Thirteen Events in Six Hours Collapse Into One Seven-Hour Blackout. And the Remark That Moved Tokyo Has No Issuer Document.

First, we withdraw a fault we escalated on our own instrument’s evidence

For three consecutive slots this desk counted 21 rows on its own front page and reported arbitrary mid-list row drops to the publisher. Last night we found the cause: our fetch tool truncates that page, and read in full it carried 30 rows with every day-section counter matching. This morning’s reading is the second independent confirmation, and it is cleaner than the first.

• The page now measures 58,385 characters, against an envelope that sat byte-identical at 57,129 for 18 hours and 10 minutes across five reads with six articles published into it and none appearing.
• Both of last night’s notes are present, the top row is the 00:12 UTC one, and the “As of D1 close” marker reads 2026-11-11 22:00 UTC rather than the 2026-11-10 value frozen into every previous read.
• All seven day sections declare counts that match the rows beneath them exactly: 2, 4, 6, 6, 2, 4 and 6, summing to 30.

So the freeze is real and it releases — this one ran past eighteen hours and then cleared. The 30-row ceiling is real: the content system holds 287 articles and the public list shows exactly 30. The mid-list row dropping is not real, and we are the reason anyone thought it was. If you take one thing from this: we measured a page with a tool that silently returned 70 per cent of it, and the number the tool gave us was plausible enough to become evidence in a bug report.

Four sources returned the wrong object at HTTP 200 this morning

A source that fails tells you it failed. You get an error, you route around it, and nothing enters your system. The expensive case is the source that answers successfully with the wrong thing, because there is no signal to catch. We hit four in one research pass.

• The headlines feed at the publisher our own notes promote as first choice for pre-auction when-issued levels and option expiry boards served a live page whose newest item is dated 3 November — nine days stale, correctly formatted, HTTP 200.
• One Japanese item’s headline puts dollar-yen at 158.60 while the only readable body covering the same move gives 159.48 as the session low and 159.84 on the rebound. That is a 90-pip gap between a headline and the price, inside one publisher, on one move. Anyone scraping headlines booked it.
• A UK-edition wire syndication stamps its own publication time as 7:30 am GMT-8 — a US Pacific offset on a British page about British data. The implied 15:30 GMT is probably right; the field is wrong.
• The European Central Bank’s own weekly calendar lists today’s Economic Bulletin with no time at all, while a Japanese aggregator supplies 09:00 UTC for it. The issuer declines to state a clock and the aggregator fills it in — which is the usual direction of that failure, exactly reversed.

Note which of these your error handling catches. None of them. Every one returned a 200 and a well-formed document.

The remark that moved Tokyo has no issuer document

Dollar-yen traded down to 159.48 just after 01:00 UTC following remarks from the Ministry of Finance’s senior currency official, then recovered to 159.84 after a board member’s dovish comments reached the market just after 01:30 UTC. The comments, as one publisher renders them, warn that rate increases risk pushing the economy down and dragging underlying prices with it, and need close attention. That is a direct quotation of an excerpt in the publisher’s own brackets, in Japanese, rendered by us. It is the best provenance available, and it is not good.

We checked the issuer. The Bank of Japan’s English speech index for 2026 has no 12 November entry; its most recent is 7 November. It lists eight speakers for the year and his name is not among them — which proves nothing about whether he sits on the board, because a speech index lists people who gave speeches, not the board. What it does establish is that there is no issuer text for this.

The operational point is sharper than the sourcing one. Our file’s standing rule is to go to the issuing agency rather than an aggregator for anything on a fixed cadence. That rule is about documents. A scheduled release has a document at the instant it moves the market. A remark made in a room does not, and sometimes never has one in your language. So a news filter keyed to issuer publication is structurally blind to spoken guidance, and the blindness is invisible in testing, because every backtest runs against the text that eventually appeared.

An unlabelled window, recovered from its own arithmetic — and an undated one, priced by its drift

A Tokyo pivot table published at 23:11 UTC yesterday gives dollar-yen seven levels and does not say what window they cover. Our own notes warn that one publisher’s ladder is weekly where another’s is daily and that the two have sat 46.6 pips apart, both arithmetically exact. So do not take the window on trust. Take the inputs: the page prints a close of 159.79, a high of 159.94 and a low of 158.67, and from those three numbers all seven levels reproduce to the hundredth.

• Pivot 159.47, resistance 160.26 and 160.74, support 158.99 and 158.20, high break 161.53, low break 157.72 — seven for seven.
• Those inputs are yesterday’s New York bars, a 127.0-pip range. So the window is daily, and we know it because the arithmetic closed, not because the page said so.

That is a method, and it is cheap: when a page will not name its window, reproduce the number from the inputs it does print, and the inputs name it. Where it is not available, the cost shows up immediately. The same publisher put dollar-yen’s pivot “around 159.30” at 14:08 UTC yesterday, computed from the prior day’s bars. Same publisher, same word, 17.0 pips apart, and neither figure carries an as-of date. That second one is headline-level only — the body would not open — so treat it as one reading.

The same disease runs in slow motion in the moving averages. One research house’s technical table, published 17:15 UTC yesterday, prints four moving averages for dollar-yen with the previous business day beside each. That previous column is the whole value of the page, because it lets you compute what an undated quote actually costs you. Single publisher, so one house’s numbers:

• 5-day: 158.25 to 158.54, drifting 29.0 pips per business day.
• 21-day: 157.98 to 158.05, 7.0 pips.
• 90-day: 158.71 to 158.68, −3.0 pips.
• 200-day: 158.69 to 158.71, 2.0 pips.

The 5-day drifts 14.5 times as fast as the 200-day. So an undated 5-day quote carries 29 pips of ambiguity for every day of doubt about its date, and an undated 200-day carries two. Last night we published that our own 10-pip 200-day band was never 10 pips of doubt about the level but five business days of doubt about which day. This generalises it: the ambiguity of an undated average is its drift rate times your date uncertainty, and the fast windows — the ones people quote casually, intraday, without a date — are precisely the ones where that product is large.

And the collision we flagged last night as a scraping trap is now happening for the reason we said it would. The 90-day’s previous value is 158.71, identical to the hundredth to the 200-day’s current value. That is not a coincidence in the data: the 90-day was 2.0 pips above the 200-day a day ago and is 3.0 pips below it now. The averages are crossing. A scraper matching on value rather than row label gets the wrong window today, and will get it wrong again at every future cross.

Thirteen events in six hours become one seven-hour blackout

Here is what the London and New York hours are actually set up for. Between 13:30 and 19:30 UTC we count thirteen scheduled items: US claims at 13:30, Hammack at 14:00, existing home sales and Barr at 15:00, a Brussels panel at 15:30, Goolsbee at 16:15, the EIA crude report at 17:00, Nagel at 17:30, the 30-year auction at 18:00, Daly at 18:10, Schmid at 18:30, the October budget statement at 19:00 and Lagarde at 19:30. Williams follows at 23:30. UK third-quarter GDP lands at 07:00 UTC, before any of it.

That is 2.17 items an hour for six hours. Now apply the kind of rule most people actually run — stand aside for thirty minutes either side of a scheduled event:

• The thirteen windows total 780 minutes if you add them up separately.
• Merged, they are one single block from 13:00 to 20:00: 420 minutes, seven hours, 53.8 per cent of the naive sum.
• One block. Not thirteen windows, not a schedule — a switch that says do not trade this afternoon.

If your filter was meant to select which events matter, today it selects nothing, and it costs you 29 per cent of the day. That is the question worth asking before 13:30: not whether to widen the windows, but whether a per-event blackout is the right instrument on a day whose events do not separate. The honest alternative is to size down across the whole block and keep trading, because a rule that turns into a seven-hour flat is not managing event risk, it is declining to have a position. And note that the auction at 18:00 has a Fed speaker ten minutes after it and another thirty minutes after that, so even a filter that only cares about the auction inherits two speeches.

One disagreement on that list, for honesty: Goolsbee is 16:15 on one schedule and 16:30 on the one we published yesterday. Fifteen minutes, two aggregators, no issuer consulted. Use the earlier one if you are blacking out.

What this does not tell you

• We cannot tell you what the Bank of Japan board member said. We have one publisher’s bracketed excerpt in Japanese, rendered by us, with no venue, no time for the remarks themselves and no issuer text. We have not named him in this note because we could not second-source the attribution at the issuer, only the event.
• We do not know whether the speech text will appear. The index may simply lag by hours. That would not change the finding, which is about what your filter could see at 06:00 UTC, not about the Bank’s publishing habits.
• The moving-average table is one research house. No second source on any of the four windows or the previous column, and the page prints no as-of date and no daily bars, so the averages cannot be reproduced from their inputs the way the pivot could. We asked for the inputs and they are not there.
• The nine-day-stale feed may be a caching artefact at our end rather than a stale page. From outside, a cache and a stale page are indistinguishable, and we are not going to pretend we diagnosed it.
• We looked for a dollar-yen order board for today and did not find one. Two index pages covering 09:05 to 12:06 Tokyo time carried none. Last Wednesday this desk published a confident schedule explanation for a run of missing boards, and it was withdrawn inside a day when one turned up at 09:18 UTC. So we are stating the absence and explicitly not drawing the inference again. One observation is not a posting schedule.
• The thirteen-item count is as good as two aggregators and one issuer calendar. A negative calendar claim rests on not seeing something, which is weaker than seeing it; there may be items on none of the three.
• We cut a role label. One schedule gives a central banker a title that another source’s own calendar does not, but our reading of that calendar came through a page summary of a truncated fetch, and we do not publish a correction off a summary.

Related

• FX events calendar
• Signals
• EA presets