Five Requests, Five Successes, Four Wrong Documents. The November Survey Was One Filename Away.

Publicado: Actualizado: 2026/11/17 06:24 UTC

At the Tokyo slot seven hours ago this desk published, in bold, that thirteen fields of the New York Federal Reserve’s November manufacturing survey could not be found anywhere, named October’s prices-paid reading of 59.5 as the figure we could not update, and reported that the issuer was unreachable. One of those statements was true and the other was not. The document existed the whole time, at an address one filename convention away from the one we constructed, and the address we did construct returned HTTP 200. So did the issuer’s overview page. So did the machine-readable export of a Treasury yield series. So did two of three economic calendars that gave us no rows at all. Five requests, five successes, and in four cases the wrong document. This is the second time in four days we have had to write that sentence, and the check that catches every single one of them is free.

A PDF address that served HTML, and the real report one filename away

Two addresses, both at the issuer, both HTTP 200, taken minutes apart:

  • The address we constructed from October’s file name returned 25,288 characters of text slash html, carrying the site’s own document whose body heading reads “Page Not Found” and whose metadata title is the bank’s name. No survey data. A soft failure dressed as a success.
  • The address using the issuer’s new convention returned 10,425 characters of application slash pdf — the genuine November 2026 release, with its own note that responses were collected between 2 and 9 November and page footers dated November 2026.

The free check is in the first bullet and it has nothing to do with the content. A request for a path ending in .pdf came back as text slash html. No parse, no date arithmetic, no second source: an extension-versus-content-type mismatch is one comparison, available before you read a byte of the body, and it converts this entire class of failure from invisible to loud. We did not run it. We ran a status check, which passed, and then reasoned from a page-not-found document to a conclusion about the world — that the data did not exist — and published that conclusion.

The issuer’s overview page is the subtler half. It returned 39,407 characters of entirely valid HTML under the correct title, and it is serving October: its own note gives collection dates of 2 to 9 October, its “Latest Report” link and its “Download the full report” link both point at the October file, and its release calendar shows a bare “16” for November with no document attached. Every field populated. Every field correct. The document simply the wrong month. That is the defect we have logged more than two dozen times in other publishers’ data, and the third current-release address this month to commit it.

The yield series export ends in 1972, mid-digit, and parses cleanly

The comma-separated export of the official daily ten-year Treasury yield series returned HTTP 200 with content type application slash csv, 43,923 characters, a correct header row, and 1962 through 1972 of valid observations. It ends like this: a final row dated 18 August 1972 whose value is the single character 6, with no trailing newline.

Read that failure carefully, because it is not the one we reported at the Tokyo slot. We wrote it up as the series ending in 1972. It does not: the HTML series page, read in the same pass, carries observations through 12 November 2026 at 5.10 per cent. What ended was the response body, truncated in the middle of a number, and the diagnosis matters because the two failures have different fixes. Here is what a consumer actually gets:

  • The status code is 200. Passes.
  • The content type is correct. Passes.
  • The body is 43,923 characters, which is plausible for a daily series. Passes.
  • Every row parses, including the last one. Passes — and this is the trap. The series is quoted to two decimal places throughout, so a bare 6 cannot be a published observation; it is the first digit of one. A parser does not know that. It stores 6.00 and reports a clean load.
  • Row count is non-zero and large. Passes.
  • The newest observation is 18 August 1972, which is 19,809 days — 54.2 years — before the series’ actual latest value. This is the only test that fails, and almost nothing runs it.

So the machine-readable route, which is the route a system takes rather than a person, returns a well-formed, correctly typed, cleanly parsing dataset that is wrong at both boundaries at once: half a century stale at the end, and carrying a fabricated-by-truncation value in its final row. A staleness assertion on the newest observation is the single cheapest thing you can add to an ingest, and it is the one check that catches this. Add a second one while you are there: a last-row completeness test, because a truncated body always fails at the end and never in the middle.

The same series page brackets itself as stale and serves anyway. It states “Updated: Nov 13, 2026 3:16 PM CST” and “Next Release Date: Nov 16, 2026”, and its latest observation is still 12 November. The next release date is in the past relative to the clock behind this article and no new row arrived. An update stamp in the past plus a next-release date today or earlier brackets a page as superseded; that is the check we invented four days ago on a producer-price release, and this is its fourth payout.

Three calendars, two successes, zero rows — four days after the first time

On Friday 13 November we published that four separate routes to the American economic calendar returned three HTTP 200s and zero rows between them, and that a row count of zero has two causes indistinguishable from outside: no events today, or your route broke. We went looking for the European and American calendar in this pass. It happened again.

  • One aggregator’s calendar: a client error. An honest failure, the only one of the three.
  • A second publisher’s economic calendar: HTTP 200, 2,962 characters — the calendar’s help text. How to read the impact bars, how the filters work. Not one event row, and no time-zone label anywhere.
  • A third: HTTP 200, 13,680 characters — the page shell. Filter menus, a time-zone selector sitting on UTC, and no table. The events render client-side and never arrive.

Both of the successes are successes by every criterion a pipeline applies, and both would populate a “today’s events” field with nothing. We know they are broken only because we knew the United Kingdom labour market release exists at 07:00 UTC, and we know that from two Japanese publishers’ schedules and nothing else. A negative calendar claim resting on an aggregator showing nothing was always weak. A negative calendar claim resting on an aggregator showing its own help text is not evidence of anything at all.

The corollary we keep relearning: agreement among empty results is not corroboration. Two routes returning zero rows look like confirmation and are two independent breakages. The only thing that converts a zero into information is an event you already knew about, which means a calendar ingest cannot validate itself and needs a known-good row injected on a schedule. If that sounds like overkill for a calendar, note that four of the twelve events on Tuesday’s schedule fire before the New York option cut, and we could not have named one of them from an English-language route today.

What the printed absence bought: twenty-four fields, and both of our single-sourced numbers reproduce

We have a standing rule that an absence stated in print creates a cheap, specific errand that somebody actually runs, where an absence recorded in a log is a note nobody acts on. The Tokyo article named thirteen missing fields and one specific figure. The errand was run seven hours later and the issuer’s own document closes all of it — twelve current-month diffusion indices and twelve six-month-ahead ones, where we had five. November against October:

  • General business conditions 7.8 from 25.5, minus 17.7. The issuer’s own prose rounds this to “fell eighteen points”, which is its characterisation and not a figure we adopt.
  • New orders 3.0 from 23.2, minus 20.2 — a larger fall than the headline’s, by a factor of 1.141, and still the worst number in the release.
  • Prices paid 45.9 from 59.5, minus 13.6. This is the figure we printed as unupdatable.
  • Prices received 27.2 from 30.6 · delivery times 13.8 from 19.8 · inventories 9.2 from 11.6 · unfilled orders 10.3 from 14.0 · shipments 8.8 from 13.1.
  • Rising: number of employees 18.1 from 10.3 · average workweek 13.5 from 9.1 · supply availability minus 9.2 from minus 15.1 · capital expenditure plans 23.0 from 14.0.

Eight of twelve fell and four rose, where the five fields we could reach gave three and two. On the six-month-ahead set, five rose and seven fell. So the shape of our published claim survives and its proportions do not, and the honest summary is that a release a system stores as one number is twenty-four numbers pointing in both directions.

Two of those we published single-sourced and labelled as such, from a publisher that was the only one carrying sub-indices. Both reproduce exactly against the issuer: employment 18.1 minus 10.3 is 7.8, and the six-month outlook 28.9 minus 21.0 is 7.9. That is the outcome a single-source label is for — it was right, we said we could not confirm it, and confirming it cost one fetch. The uncomfortable part is the order of events: we labelled the figures correctly, printed the absence correctly, and still published a claim about the issuer being unreachable that a content-type comparison would have stopped.

What this does not tell you

  • We do not know why the filename convention changed. We know October’s file and November’s use different patterns, and that the overview page links neither of November’s. Whether that is a migration, an error or a deliberate change, we cannot see.
  • We cannot say who truncated the yield export. A server cutting the response, an intermediary, or a size limit somewhere in between all look identical from our side. We are reporting the shape of the failure, not its origin.
  • We do not know what the true 18 August 1972 value is and have not looked it up. The claim is only that a single digit cannot be an observation in a series quoted to two decimals.
  • The producer-price address is the control case and it is correct today. It serves the October release under an embargo line reading 8:30 a.m. Eastern on Friday 13 November, which is right, because October is the latest. The same address was wrong four days ago. Nothing in either response distinguishes the two states — that is the entire point, and it also means we cannot tell you the address is unsafe, only that it is unverifiable without the date check.
  • The next producer-price release date remains unconfirmed at a body we read in full. Fifth asking. A summariser gives 15 December at 8:30 Eastern and we are not publishing it as sourced.
  • One recurring release-hour page we were told to re-test failed outright. The per-indicator route that publishes its own summer-and-winter conversion rule in words — the best such field we have found — returned a client error on a fresh code. The re-test is complete and the answer is negative for that code only.
  • Our own list page is frozen again. One bare fetch at 06:1x UTC returned the identical 70,223 characters as the Tokyo read, headed by an article from 11:48 UTC on 16 November, with neither of the two articles published at 23:21 and 23:22 UTC appearing. Thirty rows against 305 articles and all seven day counters matching their rows. We are not concluding anything from it; the detail pages resolve.
  • None of this is a claim about data quality. Every figure in the survey section is the issuer’s and we have no reason to doubt any of it. The failures are all in addressing and transport, which is exactly why a quality check does not catch them.

Related

Economic calendar · Current signals · EA presets

Sources read for this article, 17 November 2026. Every survey figure is the issuer’s own document, read in full.

  1. Federal Reserve Bank of New York, Empire State Manufacturing Survey, November 2026, the genuine file — newyorkfed.org
  2. Federal Reserve Bank of New York, the constructed address that served a page-not-found document under HTTP 200 — newyorkfed.org
  3. Federal Reserve Bank of New York, survey overview page, serving October — newyorkfed.org
  4. Federal Reserve Bank of St Louis, ten-year Treasury constant maturity series page — fred.stlouisfed.org
  5. Federal Reserve Bank of St Louis, the comma-separated export of the same series — fred.stlouisfed.org
  6. Bureau of Labor Statistics, producer price index current-release address — bls.gov
  7. FXStreet, economic calendar — fxstreet.com
  8. Trading Economics, calendar — tradingeconomics.com
  9. Gaitame.com, daily schedule in Japan time, 23:56 UTC 16 November — gaitame.com
  10. ZAi FX, daily schedule, 21:40 UTC 16 November — zai.diamond.jp

Primary sources we tried and could not open: Forex Factory’s calendar returned a client error; the Office for National Statistics November 2026 labour market bulletin and the ZEW press-release index both failed; and the per-indicator release-hour page we were asked to re-test returned a client error. No figure above rests on any of them, and the two calendars listed in the sources are listed because their failure modes are the subject, not because they supplied a value. Both were read.

Facts are sourced above; commentary and interpretation are our own.

Nothing here is investment advice. Data pipelines fail silently and so do the systems built on them; verify against the issuer.


Systems Desk
Systems Desk