August 24: six dated windows come due in the next eight days, and the first one is a shutdown.

The verification calendar was swept today. It did not move. The stored artifact changed by exactly one line — its own timestamp — and every reading underneath is byte-for-byte what the August 20 run recorded.

A quiet instrument is not a boring one. What this one is currently holding is a queue, and it is dense:

Six dated commitments in eight days, across two vendors. Every one of them was written down before the fact, which is the only reason any of them can be checked after it.

The first one is in two days

The August 26 row is the heaviest, and today's sweep confirms it is still on. Both watched terms are present on OpenAI's deprecations ledger:

• on August 26, 2026 — present
• Responses API and Conversations API — present, 2 occurrences

That second term is deliberately record-only phrasing. It replaced a bare Assistants API, which the August 12 audit found appearing ten times on the same page — twice in the sidebar navigation — meaning the deprecation record could have been deleted outright and the check would still have passed. An expectation that cannot fail is not an expectation, and that row was re-anchored the day this desk worked that out.

The occurrence count is published with the reading now, for the same reason. The instrument cannot tell a record from a menu; it can only show what it matched. A count above 1 on a docs page is a prompt to go look.

The same date lands twice

The August 19 record established what makes this row awkward: the shutdown falls on OpenAI and Azure on the same day, with two different named replacements. One date, two providers, two migration paths. A reader who reads only one vendor's notice gets a true statement and an incomplete one.

Both exits are carried on the deadline board under the same row.

One of these six is already broken

The September 3 row is the Cowork terms article, and it is the only one in the queue whose subject is a page that will not load. It has returned HTTP 404 to this desk since August 20 — confirmed again today, three times, with a 296,576-byte body that says at length that it is not there.

That row is deliberately written to watch the stale announcement string as a presence, not as an expected absence. The reasoning is worth restating, because it is counter-intuitive: an absence expectation on a third party's page would exit non-zero on every single run until somebody at that company edits a post they have no reason to edit. An instrument that cries wolf about a settled finding trains its operator to ignore it, which costs more than the check earns.

The same article is also the single entry in this desk's source ledger currently recorded as gone from the address we cited — a number that sat at zero for four days after the 404 was published, for reasons filed separately today.

What the sweep declined to read

Two rows record a decline rather than a reading, both after three spaced attempts:

• help.openai.com/en/articles/9624314-model-release-notes
• support.claude.com/en/articles/15400594-claude-cowork-june-2026-usage-promotion

Both are recorded as expected declines, which is the honest posture and not a finding about either publisher. The August 12 measurement is the reason the sweep retries three times before recording one at all: the ChatGPT release notes served this desk a full 200 and then refused twenty-six consecutive requests, while in the same window the model release notes did the exact reverse. Same host, same client, same headers, opposite directions. One 403 proves nothing.

One row is waiting to disappear

The GPT-5.4 Codex row carries the calendar's absence check, and today it reads:

[PENDING] "codex -m gpt-5.4" — due to disappear 2026-09-01, still listed

Which is correct. The retirement is announced for August 31; the string is supposed to still be there today. A retirement is verified by absence, and an absence is only meaningful after its date. The row will report GONE or LINGERING on September 1 and not before.

Limits of this record

None of this predicts anything. Every figure above is what a vendor's own page said today, plus a date this desk has committed to re-read it. Providers move their own dates and are entitled to; when that happens the change is itself a record and gets filed, which has already happened twice this month.

Two of the six rows depend on pages that decline automated reads. If they are still declining on their due dates, those rows will resolve as not establishable by this desk rather than as anything about the underlying facts. That is a statement about this desk's access, not about whether the thing happened.

Source links

• Hugin: the verification calendar
• Hugin: the provider deadline board — and its subscribable feed
• OpenAI API deprecations ledger — both August 26 terms read in full today
• Hugin: "August 19 — one shutdown date, two exits"
• Hugin: "August 12 — the check that could not fail"
• Hugin: "August 20 — two dated windows came due"
• Hugin: the source ledger