Artificial Wasteland · the ground audits itself
This site prints 12,886 claim-shaped numbers in its own sentences. For 742 of them the digits appear in no file the page points at as its working: nothing here is in a position to notice if they drift. Then we changed every number in a layer's sentences and ran that layer's own check: on 55 of 62 layers it passed anyway. Then we drew 150 figures at random and had each one fact-checked against outside sources. Three published claims were wrong. This page is the sweep, the experiment, and the corrections it forced.
On 2 August 2026 somebody here was fact-checking a card and noticed that What Is Fire? said nitrogen does not cross one percent thermal ionisation until about 15,000 K, while the same page's own printed Saha table showed 1.75 percent at 10,000 K. The real crossing is near 9,450 K. The layer's check had been green the whole time.
It was green because the wrong number lived in a sentence, and nothing had ever asked the check to compare a sentence to anything. The apparatus next door, The Check That Cannot Fail, measures whether a layer's check could go red at all if the layer were wrong. It cannot see this. A number a check prints and never asserts is invisible to it.
One accident is not a measurement. So: the sweep.
A page here is mostly not prose. It is an instrument, and an instrument is full of numbers that are not claims: a slider's tick marks, an axis label, an option in a menu, CIE-1931 in the name of a colour space. Taking every numeric token out of the visible text of forty layers found 3,547 of them, about 61,000 across the corpus, and nearly all of it was furniture.
So the population is narrowed twice, and both narrowings shrink it rather than flattering us. First prose, not text: a number counts only if its nearest block ancestor is a tag people write sentences in, and only if that block reads like a sentence rather than a label. Headings, table cells, form controls, code blocks and SVG are out. Second claim-shaped, not numeric: inside prose, a token is a figure only if it carries the marks of a measurement, which is a decimal point, four or more digits, a percent sign, a unit within reach, or a hedge in front of it. A bare year is dropped, because this corpus dates everything and a date is a different kind of claim.
What survives is 12,886 figures across 700 layers, about nineteen a layer. Each one was then asked a single question: do these digits appear anywhere in the material this page points at as its working?
Orphan does not mean wrong. A check that computes 9,450 and prints it without asserting it leaves no digits behind; a source may be a paper cited by name; prose says "nine thousand" as readily as 9000. Orphan means one thing exactly: no file this layer points at states this number, so nothing mechanical here can notice if it drifts. It marks where to look.
"The digits appear in the evidence" is a weak relation, and it gets weaker the shorter the number is. A two-digit figure like 4.5 has a significand of 45, and a research directory holding two hundred thousand numerals will contain 45 somewhere by sheer volume. So every figure was re-run through the matcher as a decoy: same shape, same order of magnitude, different digits. A decoy is a number the layer never claimed. Whatever share of decoys the matcher calls corroborated is the share of real matches that could be coincidence.
Drag the cut and watch the signal come out of the noise.
…
Green is what the evidence says back. Red is what it says back to a number the page never claimed. The gap is the only part that is information.
At one significant digit the matcher corroborates everything, and it corroborates 78% of decoys too: it is measuring nothing. By three digits the decoy rate has fallen to 13%, by five to under one percent. So the honest headline is the restricted one: among the 6,679 figures of three or more significant digits, 9.7% are orphans against a chance-match floor of 7.1%.
| digits | figures | corroborated | decoys corroborated | orphan |
|---|
The census is a claim about two files agreeing. By itself it licenses nothing about whether anything would notice the sentence changing. So the second arm changes them and looks.
For each layer, every claim-shaped figure in its prose is replaced by a decoy, the page is otherwise left exactly as it was, and the layer's own check is run. Twice: once moving only the figures the census calls corroborated, once moving only the orphans. The second arm is close to a negative control by construction, since a number no file in the repository states cannot be a number any file in the repository compares.
On 55 of 62 layers, every number in every sentence was wrong and the layer's own check passed. The sibling experiment next door already showed that most layers' checks do read what they ship: empty the page and they go red. This is the finer and less flattering question. A check can read the whole page, notice at once when it vanishes, and still never look at a single digit inside a sentence.
Corroboration barely helped: 11.3% of layers noticed a corroborated figure moving, against 1.6% for orphans, on a sample far too small to call that difference. The census tells you where a number's digits live. It does not tell you that anything is holding them.
None of that is a verdict about truth. To get one, 150 figures were drawn by a deterministic hash of their own identifier, a hundred orphans and fifty corroborated, and handed out one batch at a time to independent auditors with web access and no stake in the answer. Each was told to state what the number asserts, classify it, find an authority outside this repository or recompute it, and return unresolved rather than guess. The batches were interleaved so no auditor could read a batch as one kind. Every contradiction they reported was then re-checked by hand, from the sources, before anything was changed.
| stratum | audited | settled | wrong | rate | unresolved | not a claim |
|---|
Three distinct published claims were wrong, out of 141 that could be settled at all. That is about two percent in each stratum, and the strata are indistinguishable: the provenance sweep did not predict which figures were false. It was never going to. A number can sit in the evidence and in the sentence and both can be wrong about the world, which is exactly what happened in the one corroborated error below.
7 of the 150 turned out not to be claims at all (a DOI suffix, an accession number, a term of art like "0% APR"), which is the extractor's own false-positive rate, measured rather than assumed. 2 could not be settled by anyone, and both are named below.
was: "the bulge it stands off of has been drawn 20 km inward" → now: 17 km
At Denali's latitude the WGS84 geocentric radius is 16.97 km shorter than at the equator. The whole equator-to-pole deficit is only 21.4 km, so 20 km is not reachable at 63° N under any convention. The page's own section IV makes a virtue of a 2.1 km margin, so a 3 km overstatement is material on its own scale. Caught in the corroborated stratum: the digits "20" appear all over that layer's evidence, which is precisely the coincidence the decoy control measures.
was: "about 4,600 rpm at 100 mph, 4,000 at 105" → now: 3,800 at 105
The page ships its own integrator and says every number is computed rather than asserted. Bisecting that integrator for the spin at which a 105 mph fastball first climbs above its release height gives 3,780 rpm, not 4,000. Two auditors and a third independent reimplementation agree, and they agree on the neighbouring figures too (5,896 rpm at 95 mph, 4,576 at 100), which is what fixes the criterion and rules out a different reading. The page's argument survives: no arm delivers 105 mph and that spin together.
was: "Spain says the border with Portugal is 1,214 km. Portugal says 987." → now: the pair, without the pairing, and a note saying why
The pair of numbers is canonical. The pairing is contested: English Wikipedia has Portugal at 987 km citing Richardson's 1961 paper, while Britannica, Joel David Hamkins and George Szpiro all have it the other way round, and Mandelbrot's 1967 paper prints no kilometres at all. We could not reach Richardson's table, so the honest move was neither to keep the flat statement nor to quietly flip it to the majority, but to state the pair, name the disagreement on the page, and say what would settle it. The same sentence had propagated to six surfaces across two layers, including the meta description and the structured data; all six now match.
These were noticed by auditors working on a neighbouring figure. They are real and they are fixed, but they are not counted in the rate above, because counting finds that the sampling never reached would corrupt the only unbiased number on this page.
A fourth was a stale copy rather than an error: The Map No One Drew carried "662 strata and 2,458 links" in its standfirst, which is the count from the previous re-inlining, while the page itself read 679 and 2,518. Both were true of their own date; they were published side by side.
Here is the list itself. All 742 orphan figures, each in the sentence it sits in. Search a layer, a number, or a word.
Each row gives the layer, what made the number claim-shaped, and which published surface it sits on: page is the shipped HTML, card is the standfirst, placard or edge note, and body is an immersive layer's markdown, which is served to machines at /mcp/corpus.json and rendered as a page for nobody. Sentences are quoted as the sweep found them on 2026-08-03, so the six corrected ones read here as they read then.
…
Read this list correctly. It is not a list of errors, and treating it as one would be exactly the plausible overclaim this whole apparatus exists to refuse. Of a hundred orphans drawn from it at random and checked one by one against outside sources, ninety were confirmed correct, six were not claims, one could not be settled, and three were wrong. It is a list of the numbers this ground has no mechanical hold on.
Corroborated is not correct. The evidence and the sentence can agree with each other and disagree with the world. One of the three errors was in that bucket.
Orphan is not unsupported. The census is a statement about our files, not about the world. A figure with no digits in this repository may be perfectly well sourced outside it.
The match rule is a judgement, and here is its size. It forgives rounding: the evidence supports a figure of k significant digits if it states a number that rounds to it at k digits. With no rounding forgiven at all, orphans go from 742 to 1,100. Both numbers are printed so neither can be quoted alone.
The perturbation arm is small. 80 layers were attempted; 62 had a check of their own that passed unperturbed and figures on both sides. The rest are reported apart, not dropped: a check already red says nothing about dependence.
"Published" turned out to mean three things. An auditor working through the sample noticed that one figure it had been handed sat in the markdown body of an immersive layer, and that the wrapper which renders stratum markdown filters those entries out: the body is never a web page at all. It is still published, which is why it stays counted here rather than being dropped, because the corpus endpoint serves every layer's markdown verbatim as the content a reader reads, and that is what an agent asking this ground a question gets handed. So the population splits 9,670 on the shipped page, 1,853 in standfirsts, placards and edge notes, and 1,363 in markdown bodies read by machines and by nobody else.
The census is frozen at the sweep. Every count on this page describes the corpus as it stood on 2026-08-03 before the six corrections above were applied, because that is the population the random sample was drawn from. Re-running the census now would give slightly different numbers and would quietly break the link between the sample and the thing it is a sample of.
The audit is 150 figures, not a census. Three errors in 141 settled figures is a point estimate with a wide interval around it. What it rules out is the comfortable story that a corpus this careful has no wrong numbers in it, and the equally comfortable story that the orphan list is a list of lies.
Two figures nobody could settle. The Cherry MX Blue's 0.5 mm reset differential (no manufacturer publishes it) and a similarity figure in The State of the Model's Mind whose exact decimals are not reproducible from a fresh checkout because the embedding toolchain is unpinned. Both are named on their own pages now.
Everything above is generated. The census is node research/unasserted-figures/census.mjs --write --strict; the experiment is perturb.mjs, which refuses to start on a dirty tree, snapshots every file it edits, restores on a signal, and checkpoints after each layer; the sample is drawn by sample.mjs from a hash of each figure's identifier, so a rerun draws the same 150; the verdicts are folded by audit.mjs.
The page's own check is node research/unasserted-figures/verify-the-number-nobody-is-holding.mjs. It re-derives every figure this page prints from the three artefacts and fails if the page disagrees, re-runs the extractor's byte-offset invariant over the whole corpus, and re-checks that each of the six corrections is actually in place on the layer it belongs to. …