A song page on this site is checked by two programs before it is allowed to claim it is finished. One reads the structure—are the sections present, in the right order, deep enough, correctly footnoted, is the JSON‑LD parseable, is the badge the page wears justified. The other reads the citations—does every claim the page makes actually sit on the page of the book the page says it does. Both ran clean on two pages this week. Both pages were wrong.
Nine and twenty-six
The infobox on the Penny Lane page reported nine noted takes. Four lines below it, in the callout that summarises what is unusual about the recording, the same page reported twenty‑six. Nine is right. Twenty‑six is not a rounding error or an outdated draft value; it is a take number that exists, on the correct page of the correct book, belonging to a different song—Strawberry Fields Forever, whose session entry sits immediately alongside Penny Lane’s in the same chapter. The two records were made in the same weeks and are discussed in the same breath everywhere, including in the source.
The corrected value had been written into the infobox the previous week. The old value survived four lines away because it lived in a second copy of the same callout that a rewrite had left behind, and because nothing looks for two copies of a callout. Both pages worked on that week carried the duplicate; they were the only two of two hundred and fifteen that did.
Why neither checker saw it
The structural checker reads the page as a shape: sections, word counts, markers, badge state. Two callouts is a legal shape. The citation checker reads a list of claims that the session writing the page chose to write down—and nobody writes a claim to verify a number they believe they have already deleted. Between the two of them there is a gap exactly the size of a page contradicting itself, and it is not a narrow gap. The same week, a different page asserted a running time four separate times that appears in none of the mounted sources; it had been arrived at by subtracting one sourced number from another, it carried correct footnote markers, and it disagreed with its own infobox by two seconds.
The common shape is worth naming, because it is not carelessness. Every one of these numbers felt sourced. Each came from the right book. Each was attached to the wrong thing—the wrong song, the wrong arithmetic, the wrong copy of the sentence—and the attachment is precisely what a checker that reads claims in isolation cannot test.
Three cheap tests
The fix being proposed is deliberately unambitious: compare the page’s own fields against each other. Does the length in the infobox match any running time asserted in the prose? Does the stated take count match the highest take number in the session timeline? Do the recording dates in the infobox fall inside the range the timeline covers? And, added this week, does the page carry exactly one summary callout? Each is a few lines of code. Each would have blocked a real defect that shipped on a page publicly badged as reviewed.
One of them will produce a false positive immediately, which is worth stating before it happens rather than after. With a Little Help from My Friends has ten numbered takes and a timeline whose highest number is eleven, because take eleven is not a performance—it is the reduction the ten takes were bounced into. A check that compares two numbers without knowing that reduction takes are not takes will flag a page that is entirely correct. The test has to know the difference, or it will train whoever runs it to ignore it.
The adjacent column
The same failure has an older and more entertaining form, and it turned up again this week in a page being rebuilt from scratch. The entry for the Sgt. Pepper reprise had for some time reported that George asked for a few seconds of laughter to be edited onto the track, from a named tape in the studio’s in‑house sound‑effects library, and that its mono mix used heavy artificial double‑tracking. All three facts are on the right page of the right book. All three belong to Within You Without You, whose entry occupies the adjacent column.
The giveaway had been sitting in the source sentence the whole time: the laughter was edited onto the end of both mixes. The reprise was mixed to mono on the first of April and to stereo on the twentieth—three weeks apart, in different rooms. There were no two mixes to edit anything onto that night. The detail that disproves the claim is four words long and is inside the passage the claim was taken from.
Extracting text from a two‑column book interleaves the columns line by line, so a phrase search can return a true hit from a false neighbour. That hazard has been documented here for months and the verification tool prints a snippet around every match specifically so the snippet can be read. It only helps if the snippet is read to the end of the sentence.
What this costs
Two pages were re‑verified this week and eight corrections came out of them; two more were rebuilt from the sources and four long‑standing errors came out of those. That is a high yield for a routine pass, and the honest reading of it is not that the process is failing but that the process has only recently started looking in this direction. Every check on this site exists because something got through. The ones being added now exist because a page can satisfy every rule about how it is built and still tell a reader two different things about the same record.