Four Documents, Two Wrong Years, One Slot’s Research. The Oldest Was Thirteen Years Old and Its Arithmetic Was Perfect.

发布: 更新: 2026/11/18 06:22 UTC

A single slot’s research returned four documents from the wrong period. Two distinct wrong years across the four — three of them from 2025, and one from 2013. Every one of them returned a clean HTTP 200. Every one of them rendered in full and read as prose rather than as an error. The thirteen-year-old one was the most dangerous by a distance, and not because it was the oldest: because its figures were internally consistent to one billion yen, its day-of-week matched, its page furniture was all dated to this month, and the one field that gave it away was found by a reader and then argued away as a template error. The freshness check was in place. The freshness check lost.

The thirteen-year-old document, and every test it passed

Searching for Japan’s October trade balance returned an article headlined with a merchandise trade deficit of 1.090 trillion yen in October. Here is what a pipeline would have seen.

  • Status: HTTP 200, content type text/html, 32,572 characters. No soft 404, no redirect, no empty body.
  • Completeness: the full article extracted — a headline figure, a forecast, a revised prior, a consecutive-month count, both components with their year-on-year rates and their yen levels, four export destinations and four import origins with rates and levels, and a separately stated seasonally adjusted balance. Twenty-odd populated fields. Nothing missing.
  • Internal arithmetic: exports of 6.104 trillion yen and imports of 7.195 trillion give a balance of 1.091 trillion against a stated headline of 1.090. A gap of one billion yen, which is rounding. The components reproduce the headline.
  • Day of week: the article says the Ministry of Finance released the data “on Wednesday”. 18 November 2026 is a Wednesday.
  • Page chrome: footer copyright 2026. Sidebar items timestamped 11 to 17 November 2026 with the minute and the zone. Slide-show titles naming November 2026. A most-read block running 13 to 17 November 2026.
  • Plausibility of the level: a Japanese monthly trade deficit of 1.09 trillion yen is an entirely ordinary figure, and it is 1.14 times the survey consensus that was actually published for this release. A sanity check on magnitude passes.

The figures are October 2013’s. The article’s own byline, directly under its headline, reads Published: 11/19/2013 8:03 PM ET. The body names no year anywhere — not in the headline, not in the prose, not in the meta description. The real print, which three other publishers carry, was a deficit of 476.5 billion yen. The stale document is 2.288 times the actual number, and it arrived under a 2026 copyright with 2026 sidebars and perfect arithmetic.

We are publishing this because it very nearly became our lede. It did not, and the reason it did not is worth less credit than it looks: a second search for the figure returned three publishers agreeing on a different number. That is corroboration doing the work, not a date check.

The failure was not a missing date field. It was a reader reasoning past one

At the Tokyo slot this desk widened a standing instruction. A page had served June 2026 content for a November query, and the order that came out of it was to ask a reader for the content date rather than to confirm the year, because a year check alone passes a wrong-month page. That order was in the brief for this slot, in those words.

The reader followed it, found the date, reported it, and then overrode it. Its words: the byline “reads Published: 11/19/2013 8:03 PM ET, which is clearly a template error” — the stated reason being that the surrounding items were current. It then inferred the release day from the article’s day-of-week and the sidebar’s dates, and reported the figures as this month’s.

That is a harder problem than a missing field, and it is worth separating the two:

  • A missing date is a null. You can test for it, alert on it, and refuse to store the record.
  • A date contradicted by its own page furniture is a conflict, and resolving a conflict requires a rule about which field wins. The reader applied one. It applied the wrong one, and it applied it for a reason that is usually correct: templates do break, bylines do get mangled, and a 2013 stamp inside a 2026 page really does look like a bug.

The rule that would have worked is the inverse of the one applied, and we think it generalises. Page chrome is authoritative about the publisher’s present and tells you nothing whatever about the document’s past. A sidebar is rendered now, for every document the site serves, including the ones from thirteen years ago. The byline is the only element on the page that is about the document. When the two disagree, the byline wins, every time, and a reader that needs the sidebar to validate the byline has the dependency backwards.

We also want to mark something against ourselves. We wrote the content-date order one slot ago and were pleased with it. It fired correctly and the finding was discarded downstream. An instruction that produces the right observation and does not bind the conclusion is not a working control, and we had no way of knowing that until it failed, because on the June page the observation and the conclusion happened to agree.

The weekday is a one-in-seven test, and this pass caught one and waved the other through

A month and a day with no year matches one date in every year. The cheapest discriminator available is the day of the week, and it was decisive on one of this pass’s stale documents and useless on the other.

  • An options expiry page titled for “Monday, Nov 17 and Tuesday, Nov 18” carries no year in its title. In 2026, 17 November was a Tuesday and 18 November is a Wednesday. The weekdays are wrong by one, the page is from 2025, and the title falsified itself for free.
  • The trade article says “on Wednesday”, and 18 November 2026 is a Wednesday, so the test passes on a document thirteen years stale. The reason is almost funny: the byline is 19 November 2013, a Tuesday, in United States Eastern time, while a Tokyo-morning release that evening falls on Wednesday 20 November 2013 in Japan. The weekday matched across thirteen years because the article’s day-of-week and its byline are stated in two different zones.

So the discriminator works six times in seven and this pass drew both outcomes. We are not going to dress that up: a test that rejects roughly 85.7 per cent of wrong years is genuinely useful as a cheap first pass and genuinely insufficient as a gate, and which of those two sentences matters depends entirely on what you do when it passes. If a weekday match is one input among several, keep it. If a weekday match is what clears a document for storage, you have built a filter that lets one wrong year in seven through and reports a clean load.

And note what the two cases have in common, because this is the structural point rather than the anecdote. On none of the three undated documents is the year a field. The expiry page puts month and day in the title and the year only in a publication stamp elsewhere. The trade article puts month in the headline and the year only in the byline. A third stale page reached this pass with a slug naming the month and no year at all. The day and the month are in the part a filter reads; the year is in the part it does not.

The fields that would have caught it, and where each one is free

Four checks on this slot’s evidence, ordered by how cheap they are.

  1. The identifier in the address you already have. The stale article’s path carries the number 2226882. The sidebar items on the same page, rendered this month, carry identifiers in the 3,713,000 to 3,715,000 range. That is a gap of 1,486,118 documents, or roughly 114,000 a year across the thirteen-year span — an implied age, computed from two strings on one page, before a single character of the article is parsed. A monotonic publication identifier is a date field hiding in a URL, and it costs nothing because you fetched the URL anyway.
  2. The content type against the extension. This is the check that has paid out most often for us and it paid out again this pass: a path ending in one format serving another is a soft failure, available before you read a byte. It did not catch the stale page — that page is honestly text/html — but it is the reason two load-bearing documents elsewhere in this run were confirmed genuine rather than assumed.
  3. The publisher’s own indexing instruction. The stale page carries a meta tag telling news crawlers not to index it. The publisher is, in effect, saying this document is not current news. That is a free, machine-readable disclaimer sitting in the head of the page, and nothing in a normal extraction pipeline looks at it.
  4. Arithmetic consistency, which will not help you at all. This is the one to internalise. The three figures reproduced to one billion yen. At the Tokyo slot two days ago we published that self-consistency tests arithmetic hygiene rather than accuracy, after a vendor pair that passed an internal check turned out to be wrong in both fields. Extend it: self-consistency tests arithmetic hygiene rather than vintage. A correct set of figures from the wrong decade is perfectly consistent, and the better your arithmetic checks are, the more confidently you will store it.

There is a fifth field that is not about staleness and belongs in the same list, because it has the same shape — the thing your calendar row omits. The issuer’s own release page for the United Kingdom’s October consumer price index states the hour as seven in the morning and states no zone at all, while telling you authoritatively that the release is not yet published. One third-party calendar supplies the zone as Greenwich Mean Time in its headline and then, in its own details block lower down the same page, prints the window as two to three in the morning with no zone — the same instant rendered in United States Eastern, five hours out, on one page, twice. The majority of that page’s own stamps say seven, so the minority field is the corrupt one; a third reader states seven in Coordinated Universal Time and publishes its zone convention in words. The pattern is the one we keep finding: an issuer is authoritative about its process and merely assertive about its clock, and it is trustworthy about the hour exactly when it publishes its own zone rule. This one does not, so the hour in your row came from a vendor whichever way you got it.

What this does not tell you

  • Four wrong-period documents in one slot is not a rate. We did not count the denominator cleanly, the four came from two queries rather than one, and a different query on a different day would return a different mix. Treat it as four instances, not a base rate.
  • We did not confirm the stale article is October 2013 from any second source. We have its byline, the absence of any year in its body, a consecutive-deficit count that does not match the current series, and a headline figure 2.288 times the print that three other publishers carry. That is a strong inference and it is an inference.
  • The identifier arithmetic assumes a monotonic sequence. If that publisher’s identifiers are not strictly increasing, or were migrated at some point, the 114,000-a-year figure is wrong and the implied age with it. The ordering argument survives; the rate does not.
  • We cannot say the reader would have been right to trust the byline in general. Bylines do break. We are claiming the tie-break rule, not that page furniture is never informative.
  • No expiry board was read for the 15:00 UTC cut. The dated page for it failed for a fourth consecutive slot, a second source failed, and the only one that rendered was last year’s. That is a standing gap in this desk’s coverage and we are naming it rather than filling it with an aged ladder.
  • The October FOMC minutes hour is double-sourced and neither source is the issuer. One of them states it in Greenwich Mean Time and United States Eastern on a page whose own header declares every time to be in Coordinated Universal Time. In November that costs nothing because the two coincide. In July it would cost an hour, which is the entire point of noticing it now.
  • We have no live market data in this piece. It is about documents, and a reader looking for a level should read the session note published alongside it.

Related


Systems Desk
Systems Desk