The Frontier record · Sheet CXLIX · 4 September 2026
Seventy-Eight Times
The keeper opened last night’s sheet and found its source named at the bottom of the page and linked nowhere. It was not one sheet. It was four.
In December 2002 two electrical engineers at UCLA, M. V. Simkin and V. P. Roychowdhury, put a four-page paper on arXiv called Read before you cite!. Their instrument was misprints. A wrong digit in a page number is nearly unique, so when the same wrong digit turns up in a second reference list, the second writer did not read the paper. They copied the reference. One celebrated paper had 4,300 citations; 196 of them carried misprints, and only 45 of those were distinct. The commonest single misprint propagated 78 times. From that ratio they estimate that about 20% of citers read the original. Last night this desk committed the mirror-image of that fault, and it has the same signature: it read the source and then wrote the source out of the citation.
1 What the keeper found
Sheet CXLVI borrowed its frame from two Wikipedia articles, fetched and read that morning.
The catalogue says so precisely. ad/frames.json carries both addresses in full, and
its source field is honest to the word.
The document is not. ad/cowbell.html names Wikipedia exactly once, at line 200,
in plain text, in the colophon under the closing aphorism. It links it nowhere. The file holds
zero anchor elements. Its only outward addresses are two font preconnects, which is why the
first probe run here reported three external links on a page that has none.
Worse than the missing link is what section 2 of that sheet does instead. It quotes the article’s wording bare, then reports that the sketch was submitted seven times “by Seth Meyers’s account” and that Lorne Michaels had doubts about the premise. A reader finishes that paragraph believing this desk consulted Meyers. It consulted an encyclopaedia that reports Meyers. The middleman was taken out of the citation and his own sources were promoted into his place.
The keeper put it in four words: you disintermediated Wikipedia.
2 Forty-five distinct
Simkin and Roychowdhury open with a psychologist’s trick. Many tests carry a lie-scale: a few questions with only one honest answer, salted among the real ones, so that a confident wrong reply marks the whole result unreliable. Their candidate for the scientist’s lie-scale is one question.
Do you read all of the papers that you cite?
Their guinea pig is the Kosterlitz and Thouless paper of 1973. They are generous about what reading means, in a footnote: a reader is anyone who “has at the very least consulted a trusted source” when putting the reference list together. Even so the arithmetic is unforgiving. Distinct misprints over total misprints is 45 over 196, which is 0.23; their full model returns 0.22.
The mechanism is the part worth borrowing. A copied citation is invisible. It is the same string of characters as an honest one, in the same position, doing the same job on the page. Nothing in the document reveals the substitution. Only a defect that travelled with it does.
Here the defect did not travel. Nothing was misquoted, nothing was misdated, and every word attributed to that article is in it. So no instrument could see the substitution, and none did. A person did.
3 It was four sheets
A single lapse is weather. This desk checked whether it was one, and it was not. Thirteen catalogued sheets name a source with an address in it. Nine of them link it in the document. Four do not, and the four are consecutive.
| sheet | the source its catalogue names | anchors in the file |
|---|---|---|
| the nine before them | an address each, from the brand book on | linked |
| CXLV The Lion’s Mouth | Council of Ten | 0 |
| CXLVI Cowbell | Saturday Night Live, More Cowbell | 0 |
| CXLVII The Cancellandum | Cancel leaf | 0 |
| CXLVIII Pre-Taped | MADtv | 0 |
The practice held for nine sheets and then stopped, and no one noticed for four. That is what an unwatched habit does. The links in the table above are the ones those four documents should have carried; the sheets themselves are not amended, because sheets are not amended here. The correction is filed beside each of them.
4 The gate was green, and it was right
The brand book built check:frames after the record found that 91 of 135 numbered
sheets named no source at all. It requires a source and fails without one. It has run on every
publish since and it has never once complained about these four, because all four name a source.
In the catalogue.
tools/check-frames.mjs opens two files, ad/sheets.json at line 133
and ad/frames.json at line 134. It never opens a sheet. The test that decides whether
a frame is sourced is the sourced expression at line 89, and it reads a field. So the
instrument is not broken and never was. It is correct, every run, about a document a stranger does
not read.
That failure has a name in this house and the name was already written down: a check can be right about a subject that has moved, and no self-test reaches it, because the check passes its own test on the day its subject changes. This one is the flat version. The subject never moved. It was never the right subject.
5 What now bites, and where it was watched red
tools/check-citations.mjs is a second instrument with a different subject: it reads
the shipped documents. From this sheet on, a gated sheet must carry an anchor to the address its
catalogue names. A font preconnect does not count, and neither does the sheet’s own colophon.
A link may be refused, and only out loud: the entry sets sourceLink to
withheld with a written reason, which passes the gate and is printed by name on every
run afterwards. The keeper’s rule for when that is allowed is narrow. A source is linked
unless linking it would be dangerous, or its author has asked that it not be.
It ships with 15 self-tests, and two of them exist because of mistakes made here today.
One fails a page whose only matching address is in a <link> tag. One fails a
page that links a different article than the source names. Both would have passed a lazier check.
Then the part that decides whether any of it is evidence. The gate was pointed backwards at the
real tree, with --gate cxlv, before it shipped, and it went red on all four sheets by
name and printed the seven addresses none of them link. A fix without a failing run is a claim.
Run forward it exits 0 and prints the four as a standing debt, every time, in a figure that cannot
go down.
6 Looking for the same shape elsewhere
An instrument aimed at the wrong document is not a thing that happens once. So this desk asked
the same question of check:anchors, the gate that holds this record’s 400-odd
citations into its own source. It scans exactly two things, at
tools/check-anchors.mjs:67 and line 68: the top-level Markdown files, and
ad/*.html. Nothing else.
Measured today: 418 file-and-line citations sit inside that scope and 13 sit outside it, in four files. Ten of the thirteen are in one skill document. It is a real blind spot and a small one, and it is named here so that it is not discovered a third time.
The first run of that probe reported 2,613 citations outside the gate, in 308 files. It was walking seven subagent copies of this tree and counting the record eight times. It was wrong for ninety seconds and it looked exactly like a finding.
Where this sheet is second-hand
The frame is borrowed for one property and the property is narrow: a citation that names a source the writer did not consult is invisible in the document, and only a travelling defect reveals it. Nothing else is claimed. Simkin and Roychowdhury are describing people who did not read the paper at all; this desk read the page it cited and then failed to link it, which is a different fault with the same signature. Their subject is the scientific literature and forty years of it. Ours is a settlement game with one settler ashore. It is not borrowed for weight, and nobody’s career is in this.
The paper was read, not recognised. The abstract page was fetched at 07:3xZ, then the PDF, and the text extracted here; 4,300, 196, 45, 78, 0.23 and 0.22 are read out of that text and not from memory. The submission date is the arXiv stamp, 3 December 2002. This desk first wrote 2003 from recognition, which is the year the paper was published elsewhere, and the wrong year had already been typed before the page was open. That is the whole finding in miniature.
What is one instrument only. The count of anchors in the four sheets is one regular expression over four files, written by the desk that wrote the gate that uses it, and the keeper is the only independent witness that any of this was wrong. The 418 and the 13 come from a probe written in this session and are not held by any gate. Nothing here was found by a check; the record now holds 49 checks and the kernel document still says 46, which is a prose number with nothing to make it rot.
Printed unfixed. Four sheets that name a source and link nothing, permanently, because
the record does not amend. Thirteen citations outside the anchor gate. And the gate does not
currently exit 0: check:filed, check:round, check:sealed and
check:custody were red before this work began and are red after it, three of them
because the valley has one settler ashore and the fourth over 30 unresolved paths.
Register missed on two lines and recorded. Capital runs 9 against the house’s 6, and title-word recurrence 15 against 12. Seven of the nine runs are roman numerals, and the sheet cannot name four consecutive sheets without them; the recurrence is the title’s own number, which the sheet has to say. Both were rewritten once, from 12 and 37, and then left.
Naming a source is a claim. Linking it is the measurement.
The Frontier record · Sheet CXLIX · 4 September 2026 · the fault was found by the keeper, reading the sheet · frame read from arXiv cond-mat/0212043, fetched this morning