At 12:02 UTC today, two minutes after the Bank of England published its November decision, the Bank’s own page for that decision said the decision had not happened. The title read “Monetary Policy Summary and minutes of the Monetary Policy Committee meeting - November 2026 (to be published at 12pm)” and the body read, in full, that the summary and minutes “will be published on 5 November 2026.” The Monetary Policy Report page said the same. Appending a query string to the identical path returned the whole thing: Bank Rate 4 per cent, an 8–1 vote, twenty-four paragraphs of minutes. Same path, same minute, two different answers — and the one your pipeline would have polled is the one that says nothing happened.
One cache-buster is not a method, because the first one failed too
The sequence matters more than the finding. The canonical URL returned the placeholder. A first query-string variant — the same path with ?v=2 — also returned the placeholder. A second variant, the same path with a different parameter, returned the decision. So the fix is not “append something to defeat the cache”; two different appendages against the same path in the same minute gave two different documents, and only trying both told us which world we were in.
That is a nastier property than a page simply being slow. A retry against a stale edge gives you a well-formed, issuer-signed, correctly dated page asserting the negative. There is no error code to branch on and no malformed field to catch. If your handling rule is “open the window when the issuer confirms the release,” the issuer confirmed at 12:02 that there was nothing to confirm, and a conservative filter stays shut through the move it was built for.
The practical reading: an issuer URL is not a source, it is a cache with a source behind it. Two independent fetch paths against the same issuer are worth more than one fetch against the issuer plus one against a wire, because the failure you are guarding against here is not editorial. It is infrastructural, and it is invisible.
The same event, wrong a second way, in our own notes
Earlier this morning an automated read of a Japanese publisher’s rolling 24-hour calendar block returned the Bank of England policy rate at 22:30 Japan time and UK construction PMI at 21:00. A second, slower read of the same block on the same page — with the row-to-time pairing asked as the explicit question rather than left as a by-product — returned the Bank of England at 21:00 and construction PMI at 18:30. Twenty-one hundred Japan time is 12:00 UTC, which is what the issuer says and what three other readers we have say. The first reading was ninety minutes late.
Nothing was published from it, and the reason is worth stating plainly, because it is not vigilance. It is that the reading was checked against a figure we already held. Had the pairing error landed on an event with no independent timestamp in our file, it would have gone through.
What makes it dangerous is that it did not look broken. The misread put the Bank of England in 22:30 — which is exactly where that same block puts US initial jobless claims and the Q3 advance productivity figure. So the artefact did not produce an implausible orphan; it produced a plausible cluster, a central bank decision and two US prints landing together, which is a shape this desk writes about constantly and would have been predisposed to believe. A sanity check that asks “is this plausible” passes it. Only a check that asks “is this row attached to this time” fails it.
To its credit, the second read volunteered that three of its own pairings remained uncertain. An extractor reporting its uncertainty per row is worth more than one reporting a clean table, and almost nothing publishes the first kind.
Two extraction faults, zero data faults
Neither of today’s errors was a source being wrong. The Bank of England published the correct thing at the correct minute. The Japanese publisher printed the correct time next to the correct event. Both errors were introduced between the page and us — one by an edge cache, one by a table parse — and both arrived as confident, well-formed answers.
That is a different risk class from the one this desk usually writes about, and it is worse. A vendor carrying the wrong consensus is detectable by reading a second vendor. An extraction fault survives a second vendor, because the second vendor goes through the same extractor. The corroboration you buy by adding sources does not cover it; the only thing that covers it is reading the same source twice by different means, which is the expensive check nobody schedules.
If you run a news filter off scraped calendars, the audit worth running this week is not “are my sources good”. It is: take three events you already know the time of from the issuer, and check what your own pipeline thinks they are. The failure rate on that test is the number that actually governs your exposure, and it is not the failure rate of your sources.
What the stack actually is, and where the exposure sits
Today, in UTC: euro-area producer prices at 10:00, against a prior of 8.2 per cent. Norges Bank at 09:30, which held. The Bank of England summary, minutes and Monetary Policy Report together at 12:00 — one minute, three documents. A Bailey appearance at 12:30 that rests on a single Japanese calendar and that we have not been able to corroborate for two slots. US initial jobless claims and the advance third-quarter productivity figure, both at 13:30, in the same minute as the Czech central bank. Fed speakers at 13:50 and 14:10. The New York option cut at 15:00. Then Lane at 15:40, Lagarde at 16:00 and the Bundesbank president at 17:45. Tomorrow, the US employment report at 13:30, issuer-confirmed at 8:30 a.m. Eastern.
The consensus figures are thinner than the schedule. Initial claims: roughly 197,000 against a 194,000 prior, and the single page carrying that consensus states in another section of itself that no consensus has yet been published. One page, two positions, and we are quoting the figure as what that page says rather than as a consensus. The prior is the firmer number — two independent readers give 194,000. Advance productivity: 3.0 per cent against 1.4 per cent prior, on one Tokyo desk note and nothing else.
And the sizing fact, which is the only number on this page that should change what you do: one-week dollar–yen implied volatility fell to 6.23 per cent this morning, the lowest since 22 October, on one publisher’s reading. A one-week tenor priced today covers tomorrow’s payrolls. So the option market is charging a two-week-low price for the week that contains the month’s largest scheduled event, on the day a central bank hikes. That is either the cleanest carry in a fortnight or the market telling you the election and the 8 per cent collapse in crude already did the repricing. We are not going to tell you which, and we would note that a desk that cannot reliably establish what time an event is has not earned a view on whether its volatility is cheap.
What this does not tell you
It does not tell you the Bank’s page is broken, only that it served a stale copy to us. We observed two fetch paths against one path in one minute. We do not know whether the stale copy came from the Bank, an edge cache in front of it, or something between us and either. A reader fetching the same URL a minute later may well have got the decision. The useful claim is narrow and survives all of that: at 12:02 UTC a reader following the canonical URL could have been told the release had not happened, and no error state would have signalled it.
It does not quantify the pairing fault. Two reads of one calendar block is one disagreement, not a measured error rate. We do not know how often an automated read of that block mis-pairs rows, whether it is worse on dense rows, or whether the publisher’s markup changed this morning. We are reporting a near miss and the check that caught it, not a property of the publisher — and the publisher, read carefully, was right.
It does not give you a post-decision level in anything. We are publishing within half an hour of 12:00 and we have no sterling, gilt or dollar–yen print from after the announcement. Every FX level above is from the London morning before it: dollar–yen 157.79 low in the Tokyo afternoon and back to 158.10–20 by 08:32 UTC with a 158.14 ticker, around 158.30 on a 10:02 headline; euro–dollar 1.1268 high and 1.1230s low; cable 1.3501 high and 1.3465 low. The US ten-year rose from 4.84 to 4.87 per cent through that morning. None of those is a reaction to anything the Committee said.
It does not corroborate the volatility print. The 6.23 per cent is one publisher’s headline figure read at 08:27 UTC. We did not open the article behind it, we have no second reading, and we have no level for any other tenor or pair, so we cannot tell you whether one-week vol is low relative to one-month or low outright.
And it does not tell you the 12:30 window exists. That is the third slot in a row we have carried a Bailey appearance on one calendar’s authority with no issuer confirmation. If your schedule has a second window thirty minutes after the decision, it is still resting on one page.
Related
- FX event calendar — today’s stack in UTC, the 13:30 double print, the 15:00 cut and tomorrow’s payrolls.
- Signals — why an extraction fault survives adding a second source, and a data fault does not.
- EA presets — news windows keyed to a scraped calendar, and the audit that tells you what yours actually believes.
