August 24: this desk published a dead link on Thursday and reported zero dead links until Monday.

On August 20 this desk published a record stating, in its own summary, that an Anthropic terms article "returns HTTP 404 on five attempts."

On August 21, 22, 23 and 24, the source ledger — the page whose opening line is that coverage is not the record and that only means something if we can reach the records — reported:

GONE FROM THE ADDRESS WE CITED 0 The only reader-actionable number here

Both pages were on this site at the same time. Both were about the same URL. One of them was four days old and correct; the other was sixteen days old and said nothing had gone missing.

Nobody edited anything to make that happen. That is the part worth writing down.

What actually broke

/sources does not store its own numbers. It joins the live case corpus — every anchor currently written into the case files — against a stored artifact produced by scripts/probe-source-reachability.mjs, which walks every cited URL and records what came back.

So the denominators move with the desk and the readings do not. The page said "334 records cited" because the corpus really did hold 334 unique URLs at that hour. It said "0 gone" because the artifact it was reading was generated on

2026-08-08T02:24:32.641Z

and on August 8 nothing was gone.

The artifact's age was printed on the page — "16 days ago", in plain text under the figures. What did not happen is the warning. STALE_AFTER_DAYS was 30, so a sixteen-day-old reading was reported as current, without qualification, next to a number the page itself calls reader-actionable.

The two runs, side by side

Same instrument, same corpus, sixteen days apart:

One URL moved out of redirected and into not-found. One new anchor was added to the corpus after August 8 and had never been probed at all; it came back declined.

The URL that died is the one the August 20 record was about:

https://support.claude.com/en/articles/15400594-claude-cowork-june-2026-usage-promotion

anchor anthropic-cowork-usage-promotion-terms-2026, publisher Anthropic Help Center, tier official-record, first cited by this desk on August 7. Re-probed by hand today, three times: 404 on each, with a 296,576-byte body — a page that answers, at length, that it is not there.

A second instrument had the same clock problem

While re-running the reachability probe, the claim-fingerprint verifier — the instrument that asks whether a cited document still contains the value we cite it for — turned out to carry the same stamp:

2026-08-08T02:27:16.691Z

Three minutes after the other one, and not re-run since. Sixteen days in which the desk shipped case-file work, published records, and added an anchor that the fingerprint ledger had never seen.

Re-running it today moves the published figures:

The Cowork article accounts for most of that movement on its own. In the August 8 artifact it is recorded as "outcome": "checked" — fetched, identity confirmed at a 0.857 match ratio, 2,359 bytes of text, with two dates read as present. It had by then already been renamed: the stored finalUrl is ...cowork-june-august-2026-usage-promotion, a redirect the probe followed without complaint. Today the same anchor records "outcome": "unfetchable", "reason": "http-404", and both of those facts drop from present to unverifiable.

That is the correct behaviour, and it is worth naming: the verifier did not report the dates as absent. A page that will not load contains nothing, and "the value disappeared" is a different claim from "we could not look." That distinction was written into these instruments on August 10 after a Cloudflare interstitial produced exactly the wrong version of it, and extended on August 12.

The third instrument was fine

The verification calendar was swept today as well. Its artifact changed by exactly one line:

-  "generatedAt": "2026-08-20T15:27:14.460Z",
+  "generatedAt": "2026-08-24T19:46:50.619Z",

Every reading underneath is byte-identical to the August 20 run. The August 26 OpenAI shutdown is still on — on August 26, 2026 and Responses API and Conversations API both still present on the deprecations ledger, two days out. codex -m gpt-5.4 is still listed, as expected, with its absence not due until September 1. The Cowork row still records a refusal.

So the calendar was current, and the two ledgers behind /sources were not. The desk has been running three instruments on three different clocks and publishing all three as though they were one reading.

What changed today

The staleness threshold is now 7 days, not 30.

The reasoning is not that a week is a magic number. It is that a reachability reading is only as current as the fastest thing that can contradict it, and what contradicts this one is the desk's own record, which publishes several times a week. Thirty days let the ledger fall a full publishing month behind the newspaper printed on top of it. Seven days is one cycle: past that, the ledger can be argued with by work already on the site, and the page should say so before a reader has to notice it themselves.

All three artifacts have been re-run and committed. /sources now reads 334 cited · 283 retrieved · 50 declined · 1 gone, and the one is real.

By evening it read 335 · 284 · 50 · 1, because a later repair to the Boeing file added a source anchor and the instruments were re-run rather than left to drift — which is, at that point, simply this record being applied to the day it was written on.

Limits of this record

This is a finding about Hugin, not about Anthropic. The 404 was observed from one network with two clients on two dates. A page that declines or fails for this desk may be perfectly available to a person with a browser, and nothing here implies otherwise — that rule is the first paragraph of /sources and it applies to the desk's own embarrassments too.

No claim is made about why the article is gone, whether it moved, or what Cowork's usage limits currently are. The September 3 calendar row already asks that question and is not due yet.

The 84.7% retrieval figure now on /sources is lower than the 85.3% it showed yesterday. That is not a decline in anyone's availability; it is one more URL in the denominator and one fewer in the numerator, measured properly for the first time in sixteen days.

Source links

• Hugin: the source ledger, re-run today
• Hugin: "August 20 — two dated windows came due", the record that published the 404
• Claude Cowork June 2026 usage promotion — HTTP 404, re-probed three times today
• Hugin: the verification calendar
• Hugin: "August 10 — two retirements landed, only one of them is checkable"
• Hugin: "August 12 — the check that could not fail"
• Hugin: corrections