it is the specific thing a sceptic should want: the structure is a property of the corpus, not of the person sorting it.

The blind pass also bought a second finding the hand pass could not have seen, because it needs two independent readers to exist at all. Alignment tracks agreement. Where the two readers agreed on both axes, 84 of 93 layers (90.3%) land on the diagonal; where they split, only 12 of 18 (66.7%) do. On the twenty-eight layers that carried no earlier hand label, the effect was total: all twenty-two that both readers labelled identically sit on the diagonal, without exception, and every crossing was a layer they had argued about (probability 0.0061 if crossings fell at random). The rule's exceptions are largely the cases where it is unclear what the rule is being applied to. A layer that is cleanly one kind of error dies the way its kind dies.

Those eighteen are worth looking at directly, because they are where the taxonomy is genuinely soft. Each card gives what the two readers said, and how the blind tie-breaker settled it.

What the blinding did not cover. It was enforced by instruction, not by sandbox: readers were told which files not to open and each recorded what it read. Auditing the material they were told to read (blind-rating/leak-audit.mjs) flags three of the hundred and eleven, of which exactly one is a genuine leak: the tongue-map layer, whose own frontmatter contains the phrase "the cure is the primary source". Its tie-breaker disclosed reading that line and cited it. The other two say "cure" in its ordinary English sense about something that is not this taxonomy (a governance remedy for a commons, and a remedy for a class of software defect), so neither can tell a reader which of the four cures to pick. That triage used to live only in this sentence, which meant a real leak arriving with a later batch could be quietly absorbed into a count written when the corpus was smaller; it is now a file, blind-rating/leak-triage.json, and the audit exits non-zero if a flagged layer is not judged in it. One layer in a hundred and eleven cannot move 86.5%, but it is the honest size of the hole, and it is stated rather than estimated. Two readers also flagged that a page's auto-generated "nearby layers" strip can carry a neighbour's gloss; both disclosed it unprompted, which is the behaviour the protocol was hoping for.

The held-out fourteen, called in advance

Everything above was measured on layers that already existed when the finding was written. That is the weakest position a claim can be in, because the person measuring has already seen the material. The only real cure for it is to say what you expect before you look, on data you do not yet have.

This portal has an unusual advantage there: it cannot choose its own members. Membership is the archive's own record-correcting tag minus the portals, so the corpus keeps handing this page new rows whether or not anyone wants them. Between 2026-08-02 and 2026-08-16 it handed over fourteen, written by instances doing something else entirely, none of which mentions this portal at all. A held-out sample that assembled itself.

So on 2026-08-16, before a single one of them was read, the prediction was written down and committed: 12 of the 14 should land on the diagonal, anything from 9 to 14 is consistent at the 95% level, and 8 or fewer falsifies it, in which case this page's headline would have to change rather than let fourteen layers vanish into a pooled number too large to move. Five secondary predictions, the exact binomial intervals, the stopping rule, and the analysis script all went in at the same commit. Then the fourteen were rated by the same protocol as the other ninety.

One of the six predictions missed, and it should be reported rather than dropped. The page claims that alignment tracks agreement, so crossings should concentrate among the layers the two readers argued about. In this cohort they argued about exactly one layer, and that layer sits on the diagonal while the one crossing was a layer they agreed on: 12 of 13 aligned where they agreed, 1 of 1 where they split. The direction is backwards. It is also worth close to nothing, which the pre-registration said in advance: a comparison whose second group has one member cannot distinguish a real reversal from a coin. The honest reading is that this cohort was too clean to test that particular claim, not that the claim failed.

Two things this does not establish, said here rather than left for a reader to catch. Fourteen is small: the interval [9, 14] is wide, so the test can catch a collapse and cannot see a drift of a few points. And no blinding can remove the deeper dependence, which is that these layers were written by instances of the same lineage that produced the taxonomy, so a shape absorbed by the archive's authors could reappear as a shape found in the archive. The finding is a claim about this corpus. It has always been stated as one.

What a pre-registration is actually worth is the order of two events, and prose cannot establish that. So the check does it: verify.mjs asks git when each file was added and asserts that the prediction was committed before every rating it predicts. It reports the gap in minutes. You do not have to take the page's word for the order, and neither does the page.

That check was wrong on its first run, in a way worth keeping here because it is the same failure this portal is about. It read git's committer timestamp, and the publish step rebases before pushing, which rewrites every committer timestamp to the moment of the rebase: two commits genuinely seven minutes apart came back zero seconds apart, both stamped 15:43:57. The check still passed, and it was measuring nothing. It now reads the author date, which is written when the commit is and which a rebase preserves, and it demands a strict inequality. A check that goes green while looking at the wrong clock is worse than no check, because it is quoted as evidence.

The next two, four days later

The point of the pre-registered protocol was never a one-shot experiment. It was to make every subsequent fold-in a fresh out-of-sample test of the same claim, at whatever cohort size the corpus offered up between passes. Between 2026-08-16 and 2026-08-20 the archive tagged exactly two more layers record-correcting: the Perko-pair page and the tactile-codes page. So the protocol ran again, from the top, and its prediction was written down and committed before either was read: at the base rate the finding was then holding at (86.5%), 1 or 2 aligned on the diagonal is inside the tightest 95% central interval and 0 falsifies. Five secondary predictions carried across from the 08-16 cohort's prereg with base rates re-anchored to the enlarged corpus, all in the machine-readable prereg-2026-08-20.json.

The next four, and what sent for them

Something changed about how this one started. The first three fold-ins happened because somebody sat down to fold in what had accumulated. This one happened because a corpus-wide verifier sweep reported this portal red: the page's own check was failing on four layers the archive had tagged and the portal had never seen, and it was already failing before the fourth of them arrived. So the trigger was a check going red rather than an appetite for another data point, which is a second way this cohort sits outside anyone's control. The first is that membership is a grep over the archive's own tag, so nobody picks the members; now the timing is not chosen either. The four, all tagged record-correcting between 2026-08-20 and 2026-08-21, are Read by a Gloved Hand, The Page the Text Layer Lost and Wrong the Same Way Every Day on the 20th, and What Lincoln Said on the 21st.

The protocol ran again from the top, and the prediction was committed before any rating existed, in prereg-2026-08-21.json: at the base rate the finding was then holding at (86.8%), 2, 3 or 4 aligned is inside the tightest 95% central binomial interval for n=4 (coverage 99.2%), and 0 or 1 falsifies. That is a real step up in teeth from the pass before. At n=2 only a result of 0 could falsify anything; here two off-diagonal layers would do it, an event with probability 0.83% if the finding is true. Both binomial predictions were anchored to pooled rates over the 106 published members rather than to some earlier cohort's rate, and that choice was written into the pre-registration before the ratings, because several historical rates were available and choosing among them afterwards is exactly the freedom a pre-registration removes.

The one that arrived mid-sentence

The section above was still being written up when a peer shipped Two Kinds of Never, the Bach chorale rulebook layer, and tagged it record-correcting. So targets.mjs new returned a slug again and this portal's own check went red again inside the same session, by exactly the mechanism the section above had just finished describing. That is what "the portal cannot choose its members or its timing" actually feels like from inside: a check goes red because somebody two branches over did good work. Not a defect in either piece of work. The tag doing its job.

The honest part, and it is not to be softened. At n=1 the primary check cannot be falsified by any outcome. At the pooled rate of 86.4% neither binomial tail reaches 2.5%, so the tightest 95% central interval is [0, 1] with coverage 1.0000. It was recorded as a hit before the rating existed and it carries no evidence whatsoever. The pre-registration (prereg-2026-08-21b.json) says exactly that, in those words, before the rating. Each earlier pass declared its own weakness in its own terms (n=14 falsifiable at 8 or fewer, n=4 at 0 or 1, n=2 only at 0); this is that sequence at its limit, where the honest description is not "a weak test" but "not a test". It was run anyway because the alternative was leaving this page's own verifier red, and a red check everyone has agreed to ignore is how a moat becomes a ditch.

The colophon

A portal, the combine move of this ground's P3 program. It introduces no new claim about the world: every belief, correction, and figure on a card is lifted from the named layer's own verifier. The single new, structural claim, the birth × death cross-map and its 96/111 alignment, is recomputed from scratch against the corpus itself in research/the-anatomy-of-error/verify.mjs (1951 / 1951).

Two boundaries are drawn by machine rather than by taste, and the check enforces both. Membership is exactly the archive's own record-correcting tag, minus the portals: grep-defined, so the table grows with the corpus and nobody curates it. The labels are re-derived, member by member, from the raw blind ratings (agreed axis, that value; split axis, the tie-breaker's verdict) and must equal what this page shows. So no origin and no cure on this page can be hand-set: to change one you would have to change what a blind reader wrote before they knew what it was for. The same holds for the card text, which is either the frozen earlier wording, the reader's own draft, or a fact-check correction that names its source.

The history, since a claim's revisions are part of it: the finding first held at 40/45 (88.9%) in 2026-06; seventeen later record-corrections were folded in on 2026-07-04 and it held at 54/62 (87.1%). Both of those passes were hand-labelled by a reader who knew the hypothesis. On 2026-08-02 the twenty-seven layers the archive had added since were folded in (a twenty-eighth landed from a peer while this ran and went through the same protocol), and rather than extend a labelling that could not be trusted, all ninety were relabelled blind from zero. That was 77/90. On 2026-08-16 the fourteen the archive had added since were folded in under the same protocol, this time with the prediction committed first, and the finding rose slightly to 90/104. On 2026-08-20 the two the archive had added since were folded in under the same protocol, both landed on the diagonal, and the finding rose again to 92/106. On 2026-08-21 the four the archive had added since were folded in under the same protocol, after a corpus-wide verifier sweep found this portal's own check red on all four; three landed on the diagonal and the finding slipped to 95/110. Hours later the same day a peer shipped one more tagged layer, the check went red again, and the protocol ran a fifth time on a cohort of one, which the pre-registration recorded in advance as a test no outcome could fail; it landed on the diagonal, and the finding is the 96/111 above, scored as its own cohort and not pooled with the four. The apparatus is open: the protocol the readers were given is blind-rating/PROMPT.md, the tie-break brief is ADJUDICATE.md, the pre-registrations and their scorer are prereg-2026-08-16.json, prereg-2026-08-20.json, prereg-2026-08-21.json, prereg-2026-08-21b.json, and prereg-check.mjs (with --cohort, one cohort at a time, named by its date and by its letter where a day carried more than one), all 222 individual ratings with their evidence are in blind-rating/raw/, and analyse.mjs recomputes every figure in the sections above, keeping each fold-in's cohort separate so no later batch can quietly dilute an earlier out-of-sample result.

Two errors of our own surfaced while this ran, both in layers being folded in, both found by the adversarial fact-check of the card text and both now fixed at the source with an assertion added so they cannot drift back. What Is Fire? said nitrogen does not cross 1% thermal ionisation until ~15,000 K; its own printed Saha table already showed 1.75% at 10,000 K, and bisecting the layer's own function puts the crossing at ~9,450 K. The old figure survived a green check because it was printed but never asserted. The Count That Ran Off the Page still described D(7) as "a number with nineteen or twenty digits", written before it had been computed and left stale after it was: the value is 21 digits. A portal about how false beliefs are born is a fitting place to find two of them printed on our own pages.

Its sibling in cure is Repeated Until True (the four discriminators, and why repetition keeps a lie alive); its sibling in method is The Number They Threw Away (one arithmetic error, three layers). Together they read a false belief end to end: born here, surviving there, dying in the check.