Corrections
Every entry on this site whose factual claim turned out to be wrong, what it claimed, and the day it was corrected.
A corrected entry keeps its original text. The correction is added above it and the error is left readable underneath, because an archive you can audit is worth more than an archive that looks clean. This page exists so that record is one click from the front of the site rather than buried inside the post it amends.
Two kinds of correction are listed. The first is a written entry whose claim was wrong. The second is a case record — a source anchor or a dated row inside a case file. Those are published too, on the case pages and over the API, and until August 2026 a wrong one could be quietly edited with nothing to show it had ever been different. That is the exact failure this desk exists to point at in other people’s archives, so they are listed here with what the record used to say.
- Journal
Promtly is open source where it touches your machine: paste one prompt into Claude Code to install it, then tell Claude what you want it to become.
September 19: The public repository contains the bridge, not the whole hosted product. The browser sandbox does not remove the need to trust the page handling prompts. The original chairs section incorrectly said one briefing seats both Parley models; the chairs require different keys and MCP connected in each client. The text below now distinguishes bridge capability from the tested product flow.
- Journal
Two models will agree to anything that cannot be checked. So the styleguide was written as rules a machine can fail, and the second model's chair is waiting on a usage reset.
September 19: This initial account overstated completion and universal client support. It also proposed changing the version after signing, which changes the signed bytes. Fix the version before signing. A matching hash identifies text, not a verified model identity or a correct document. The original rehearsal remains archived; Parley links the subsequent Astra–Fable review.
- News
August 31: OpenAI's deprecations ledger carries four future dates. This desk was watching one.
Published saying the deprecations ledger “carries four dated retirements still ahead of today” and listing them. It carried at least seven, and one of the four was itself two dates. Missing from the count were September 24, 2026 (the Videos API and the sora-2 family, announced March 24, 2026), September 28, 2026 (four legacy completion models, announced September 26, 2025) and October 23, 2026 (seventeen rows including gpt-4-turbo, gpt-4, o1 and gpt-image-1, announced April 22, 2026) — every one announced before the August 31 read. This desk kept no copy of the page as it stood that day, so the three dates are established from the September 7 reading and from the announcement dates the page carries. What settles it is closer to home: this desk had already filed all three. Its public deadline board has carried them since August 10, and the card for the September 24 group says in those words that the replacement column is empty; the AI Release Receipts case file recorded the same three that day. So the error was in this record, not in what the desk had been told. The entry's central point stands and was understated — the verification calendar was watching one of seven, not one of four.
- News
August 10: two retirements landed. Only one of them is checkable.
Published saying the developer documentation is “wide open to an identified machine” while “every page a person would have been sent to” is not, and attributing the refusals to the desk identifying itself. Measured on August 12, that split does not hold: help.openai.com served this desk's identifying client a full 144KB document on ten consecutive spaced requests and refused the same client on the same page later the same hour, and a sibling article swapped states in the opposite direction inside twenty minutes. learn.chatgpt.com — also consumer-facing — serves the identifying client consistently. The readings in this entry are unchanged and the Atlas pages have refused every attempt across three days, so the “not establishable” resolution stands; the stated cause does not. Full measurement in the August 12 entry.
- News
August 4: the resets stopped, and August is filling up with end dates.
Published saying the 3.9-day gap was “1.5x July’s average interval”. The figure divided into it — 2.58 days — is July's density, twelve resets across a thirty-one-day month, and not the mean interval between consecutive resets. Those twelve resets occupied a span of 19.3 days rather than the whole month, so the true mean July interval is 1.75 days and the gap was 2.2x it, not 1.5x. The error understated this desk's own comparison, which is the direction that flatters the calmer reading the entry was arguing for. The same paragraph said “all forty intervals in the record”; forty resets yield thirty-nine intervals. The rank it reported was right: 3.9 days ranked twentieth of thirty-nine, which is the exact middle. One further detail found in the same re-derivation and recorded rather than smoothed: a 3.9-day gap corresponds to 01:08 UTC on 5 August, about three hours after this entry's own 22:00 UTC dateline, so “as of this writing” ran later than the publication slot. Corrected 7 August 2026 while re-deriving the series from the status ids for that day's entry.
- News
August 4: the year that was missing from our own Epstein file.
Published saying the en banc Eleventh Circuit held that the Crime Victims' Rights Act “confers no rights enforceable before the government files charges.” That is the one question the court expressly refused to answer — it wrote that it “needn’t decide” whether those rights attach pre-charge, and both dissents criticise the majority for declining it. What the court actually held, 7-4, is narrower: the Act provides no private right of action to enforce CVRA rights in a freestanding civil suit outside a preexisting proceeding. The entry and the matching case-file record were corrected on 5 August 2026 after re-reading the opinion.
- Journal
July 29: a limit with no published date is not a schedule.
Published under the heading “Nobody published the date.” A dated OpenAI staff post from July 28 had already committed to restoring the five-hour limit the next day. The entry verified that the July 12 record carried no end date and reported that as the absence of any record.
Case-record corrections 8
Changes to a source anchor or a dated timeline row inside a case file, with the superseded text kept readable.
- Case record
Epstein public records
This is the worst error a board about unanswered requests can make: it recorded a silence that did not happen, and it had done so for eleven months. The response was reported twelve days after the request and was findable the whole time. “No response located” is written to be a statement about what this desk can see rather than an accusation, but a reader looking at a row that says nothing came back does not draw that distinction, and the burden was never on them to. Found by an adversarial audit of this desk's own board, in which every live row was handed to a reader instructed to assume it was wrong. The general lesson is now a standing rule: a ledger of non-response has to be re-checked for responses that arrived, because absence is the one status that decays into a false claim on its own.
Was: Status: no published response located. “No published Treasury response to this request has been located by this desk. Subsequent committee releases have concerned records obtained from other custodians.”
Now: Status: answered. Treasury responded; reporting on 12 September 2025 quotes a Treasury letter to the Committee saying it “plans to fully cooperate with the Committee and intends to provide all information responsive to your request”. The row now states that the letter itself is unpublished, so the response is known through reporting rather than from the document, and that whether records were produced is not settled here.
- Case record
Federal improper payments oversight
The current total, 53, had been copied into the prior-count slot, and the prior letter was misdated by a month. GAO-26-108980's own footnote 3 cites GAO-25-108402 of 2 June 2025 identifying 79, and the letter's arithmetic corroborates it: 79, less 17 implemented, less 16 that lost priority status, plus seven newly added, is 53. The error understated the department's own record — the true figures show the largest single reduction on the board, from 79 to 53. A pinned test was supposed to prevent exactly this drift and instead froze the mistake, because it had been written from the row rather than from the letter.
Was: Prior letter: May 2025, identifying 53 priority recommendations.
Now: Prior letter: June 2025, identifying 79 priority recommendations.
- Case record
Epstein public records
9 July was the date of the press release, not of the letter it published — the department's own link names the file 6-30-26. Attributing a quotation to the wrong document by nine days is small, and it is exactly the kind of thing a reader checking this desk's work would find first.
Was: “The state's 9 July 2026 letter states that 'more than 130 days have now elapsed since the NMDOJ's initial request'.”
Now: The letter is dated 30 June 2026 and was published by the department on 9 July 2026. The note now also records that reporting puts the first request at 13 February 2026 and a formal Touhy submission at 3 May 2026, neither carried by a document this desk has read.
- Case record
Epstein public records
Ranking Member is the senator's title as of the August 2026 report, not as of the April 2024 letter — he chaired the committee then. Carrying the later title on the earlier ask misstates who held what office when, which is a detail an opponent of the record checks first. The status was also resting on a sentence about all the banks rather than the one about this one; the specific sentence is both stronger and narrower. The same report is a ranking member's staff product rather than a committee-adopted document, and both rows drawn from it now say so.
Was: Asker: “Senate Finance Committee, Ranking Member”. Status rested on the report's sentence that “each and every bank refused to cooperate”.
Now: Asker: “Sen. Ron Wyden, then Chairman, Senate Finance Committee”. The declined status now rests on the report's row-specific sentence, that attorneys for the bank said they would not cooperate unless subpoenaed, in a reply the report footnotes to 6 May 2024.
- Case record
Epstein public records
The old sentence was not in the cited release, nor in the eighteen-page staff memorandum it links. The release concerns one bank, not several; says nothing about payments being reported years late; and its only $170 million figure is an individual's payments to Epstein, not a bank's reporting failure. The wrong version was authored on 1 July 2026 and had no source from the day it was written. It was found by an adversarial audit of this desk's own case files and confirmed against the release and the memorandum before being replaced. This desk's automated claim checker had marked the sentence confirmed, because the string $170 million does appear on the page — attached to a different subject. That limit is now stated on the source ledger.
Was: The committee investigation states several banks reported large Epstein payments only after he was charged, one flagging more than $170 million in payments years late.
Now: The release states Senator Wyden obtained Leon Black's settlement with the government of the U.S. Virgin Islands and released new information on Black's payment of $170 million to Epstein over several years, which the release describes as ostensibly for tax and estate planning services.
- Case record
Federal AI safety governance
A press release cannot report events that happened after it was published — in this instance by 453 days. A reader following the citation to check the 2025 vote would have found a 2024 announcement that says nothing about it.
Was: One anchor for the September 2024 Operation AI Comply announcement, carrying facts about the October 2024 Rytr consent notice and the December 2025 vote to set that order aside.
Now: Three anchors: the September 2024 announcement, the Federal Register consent notice at 89 FR 80565, and the Commission's own record of the order setting it aside.
- Case record
Epstein public records
The Eleventh Circuit's en banc opinion in In re Wild is stamped 'Date Filed: 04/15/2021' on every page and carries '(April 15, 2021)' on its caption. The record had been one day out since the anchor was added, and was caught only by opening the PDF rather than reading a summary of it.
Was: Published 16 April 2021.
Now: Published 15 April 2021.
- Case record
Boeing 737 MAX safety oversight
One entry carried two accidents at the date of the first, so the file recorded nothing at all for 2019 — the year of the second crash and of the grounding. A case file about FAA oversight of Boeing did not contain the FAA grounding the aircraft. Found by a check for calendar years that hold no entries while the years on both sides do.
Was: A single timeline entry, 'Two 737 MAX crashes prompt federal investigations', dated 29 October 2018.
Now: Separate dated entries for Lion Air flight JT610 on 29 October 2018 and Ethiopian Airlines flight ET302 on 10 March 2019, plus the FAA's emergency grounding order of 13 March 2019 and its rescission on 18 November 2020.