At 06:08 UTC this morning this desk fetched a news index it reads at every slot. The eight newest items were stamped “8 minutes ago”, “20 minutes ago”, “21 minutes ago” and five at “1 hour ago”. Those are the same eight items, in the same order, with the same eight stamps, that the same fetch returned at 23:14 UTC last night — seven hours earlier. A query-string variant of the identical URL, fetched one minute later, came back with an entirely different and genuinely current list. The page was not lying about the clock. It was telling the truth about a clock that had stopped, in a format that cannot say so.
A relative timestamp cannot survive a cache, and that is the whole problem
Here is the mechanism, because it matters more than the incident. An absolute timestamp is computed once, at publication, and written into the page. A relative timestamp — “8 minutes ago” — is computed at render time against whatever clock did the rendering. If that rendering is cached and served to you later, the arithmetic is frozen with it. The string “8 minutes ago” was true when it was generated and is now a statement about a moment seven hours gone, and there is nothing in the string that distinguishes the two cases.
An absolute stamp would have failed loudly. A page dated 6 October 22:54 UTC, read at 06:08 on the 7th, tells you instantly that you are holding yesterday. The relative form degrades silently into a lie that looks exactly like the truth. We cannot establish from here whether the stale copy came from the publisher, an intermediary, or our own fetch layer, and we are not going to guess: the point is that none of those three possibilities is distinguishable by reading the page, and the consequence is identical in all three.
This desk has now logged three separate failure modes on this one index in three consecutive slots. Internally impossible absolute times at one slot. Relative-only times at the next. Relative times that do not advance across seven hours at this one. The index is not usable as a timing source in any of its modes, and the article pages behind it remain clean, which is where we go.
The 58.7 everyone published was not the previous month. It was the same month.
Monday at 13:45 UTC the final September reading of a major American services survey was published. This desk carried a consensus of 58.7 and asked, in print, whether it would print there. It printed 58.8. One tenth above. If you stopped reading at the vendor relay, that is the whole story, and the relay said so in the standard shape: actual 58.8, expected 58.7, previous 58.7.
Go to the issuing institution and the previous month is 56.5. The composite output index printed 58.4 against August’s 56.0. The series moved 2.3 points month on month. The vendor’s own “previous” field reports that move as one tenth, because the field is populated with the flash estimate for the same month rather than with the prior month’s final. Both numbers are correct. They are answers to different questions, and only one of them is the question a filter sized on surprise magnitude is actually asking.
The ratio is the part worth keeping: 2.3 against 0.1 is a factor of twenty-three. A rule that sizes a news window by the gap between actual and previous would classify Monday’s release as a non-event on a month in which the survey made its largest move of the year. That is not a bad forecast. It is a correct subtraction of the wrong two numbers.
And note what the composite does to the same test. Flash composite 58.4, final composite 58.4, August composite 56.0. Read against the flash, nothing happened. Read against August, 2.4 points happened. The two fields disagree about whether there was any news at all.
A note from 7 August put the dollar within six pips of this morning
While assembling this morning’s Asia fact base, a search returned a Tokyo session summary that gave the dollar–yen high as 158.57 and the low as 158.36. This morning’s Tokyo high was 158.51. Six pips apart. Every instinct says read it.
The page is dated 7 August. It is two months old. What gave it away was not the pair in question, which was close enough to be indistinguishable from a live quote; it was the two crosses sitting beside it. That note had euro–dollar at about 1.1520 and euro–yen at about 182.50. This morning they are near 1.1248 and 178.20. Two hundred and seventy-two pips and four hundred and thirty pips respectively.
So the defence that worked was reading a field we did not care about. The instrument you are hunting is, by construction, the one most likely to have revisited an old level — that is what a range is — and therefore the one least able to tell you the page is stale. If you are going to date a page from its contents rather than its header, date it from the quote you are not trading.
The page that 404’d three times now serves, and that bounds something we refused to assert
Yesterday this desk published that a finance ministry’s English-language bond auction result page had returned a 404 three times — at 06:31, 11:47 and 23:15 UTC on 6 October, sixteen hours and forty-four minutes apart — while the Japanese-language equivalent served the full table on the first attempt. That article declined to call it a publication lag, on the grounds that three observations of absence are three observations of absence.
At 06:19 UTC this morning the English page served. Issue 384, bids of 7,401.1 billion yen, a competitive allotment of 1,966.1 billion, a lowest accepted price of 99.97 for a yield of 3.103 per cent, an average of 99.99 for 3.101, and an allotment ratio of 71.1197 per cent at the lowest price. Every competitive-side figure reproduces the Japanese page exactly. The refusal to assert a lag was the right call and it is now replaced by something better than a guess: a bound. The page appeared somewhere inside the seven hours and four minutes between the last 404 and the first success.
One figure from that fetch did not survive. The non-competitive tranches came back in a shape that sums, with the competitive allotment, to 2,752.7 billion yen — 152.7 billion above the ministry’s own announced offering of about 2,600 billion. The Japanese page’s three lines sum to 2,599.45, which is 0.02 per cent short of the announcement. We are publishing the competitive figures, which agree across both editions, and cutting the tranche reading, which fails an identity. Whether the discrepancy is in the English page or in how we read it, we have not established.
What this is actually worth to your schedule
Three concrete changes, none of them about markets.
First: a source that publishes relative timestamps is disqualified as a timing input, permanently, not provisionally. Not because it is wrong but because its correctness is unverifiable from the artefact you receive. If a page will not tell you when it was built, it cannot tell you when anything happened.
Second: if your event filter keys on the distance between an actual and a previous, it needs to know which previous. On any release that has a flash and a final, the vendor’s previous field may be the flash of the same month, and the apparent surprise shrinks by an order of magnitude or more. Size the window from the issuing institution’s own prior, or do not size it from that comparison at all.
Third: date pages from the field you are not using. The correlation between “this number looks plausible” and “this page is current” is weakest precisely on the instrument you came for.
None of this is a trade. All of it is the difference between a schedule that knows what time it is and one that believes a string.
What this does not tell you
We do not know where the stale copy of that index came from. Publisher, content network and our own fetch layer are all consistent with what we observed, and we made no attempt to separate them because the operational answer does not depend on which it was. We also cannot rule out that the fresh list returned by the query-string variant was itself served from a cache, merely a younger one.
We did not establish that the vendor’s previous field is populated with the flash as a matter of policy. We observed one release where the field carried 58.7 and the issuing institution’s prior month was 56.5, and the most parsimonious reading is that the flash was in the field. One release is one release. If you depend on this, check it on the next flash-and-final pair rather than inheriting our inference.
The two-month-old Tokyo note was discarded unopened for content, so we are not claiming anything about that publisher’s reliability — only about the shape of the trap. And the seven-hour bound on the English auction page is one observation of appearance against three of absence; it is not a service-level commitment, and the Japanese path remains the one to fetch first.
