A checker that reports a problem which is not there is worse than no checker at all, because a false alarm costs the one thing the check was supposed to protect. This week ours raised two, on the same page, within a minute of each other—and if we had believed it, we would have deleted two correctly‑sourced sentences from a Beatles page in the name of accuracy.
The page was The Word, the John song at the end of side one of Rubber Soul. It was being rebuilt because it was too thin: about half the length of the page we use as the benchmark, with a Studio Notes section barely half of what the standard asks for. Rebuilding a page means re‑checking every claim already on it against the books, and two of the claims came back as not found.
What the verifier actually checks
Every factual sentence on a song page has to be traceable to a page of a book we hold, and the check is mechanical: a script extracts the text of the cited page from the PDF and looks for the exact phrase. It returns one of four verdicts. Found means the words sit on that page in that order. Partial means all the words are on the page but not consecutively. Miss means they are not on the cited page, and a second mode widens the search to the whole book.
The two suspect sentences both came back Miss against the whole book. Not the wrong page—absent. A quotation in quotation marks, attributed to a named source, that does not exist anywhere in that source is the single worst defect this site can carry, and we have found six of them since we started looking. The reflex is to cut the sentence and log it. That reflex would have been wrong.
The column break
The book in question sets its text in two columns. The extraction tool has two modes, and neither of them reads a two‑column page the way a person does. One mode preserves the visual geometry, which means it reads across the page, taking a line from the left column and then a line from the right, interleaving two unrelated sentences. The other reads the columns in the correct order but can silently drop boxed material and captions. We run both and count a phrase as present if either mode finds it, which catches most of the damage.
It does not catch a word broken across a column. One of the two sentences contains the word stumbled, and on the printed page that word is hyphenated at the end of a line: stum‑ at the foot of one column, bled at the head of the next. Both extraction modes produce a string in which the word does not exist. The phrase is on the page, in print, exactly as quoted; the text the checker searched had a hyphen and a column boundary sitting in the middle of it.
The second sentence failed for the plainer version of the same reason: it starts in one column and finishes in the other, so no contiguous run of text ever contains it.
Reading the page
What settled it was opening the page in the mode that preserves the layout and reading the paragraph as printed—which is what a person would have done first, and what an automated check is meant to save us from doing every time. Both sentences are there. Both are quoted correctly. The count of fabricated quotations on this site stays at six.
The rule we have written down from this is narrow on purpose: a miss on a page whose subject‑matter profile says the material is there is a reason to read the page, not a reason to cut the sentence. We have a separate, cheap check that reports which technical topics a given page discusses. When that check says a page is about overdubbing the bass and the phrase search says a sentence about overdubbing the bass is not on it, the two results disagree, and the disagreement is information. A phrase search that misses on a page which is demonstrably about the subject is far more likely to be an extraction artefact than a forgery.
Why we are not softening the check
The obvious response is to make the matcher more forgiving—strip hyphens, ignore line breaks, allow words to be separated by arbitrary junk. We are not doing that, and the reason is the failure it would let through.
The purpose of the phrase check is not to confirm that a set of words appears somewhere near each other in a book. It is to confirm that a specific sentence was written by a specific author on a specific page. A matcher loose enough to reassemble stum‑bled across a column break is also loose enough to assemble a sentence nobody ever wrote out of words scattered down a page, and to report it as verified. We would have traded a false alarm for a false clearance, and only one of those two errors puts something untrue in front of a reader.
So the check stays strict and the human step stays in. The cost is a handful of misses per rebuilt page that have to be looked at by eye. The benefit is that when the checker says a quotation is real, it means the words were found consecutively, in that order, on that page.
The rest of the week
Two pages were rebuilt to standard, both from the November 1965 Rubber Soul sessions: The Word and I’m Looking Through You, which were recorded in the same all‑night Studio Two booking on the 10th. Between them they carry 139 individually verified citations, and both cleared the depth standard they had been failing.
The rebuild also turned up two things worth having. The American stereo pressing of Rubber Soul carries an earlier stereo mix of The Word than the British one—not a remaster, a different mix, made four days before the one Britain got, with the double‑tracked vocal at full strength where the UK mix has it dropping in and out. Our page had said the earlier mix was superseded and never released. It was released; it was just released somewhere else. And on I’m Looking Through You, the two false guitar starts that open the American album are absent from the British one because Capitol was sent the mix untrimmed. Neither difference had a line on either page before this week.
Both corrections came from the same discipline as the fabrication check: bind the catalogue number to the mix by its position on the page, then look at what is actually printed. The tools do the reading. The judgement is still ours.