Corrections

Everything we shipped to readers and later found to be wrong is listed here: factual errors (a number, a quote, a date, an attribution) and wording in an inference or summary that said something the evidence did not support — with which layer failed, and what we changed because of it. There is no filter on this list.

What belongs here, and what doesn't

In scope: things we got wrong ourselves and had already sent out — a factual error in a number, a quote, a date or an attribution, or wording in an inference or summary that said something the evidence did not support. Found means fixed, and fixed means listed.
Not someone else's correction: when a source, an analyst or a vendor corrects themselves, that is news content, not a correction by us.
Not errors caught before publication: no reader ever saw them, so listing them here would invent a past that never happened. Those are fixed directly in the draft and the research graph.
Whether a forward call came true is not on this page: a call that misses is not a correction, and it is not rewritten afterward — it stands on its verdict card, is checked against the rule written in advance on its scheduled date, and the result goes on that card. If a rule itself was ambiguous, the correction is dated and explained on that card.

Corrected 2026-10-22Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The October 18 brief set CoreWeave's self-reported short-lease price of about US$40 million per MW per year against our own estimate of US$10 million to US$16 million for five-year leases, and called the 2.5 to 4 times gap a term premium.

What was actually the case
The two sides do not match on GPU generation, capacity basis or signing date (the body says so), so we cannot tell how much of that multiple is term premium.

Which layer failed
The long-lease figure was our own estimate from total contract values, on a different basis from the self-reported short-lease figure, yet the multiple appeared in the headline, At a glance and key point without the caveat.

What we did about it
On 2026-10-22 a correction-and-addendum note was added to the three-language archive pages; the body is unchanged.

Editions affected: /issues/2026-10-18/, /en/issues/2026-10-18/, /ja/issues/2026-10-18/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-22Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The October 17 brief's Main line 3 said the entry-level share of Anthropic's technical staff fell from 21% to 14%, and that the share fell because sales, finance and recruiting grew faster.

What was actually the case
In the report, that reason explains the technical share of all Anthropic staff (61% to 56% in a year), not the drop in the entry-level share inside the technical staff; we did not find the report's explanation for the entry-level drop. The 21% to 14% and 27% to 9% figures are unchanged.

Which layer failed
When paraphrasing the source, the body joined two shares with different denominators (technical staff of all staff, entry-level of technical staff) in one causal sentence.

What we did about it
On 2026-10-22 a correction-and-addendum note was added to the three-language archive pages; the body is unchanged.

Editions affected: /issues/2026-10-17/, /en/issues/2026-10-17/, /ja/issues/2026-10-17/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-21Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The August 13 issue, in Chinese, English and Japanese, said Nebius's adjusted EBITDA turned positive in the second quarter: positive US$236.2M for the quarter against negative US$53.7M in the previous one. The At a glance item and the headline of Main line 3 also called it the first time.

What was actually the case
In Nebius's first-quarter filing (Form 6-K, May 13, 2026), the Adjusted EBITDA row of the financial highlights table reads "(53.7) 129.5", and the first column is the first quarter of 2025: the US$53.7M loss was a year earlier, and the first quarter of 2026 was positive US$129.5M. The company's fourth-quarter filing says the US$15.0M of the fourth quarter of 2025 was its first quarter of positive adjusted EBITDA at the group level. The second quarter was the third positive quarter in a row, US$106.7M above the first, not the first. The same item's capex of about 9.7 times revenue and the unraised full-year revenue guidance were right.

Which layer failed
Bad data: the issue did not retrieve the filings and carried over an older figure in our records that had taken the year-earlier column of the financial highlights table for the previous quarter.

What we did about it
On 2026-10-21 a correction note was added at the top of the Chinese, English and Japanese archive pages, withdrawing the "first time" half and stating that the other half of the conclusion is unchanged; the text itself is left as published. The matching quarterly records in our knowledge graph were corrected the same day. The emails and the Japanese note post of August 13 are the original versions.

Editions affected: /issues/2026-08-13/, /en/issues/2026-08-13/, /ja/issues/2026-08-13/
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-21Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The October 20 issue set the 20 to 33GW our lease conversion produced against SemiAnalysis's forecast of about 50GW added in 2027, wrote that leases "can't carry 50GW", and set a rebuttal condition that 50GW fails if capex guidance grows less than in 2026.

What was actually the case
Leases are only one of SemiAnalysis's bases. The post also lists sites in late-stage construction, mostly delivering in 2027, plus self-builds and construction outside the US, with detail in the paid section we did not read. So 20 to 33GW cannot rebut 50GW, and a growth rate in capex cannot decide a level of capacity.

Which layer failed
Inference layer: a part was treated as a rebuttal of the whole, and a growth rate was used to judge a level.

What we did about it
On 2026-10-21 we rewrote the headline, the 30-second summary, the judgment update and the key point on the three archive pages, and changed the rebuttal condition to whether the earnings calls disclose comparable completed capacity, settling as cannot be determined if they do not. A correction note was added. The emailed issue and the Japanese note of October 20 are the original versions.

Editions affected: /issues/2026-10-20/, /en/issues/2026-10-20/, /ja/issues/2026-10-20/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-19Edition with the error: English editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The English and Japanese versions of the verdict card on whether model companies are becoming do-everything integrators (call made July 27) said Anthropic's officially confirmed $2.5B in Claude Code revenue was one-fifth to one-third of a total revenue estimate spanning $14-30B (the Japanese said 10–30%).

What was actually the case
$2.5B divided by $14-30B is about 8% to 18%, so one-tenth to one-fifth; divided by the $14B official annualized figure alone it is about 18%. The Chinese card and the July 27 deep column, in all three languages, already said one-tenth to one-fifth.

Which layer failed
Bad inference — the inputs were right, but when the English and Japanese cards were rewritten into plain language the share was overstated and not recomputed against both ends of the band.

What we did about it
On October 19, 2026 we changed the share on the English and Japanese card pages to one-tenth to one-fifth and added a correction note at the end of the call section. The direction of the call does not change. The Chinese card and the deep column were not affected.

Editions affected: /en/cards/v-2026-07-27-01/, /ja/cards/v-2026-07-27-01/
Language editions affected: English, Japanese
Corrected 2026-10-18Edition with the error: English editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
Chips & semiconductors item 1 in the October 10 English edition called Lite-On's US$350M project in McKinney, Texas, "a high-voltage DC power plant," which reads as a power station.

What was actually the case
The post relayed by Dan Nystedt says Lite-On signed a US$350 million (NT$11.06 billion) contract for construction of its McKinney plant, which is aimed at high-voltage DC power racks and AI infrastructure production. It is a factory, not a power station. The Chinese edition called it a high-voltage DC power-supply factory and the Japanese edition a power-supply factory; both were right.

Which layer failed
Inference layer: the English rewrite turned "power-supply factory" into "power plant," which in English means a power station, and so changed what the facility is.

What we did about it
On 2026-10-18 the line on the English archive page was rewritten as "has separately signed a US$350M construction contract for a plant in McKinney, Texas, that will make high-voltage DC power racks," with a correction note at the top of the page that links the original post. The Chinese and Japanese editions were not affected. The English email sent on 10/10 is the original version.

Editions affected: /en/issues/2026-10-10/
Language editions affected: English
Corrected 2026-10-18Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
Entry COR-2026-07-09-01 on our corrections page in Chinese, English and Japanese (missing the court rulings that had blocked three federal pressure moves against Anthropic) said in its fix that "court rulings became a mandatory source for this topic."

What was actually the case
No such rule has ever been in our writing rules or source lists. The rule that does exist is the current-status check added on July 8: reversible states such as bans, suspensions, lawsuits and personnel moves must be checked against same-day official or mainstream reporting before they run. It does not require checking court rulings.

Which layer failed
Bad data: the entry was written when the corrections page launched on August 15, and its fix described a rule that was never put in place; the rule files were not checked.

What we did about it
On 2026-10-18 this entry records what is actually in place. The original entry is left as published, since entries on this page are added, never rewritten. There is currently no rule making court rulings a mandatory source for this topic.

Editions affected: /corrections/, /en/corrections/, /ja/corrections/
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-18Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
Entry COR-2026-09-30-07 on our corrections page in Chinese, English and Japanese (two sentences added at the end of the first "Chips and semiconductors" item in the 9/30 Japanese edition) gave as its cause that "sentences that are in neither the master text nor the sources were added during translation, and the translation fidelity check did not catch them."

What was actually the case
They were not added in translation. In the first round of the Japanese edition's review panel, the chair accepted a note from a reader seat and told the rewriter to add a sentence drawn from a longer internal research draft of ours; the rewrite added two. The export-control claim came from that internal draft and is in neither the Chinese master text nor the sources the item cites. The fidelity check compares only numbers and links, and it logged the rewrite as an instruction carried out.

Which layer failed
Inference error: when that correction was written, the cause was inferred from the symptom (the sentences were in the Japanese edition and not the Chinese one) without checking the review panel's records.

What we did about it
On 2026-10-18 this entry records the actual cause. The original entry is left as published, since entries on this page are added, never rewritten. The removal of the two sentences from the 9/30 Japanese archive page stands. The Japanese review panel now takes material only from the Chinese master text.

Editions affected: /corrections/, /en/corrections/, /ja/corrections/
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-18Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The opening note on sources in the October 9 Chinese edition said the body used 24 outside sources with links, while the same issue's "Sources & accounting" section and footer said 25.

What was actually the case
Recounting by the rule the "Sources & accounting" section states (only links the body actually cites, excluding our own domain), the public body has 25 distinct outside links. 25 is right; the 24 in the opening note was wrong.

Which layer failed
Bad data: the opening note's count didn't match the other two places in the same issue. The pre-publication count-consistency check flags this kind of mismatch, but it doesn't block publication on its own, and the issue went out.

What we did about it
On 2026-10-18 the opening note on the Chinese archive page was changed from 24 to 25, with a correction note at the top; the body and conclusions are unchanged. The English and Japanese editions said 25 in all three places and were not affected. The Chinese email sent on 10/9 is the original version.

Editions affected: /issues/2026-10-09/
Language editions affected: Chinese
Corrected 2026-10-18Edition with the error: English editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
In main line item 1 of the October 17 English edition (Anthropic's Claude Code team on how it reviews code), the "Not proven yet" line said "the 40%-plus drop in reverts is self-reported," which reads as a fall of more than 40% in the number of reverted changes.

What was actually the case
Anthropic's post says that from June to September "merged pull requests to Claude Code's codebase more than doubled while the share that had to be reverted fell by more than 40 percent." What fell by more than 40% is the share of changes reverted, not their number. With merges more than doubling, the number of reverts could have risen or fallen, and the post doesn't give it. The quote two paragraphs earlier in the same item, and the Chinese and Japanese editions, all describe a share.

Which layer failed
Inference layer: the English rewrite shortened "the share that had to be reverted" to "reverts" and so changed what the figure measures. The English edition's pre-publication fidelity check compares numbers and links; the number was right and what it measured was not, so the check passed.

What we did about it
On 2026-10-18 the line on the English archive page was rewritten as "the more-than-40% fall in the share of changes that had to be reverted," with a correction note at the top of the page. The Chinese and Japanese editions were not affected. The English email sent on 10/17 is the original version.

Editions affected: /en/issues/2026-10-17/
Language editions affected: English
Corrected 2026-10-18Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
On the Chinese and Japanese archive pages of the October 18 brief, the caption for item 10 (Chollet on agents writing the code while engineers get busier) called "the bottleneck may just move elsewhere" a personal judgment by Chollet and Box CEO Aaron Levie, crediting both men with the bottleneck view.

What was actually the case
The bottleneck explanation is Chollet's alone. Levie quoted Chollet's post to reject its premise that software demand is fixed and automation cuts labor demand, arguing that software demand isn't fixed; neither attached data, and pairing the two posts was our choice. The item's own text said this correctly; only the caption was wrong.

Which layer failed
Inference layer: an internal summary line attached to the item merged the two men's views into one judgment, the caption was written from that line, and the pre-publication cold reads of the Chinese and Japanese pages did not check the attribution against the item. The English page's cold read caught the same error before it went live.

What we did about it
On 2026-10-18 the captions on the Chinese and Japanese archive pages were rewritten to credit the bottleneck view to Chollet and give Levie's rebuttal of the premise, then re-reviewed and republished; the Chinese and Japanese versions of the internal summary line were corrected as well. The English version of that line was left unchanged; that line is never published. The gap that let captions copy attributions from that line is being fixed separately in the illustration pipeline. The English edition never ran the wrong caption.

Editions affected: /issues/2026-10-18/, /ja/issues/2026-10-18/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-18Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The August 16 column "Are model companies all buying share at a loss, with nobody daring to raise prices?" said in two places, in English, Chinese and Japanese, that Anthropic had said profitability may not last the full year. The prediction card opened that day and due October 31 (did Anthropic's first profitable quarter materialize, with no return to a loss in Q3?) said the same on its card page in all three languages.

What was actually the case
Going back to the May reports on Anthropic's second-quarter figures from CNBC, Reuters and The Decoder, none contains that statement or a company warning to that effect; the Wall Street Journal article is paywalled and we did not read it. We cannot find a source for the statement.

Which layer failed
Data layer — the line entered our database on July 8 as a "key caveat" written by a model verification batch, with no original quote attached and no one re-checking it; the column and the card repeated it as something the company had said.

What we did about it
On 2026-10-18 both mentions were removed from the column's three archive pages, and the sentence was removed from the card page in all three languages with a correction note. The card now also states that reports tracing back to the same origin count as one; that is our existing admission rule, and the question and how it is answered are unchanged. The database line is now marked as unsourced and not to be cited. The email sent on August 16 is the original version.

Editions affected: /column/2026-08-16/, /en/column/2026-08-16/, /ja/column/2026-08-16/, /cards/v-2026-08-16-02/, /en/cards/v-2026-08-16-02/, /ja/cards/v-2026-08-16-02/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-18Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The July 20 deep dive "Selling Tokens Finally Makes Money" in English, Chinese and Japanese credited CNBC's May 20 report for compute cost per dollar of revenue falling from 0.71 to 0.56 and for the quarter swinging "from roughly breakeven to $559 million in operating profit," in wording that read as if the profit had already happened. The July 20 brief's deep-dive section (English and Chinese) and item 5 of the July 31 brief (English, Chinese and Japanese) also credited CNBC for the 0.71-to-0.56 figure.

What was actually the case
That CNBC story only confirmed the $10.9 billion second-quarter revenue target and that hitting it would mean a first profitable quarter; neither figure is in it. The Wall Street Journal reported both first that day, and Reuters wrote the same day that the figures came from fundraising materials the company gave investors, with $559 million as expected second-quarter operating profit. 0.71 is the first-quarter figure; 0.56 is a second-quarter projection.

Which layer failed
Data layer — our database held a quote tagged with the CNBC link that a model verification batch wrote on July 8; refetching CNBC was blocked that day and nobody re-checked it. The pieces relied on it, and no step before publication checked figures and attributions against the original.

What we did about it
On 2026-10-18 the eight archive pages (the July 20 deep dive in three languages, the July 20 brief in English and Chinese, the July 31 brief in three languages) were changed to present the figures as a company projection reported by the WSJ, with the sources replaced and a correction note at the top of each page. The deep dive's conclusions don't rest on this sentence and are unchanged. The database entry that had been upgraded to a confirmed claim on this basis was downgraded to a single origin. The emails sent on July 20 and 31 are the original versions.

Editions affected: /deep/2026-07-20/, /en/deep/2026-07-20/, /ja/deep/2026-07-20/, /issues/2026-07-20/, /en/issues/2026-07-20/, /issues/2026-07-31/, /en/issues/2026-07-31/, /ja/issues/2026-07-31/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-17Edition with the error: Japanese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The Japanese edition of the July 15 issue (lead item 1, positioning note) said demand is backed by firm orders while execution “has no backing yet”, which contradicts the sentence before it saying the evidence is firmer on the execution-bottleneck side.

What was actually the case
The Chinese original says demand is already locked in by firm orders while execution (power, cooling, delivery) has not caught up, which is why the execution-bottleneck call has firmer evidence than a call on demand direction.

Which layer failed
This Japanese page was added to the site on October 17 as a back issue; a rewrite step during review turned “has not caught up” into “has no backing yet”.

What we did about it
The Japanese page now says demand is locked in by firm orders while execution has not caught up. The call itself is unchanged in all three languages.

Editions affected: 2026-07-15
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-17Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
Model watch item 1 and "Still missing today" in the 10/17 issue listed Epoch AI's conversion method and full leaderboard for its Cyber ECI cyber-capability index among other gaps and said each "either hasn't been reviewed or doesn't exist publicly."

What was actually the case
Epoch's data page published the same day, October 16, gives the method: it takes the two closed-weight models whose scores most closely bracket GLM-5.3 — GPT-5.4, released March 5 and scoring below GLM-5.3, and Anthropic's Mythos Preview, released April 7 and scoring above it — which were seven and six months old at Epoch's data cutoff. The page lists GPT-5.4 at 157.04 and GLM-5.3 at 160.10, with a CSV download and a CC BY license.

Which layer failed
Data-layer miss — only Epoch's post on X was read (it carried an image and no link to the page), Epoch's own same-day page wasn't opened, and the method was listed as possibly nonexistent.

What we did about it
On 2026-10-17 a correction and addendum was added to the archived issue page in English, Chinese and Japanese, with the method and both scores at the end of Model watch and Sources & accounting. A writing rule now requires opening the institution's own page when a claim comes from its social post; if it wasn't opened, the issue says that page hasn't been reviewed, never that the information doesn't exist. The email sent on 10/17 is the original version.

Editions affected: /issues/2026-10-17/, /en/issues/2026-10-17/, /ja/issues/2026-10-17/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-16Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The verdict card on whether retention of AI subscriptions over $250 a month will drop below the warning line (call made July 31) said the call would be settled by "the next a16z/ChartMogul benchmark report" and described the 70% gross retention and 85% net retention baseline as figures cited in an earlier SecondSource call.

What was actually the case
The 70% gross revenue retention and 85% net revenue retention for the above-$250-a-month tier come only from ChartMogul's December 10, 2025 retention report. a16z's report, The Great Expansion, does not contain these tiered figures, and the two firms have not published a joint benchmark report.

Which layer failed
Wrong source — the card reused an earlier call's citation of these figures without reopening and checking both originals, and merged two sources into one.

What we did about it
On October 16, 2026 the card pages in all three languages were changed to name ChartMogul's retention report as the source, with a dated correction note on the card, and the ChartMogul report is now listed as a source readers can open and check. The call itself is unchanged. The conflict between the card's threshold and its baseline (the 85% baseline is already below the 100% threshold) is being revised separately; once settled, the card will state the date, the original wording and the reason.

Editions affected: /cards/v-2026-07-31-05/, /en/cards/v-2026-07-31-05/, /ja/cards/v-2026-07-31-05/
Whether other language editions were affected: not stated in this record.
Corrected 2026-10-16Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
Recapping earlier issues, the judgment update in main-line item 2 of the 9/27 issue said the piece added on 9/26 was that 'the lab itself says disclosure is by design not the full picture'.

What was actually the case
OpenAI's CEO said only that holes found by OpenAI agents at other companies are those companies' call to disclose; those companies can still disclose the holes, so the lab's account is not necessarily incomplete. The accurate reading is that the lab's public account can't be assumed to cover every incident, the same reading the 9/26 issue has carried since its 9/30 correction.

Which layer failed
Bad inference — recapping the previous issue, the 9/27 issue kept the strong reading the 9/26 issue went out with, turning 'the other company decides whether to disclose' into the lab conceding its disclosure is not the full picture, overstating both who said it and how certain it is; when the 9/26 issue was corrected on 9/30, this sentence was not corrected with it.

What we did about it
On 2026-10-16 that sentence on the archive page was changed to say that OpenAI's CEO said holes found by OpenAI agents at other companies are those companies' call to disclose, so the lab's public account can't be assumed to cover every incident, with a note at the top; the English and Japanese editions carried the same sentence and were corrected the same day with their own notes. The conclusion of main-line item 2, that what the outside world sees of incidents depends on what outsiders can find, is unchanged. The email sent on 9/27 is the original.

Editions affected: /issues/2026-09-27/, /en/issues/2026-09-27/, /ja/issues/2026-09-27/
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-15Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The investor note in main-line item 1 of the 9/12 issue said that of the $530.5 billion, only $108.5 billion is credit exposure and that "the rest is procurement and investment commitments Nvidia will pay for itself."

What was actually the case
Nvidia's quarterly report (filed 2026-08-26) also lists, in the two commitments tables, $45 billion of data center leases not yet commenced: $25 billion for Nvidia's own use and $20 billion signed for third parties, which Nvidia expects to reassign to third parties. The $36 billion of AI cloud agreements in the additional-commitments table shrink as third parties use the capacity. The point that only $108.5 billion is guarantees is unchanged.

Which layer failed
Bad inference — the figures in the filing were read correctly, but every non-guarantee item in the two commitments tables was lumped together as "procurement and investment," leaving out leases as a category, and the portion that shrinks as third parties take up capacity was written as certain to be paid by Nvidia.

What we did about it
On 2026-10-15 the sentence was changed on the archive pages in all three languages to "procurement, lease and investment commitments," noting that the AI cloud agreements shrink as third parties take up capacity and that Nvidia expects to reassign the leases signed for third parties; each page gained a note at the top. A table listing what each of the three tables contains was added to main-line item 1 of the Chinese and English editions. The email sent on 9/12 is the original.

Editions affected: /issues/2026-09-12/, /en/issues/2026-09-12/, /ja/issues/2026-09-12/
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-12Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The 10/11 main item said we had not read the security notice linked from Fireworks' CEO's post and, as of early October 11, had found no follow-up from the company. The Open SWE item in Product moves linked to a different post by Harrison Chase and said the models used before and the quality bar had not been published. The Chips column labeled Terafab's US$16.8 billion first phase 'a single post, unverified', and Named commentary said we couldn't find the exact publication date of Jordan's interview. The Chinese edition's main item also called the Altman remark one made 'the day before'.

What was actually the case
Fireworks' security notice says the company identified the incident on October 6 and that API keys and tokens stored as secrets may have been captured in certain affected accounts; the notice was updated on October 10. The 64% post is at a different address, and LangChain's October 1 blog post sets out the control and the quality measure. The Texas governor's office announced on August 6 that Terafab's first phase is a capital investment of more than US$16.8 billion. The English translation of Jordan's interview says Libération published it on October 4. October 9 is the day the Altman quote was reposted; when he said it is not known.

Which layer failed
Bad data — before publication we did not open the notice and the blog post that the posts linked to or check dates against the original pages, and the labels on two post addresses were swapped in the draft.

What we did about it
A correction line is queued for the October 13 Chinese edition. On 2026-10-12 the three archive pages were corrected (the Altman point appears only in the Chinese edition) with a note at the top, and their illustrations were reviewed again and put back. The judgments are unchanged. The email sent on 10/11 is the original.

Editions affected: 2026-10-11
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-06Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The 7/25 deep column (No. 14) said "only 11% of teams switched model vendors in the past year; 66% upgraded to newer models inside their existing vendor," linking to Menlo Ventures' year-end 2025 report, and said Anthropic's inference-serving gross margin, "per internal and analyst accounts," climbed "from 38% in 2024" to around 65% in 2026. The deep-dive section of the 7/25 brief and the 8/4 brief linked the same 11% to the same year-end report.

What was actually the case
In the year-end report, 11% is the market share of open-source models; the switching figure comes from Menlo's 2025 mid-year report: "Only 11% of teams report changing model providers in the past year. But 66% upgraded to newer models from their existing vendor." The margin figures are an estimate by the research firm SemiAnalysis: "inference gross margins are now in the mid 60s, up from 38% in 2025 and -94% in 2024" — 38% is for 2025, not 2024.

Which layer failed
Bad data — the numbers were right, the source and year were not: the 11% was attached to the URL of a different report from the same firm, the margin year was copied one year off, and an analyst firm's estimate was described as "internal and analyst accounts." The trend tree in the same piece says -94% for 2024, and the draft was not checked against it.

What we did about it
On 2026-10-06 the archive page moved the 11% link to Menlo's 2025 mid-year report, changed the margin passage to "per an estimate by the research firm SemiAnalysis, from 38% in 2025" with the SemiAnalysis source cited, and added a note at the top; the Chinese edition was corrected the same day with its own note. The same link in the deep-dive section of the 7/25 brief (all three languages) and in the 8/4 brief (all three languages) was corrected the same day with notes. The conclusion does not rest on either passage and is unchanged. The emails sent on 7/25 and 8/4 are the originals.

Editions affected: /deep/2026-07-25/, /en/deep/2026-07-25/, 2026-07-25, 2026-08-04
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-05Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The verdict card on whether large models will become swappable commodities (call made July 25) had a Chinese headline saying “all three numbers currently point the other way,” and its background said we “checked three real numbers and take the opposite side”; the English and Japanese cards said the thesis fails a stress test against all three of its own watchpoints, while the badge on the page says the call is pending.

What was actually the case
Only one of the three watchpoints has a before-and-after comparison: Menlo Ventures measured Anthropic rising from 32% in mid-2025 to 40% in December, with the top three at 88% combined. The 11% switching figure comes from a single Menlo survey in mid-2025, with no earlier figure to compare against. The margin item gave no number on the page; what exists is an outside estimate by the research firm SemiAnalysis (inference gross margin rising from 38% in 2025 to the mid-60s in the first quarter of 2026), unaudited. The card does not resolve until July 25, 2027.

Which layer failed
Bad inference — the material was right: the July 25 deep column said each of the three readings “comes with a caveat that cannot be ignored,” that the 11% has no reliable baseline, and that the margin figures are unaudited. Rewriting it into the card's plain-language layer turned three watchpoints into “three numbers” and dropped the caveats, and the English and Japanese cards stated our call as an outcome that had already happened.

What we did about it
On 2026-10-05 the card pages rewrote the headline and the call in all three languages: the call is stated as ours and resolves on July 25, 2027; each of the three items now gives its number, its baseline and what kind of source it is; and the call section carries a correction note. The last line of the Chinese background was changed to match. The same day the card added the readings from the September 30 interim check: Menlo has not published a new report, and Claude Sonnet 5's introductory price became its standard price, so under item (3) of the receipt the price item was revisited. The direction of the call (only a layered version holds, not across-the-board commoditization) is unchanged.

Editions affected: /cards/v-2026-07-25-01/, /en/cards/v-2026-07-25-01/, /ja/cards/v-2026-07-25-01/
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-05Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The 7/28 headline, the first At a glance item and core item 1 said that of ten Chinese labs exactly one, MiniMax, had "followed" K3 with a gate of its own, and that MiniMax's license was "published July 23"; the third At a glance item and the headline of core item 3 said the "pick a model for me" distribution layer "has been shown to make money."

What was actually the case
Hugging Face's file history shows the MiniMax M3 license was uploaded with the model page's first commit on June 12 and has not changed since; July 23 is only when the page's description was last updated. It came more than a month before K3's license (July 27), so MiniMax did not follow K3. The only evidence for the distribution-layer item is an OpenRouter board member's own account, with no figures, and the item itself says confidence does not move.

Which layer failed
Bad data — the license date was taken from the model page's last-updated date rather than the date the license file was uploaded, and "followed" was written from that wrong date; the same day's deep column already said MiniMax came before K3, and the brief was not checked against it. The distribution-layer headline said more than its own item.

What we did about it
On 2026-10-05 the archive page changed the three passages to say that besides K3 only one lab has a similar gate and that it came first, gave the June date, changed the two distribution-layer headlines to "is said to make money," and added a note at the top; the finding that only two of the ten put a revenue gate in their license stands. The Chinese and Japanese editions were corrected the same day with their own notes, and the Japanese deep column of the same day, which put MiniMax's license at July 23 and four days ahead of K3, was corrected with a note. The email sent on 7/28 is the original.

Editions affected: 2026-07-28, /ja/deep/2026-07-28/
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-05Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The 8/29 deep column (No. 26) said the numerator and denominator of the "two-to-five-times gap" (a trading firm capturing US$200M–500M per MW, and a model company earning US$50M per MW) came from two different companies, that the trading firm was not that model company's customer, citing the transcript; its headline and summary said we had made the judgment "yesterday" in our August 28 issue, and main-line item 1 of the 8/29 issue said our August 28 issue argued it.

What was actually the case
The half-sentence the quotation dropped with an ellipsis reads "or Jane Street where they're one of Anthropic's biggest customers": the speaker calls the firm both a contract customer of OpenAI's fast mode and one of Anthropic's biggest customers. US$200M–500M divided by the US$50M on the page is four to ten times; "two to five times" divides by another figure the speaker gave late in the episode, US$100M per MW. The judgment was logged in our ledger on August 28, and that day's public issue did not carry it.

Which layer failed
Bad inference — the material was right: the transcript gives both denominators and calls the firm a customer of both labs. The draft took only the US$50M from the opening, the ellipsis happened to remove the half-sentence that contradicted the paragraph's point, and "we said yesterday" was written from the ledger's entry date without checking that day's public issue.

What we did about it
On 2026-10-05 the archive page rewrote section 1, the summary, the route paragraph, the section 2 line on the current ceiling and falsification condition 5 to follow the transcript, changed the headline and summary to "the judgment we logged on August 28" with a note that the issue did not carry it, and added a note at the top; the conclusion (the leading indicator cannot discriminate, watch list prices and quotas) does not rest on these and is unchanged. The Japanese edition was corrected the same day and its headline now separates the circulating consolation line from our judgment; main-line item 1 of the 8/29 issue was corrected in all three languages with its own notes. The email sent on 8/29 is the original.

Editions affected: /deep/2026-08-29/, /ja/deep/2026-08-29/, 2026-08-29
Language editions affected: Chinese, English, Japanese
Corrected 2026-10-04Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The 10/3 headline and email subject said that since July's Hugging Face incident all six parties that have acted 'are using existing law, and none is waiting for Congress to write a new one'; the main-line item 1 headline added that the federal side treats it as a reason no new law is needed.

What was actually the case
In the same issue's table, New Mexico's attorney general, according to a Politico reporter, is set to propose a new state bill, and the California attorney general's press release cited in that issue says he joined a bipartisan group of state attorneys general last month in asking Congress to legislate on large AI models now. The Washington Post sentence quoted in the item says only that the investigation 'could provide backing' for the Trump administration's position; that is the Post's reading, and we have not seen a federal official say so.

Which layer failed
Bad inference — the headline turned what most of the table shows into all six, and turned a newspaper's 'could' into something the federal side is already doing.

What we did about it
The Chinese and Japanese editions of October 4 carried a correction; on 2026-10-08 both headlines on the three archive pages were rewritten (most of the six used existing law and powers; the Post's 'could' kept) with a note at the top, and the illustration whose caption repeated the same sentence was taken down until a corrected one is made. Our working read, the item text and the table are unchanged. The email sent on 10/3 is the original.

Editions affected: 2026-10-03
Language editions affected: Chinese, Japanese
Corrected 2026-10-04Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The 'What happens when we're wrong' section of our methodology page said we had 49 calls tracked and 20 public as verdict cards, that nothing had come due yet, and that the table would start keeping score in September; the 'What we haven't done yet' section on the same page said the scoreboard hadn't settled a single call — zero track record.

What was actually the case
The ledger kept growing after that, so the count had long stopped being 49; the first call came due on 2026-08-24, and by 10-04 several had come due, each with its result on its own card. The admission-log and live-call counts in the gates section of the same page were also hard-coded at the time and had not been updated since.

Which layer failed
Bad data — the ledger figures and the due status were text hard-coded into the site's code in early August, and nothing updated them as the ledger grew and calls came due.

What we did about it
From 2026-10-04 the due-status text on the methodology page is generated from the call ledger on every site build: it says only whether calls have come due and where to read the results, with no counts; the card-link blurb was rewritten to match. The 'What we haven't done yet' item now says there is no public overall score yet and why; the gates section drops the hard-coded admission-log count, live-call count and top confidence score, and its line that more items get blocked or parked than admitted, which holds or not depending on how decisions are classified, now says instead that blocked and parked items are logged the same way and can be traced. The Chinese and Japanese pages were corrected the same day.

Editions affected: /methodology/, /en/methodology/, /ja/methodology/
Language editions affected: Chinese, English, Japanese
Corrected 2026-09-30Edition with the error: Japanese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
At the end of the first 'Chips and semiconductors' item in the 9/30 Japanese edition, two sentences had been added: one restating the preceding speculation, and one about export controls that appears neither in the Chinese edition nor in the sources.

What was actually the case
The export-control sentence cannot be found in the Chinese master text or in the sources the item cites; the restating sentence only repeated what came before.

Which layer failed
Bad data — sentences that are in neither the master text nor the sources were added during translation, and the translation fidelity check did not catch them.

What we did about it
On 2026-09-30 both sentences were deleted from the Japanese archive page, with a note at the top (the same note as the correction to that issue's Chinese edition). The Chinese and English editions did not have them and were not affected. The Japanese email sent on 9/30 is the original.

Editions affected: 2026-09-30
Language editions affected: Japanese
Corrected 2026-09-30Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
Main-line item 1 of 9/30 said second-quarter revenue times four, about $46B, matched May's self-reported $47B run-rate — 'the first time … can be checked against numbers in financial-statement form, and they line up' — and that 'most of that $42B isn't cash burn'; item 2 said Pichai confirmed 'a second document' was signed the same day.

What was actually the case
A quarter times four is an average pace and a run-rate is the pace at one point in time, and a leaked draft is not a filed, audited statement, so being close only means no obvious contradiction. SiliconANGLE calls the roughly $34B only an 'accounting charge related to the funding', without saying it is non-cash, and no report gives a cash-flow figure. The full text read on October 1 is titled 'Joint Commitment on Frontier Responsibilities': that is the accord's own name, not a second document.

Which layer failed
Bad inference — closeness between two differently measured numbers was treated as verification, 'accounting charge' was taken to mean 'not cash', and a phrase quoted in the press was read as a separate document.

What we did about it
On 2026-09-30 the first two passages on the archive page were rewritten (a net loss is not itself a cash-burn measure and the reports don't establish actual cash burn), 'What would prove this wrong' and 'what to track' were revised, outside sources went from 27 to 29, and a note was added at the top; on October 4 the second-document sentence was rewritten and 'second document' was removed from the Verification and sourcing sections. The English and Japanese editions were corrected the same day with their own notes. The email sent on 9/30 is the original.

Editions affected: 2026-09-30
Language editions affected: Chinese, English, Japanese
Corrected 2026-09-30Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The judgment update of main-line item 4 on 9/29 said xAI's bot 'shares memory and skills'.

What was actually the case
The official wording quoted in the same item says 'The Bot keeps separate context and memories for each user while drawing on the skills shared across the team': skills are shared, while each user's context and memories are kept separate.

Which layer failed
Bad inference — the quote was right, but 'shared skills, separate memories' was read as both being shared.

What we did about it
On 2026-09-30 the archive page was changed to say the bot shares skills and credentials, with the qualifier that, since xAI says context and memories are separate, there is no evidence yet that such a path exists, and a note at the top. The English edition was corrected the same day with its own note; the Japanese edition had already fixed the misreading at publication and the same qualifier was added that day. The email sent on 9/29 is the original.

Editions affected: 2026-09-29
Language editions affected: Chinese, English
Corrected 2026-09-30Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The opening and the 'Sources & accounting' section of the 9/27 issue called the Intel 18A chip teardown the fourth main-line item.

What was actually the case
The issue has only three main-line items; that one had already moved to item 5 of the briefs at publication, and we had not verified it.

Which layer failed
Bad inference — when the item was moved out of the main line, only the item itself moved; the pointers in the opening and the inventory, built from the main-line list, were not recomputed (the same pattern as the 9/14 issue).

What we did about it
On 2026-09-30 both places on the archive page were changed to point to brief item 5, with a note at the top; the Japanese edition had the same two places and was corrected the same day with its own note. The English edition already said item 5 and was not affected. The email sent on 9/27 is the original.

Editions affected: 2026-09-27
Language editions affected: Chinese, Japanese
Corrected 2026-09-30Edition with the error: English editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The headline of main-line item 1 in the 9/26 English edition said an agent posted '53 ChatGPT users' images', which reads as images from 53 people.

What was actually the case
The count is of images, not people: OpenAI's agents posted 53 images from ChatGPT users, and neither OpenAI nor the reporting says how many users they came from, as the verification paragraph already noted; the images went to several outside image hosts as unlisted links.

Which layer failed
Bad inference — the possessive in the English headline turned a count of images into a count of people, and the body described the hosts as a single site.

What we did about it
On 2026-09-30 the English archive page headline was changed to '53 images from ChatGPT users' and the body now says several outside image hosts, matching Fortune, with a note at the top. The Chinese and Japanese editions count images with a measure word (53 張, 53 枚) and were not affected. The English email sent on 9/26 is the original.

Editions affected: 2026-09-26
Language editions affected: English
Corrected 2026-09-30Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The 9/26 headline, the first At a glance item, several parts of main-line item 1 and the takeaway said OpenAI 'concedes' its disclosures of agent incidents 'won't be the whole picture'.

What was actually the case
OpenAI's CEO said only that vulnerabilities its agents find at other companies are those companies' call to disclose; those companies can still disclose them, so the lab's account is not necessarily incomplete. The accurate reading is that we can't assume the lab's public account covers every incident.

Which layer failed
Bad inference — the quote was right, but 'the other company decides whether to disclose' was read as 'OpenAI admits its disclosures are incomplete', overstating both who said it and how certain it is.

What we did about it
On 2026-09-30 those lines on the archive page were changed to say we can't assume the lab's account covers every incident, with a note at the top; the English and Japanese editions were corrected the same day with their own notes. The email sent on 9/26 is the original.

Editions affected: 2026-09-26
Language editions affected: Chinese, English, Japanese
Corrected 2026-09-30Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
Main-line item 3 of 9/25 said 'measured by contract value, the active share is 2.5%, lower than August's 8.7%'; item 1 said the State Land Office 'repeatedly rejected the route permit' and the campus was 'stuck' on it, and presented an Investing.com paraphrase as Nscale's filing wording.

What was actually the case
2.5% is a share of contract value and 8.7% a share of GPUs, with different denominators and units, so neither is lower; on the S-1's like-for-like figures, the contract-value share rose from about 1.3% to about 2.5%, and the chip-count share was about 5.4% at the end of August. Bloomberg reported denials for one proposed pipeline route, later changed; Oracle's notice preserves its right to defer payments if the campus is not online in 2028, and Oracle says such notices do not by themselves mean a delay.

Which layer failed
Bad inference — two shares measured differently were compared directly, qualifiers in the reporting (a proposed route, if it fails to come online) were read as settled facts, and a paraphrase was not marked as one.

What we did about it
On 2026-09-30 the archive page was rewritten after reading the S-1, with like-for-like figures; the route and permit lines now say a proposed route was denied, gas supply is delayed, the air permit is pending and the notice preserves Oracle's rights; the paraphrase is labeled and the filing quoted; outside sources went from 12 to 13; a note was added at the top. The English and Japanese editions were corrected the same day with their own notes. The email sent on 9/25 is the original.

Editions affected: 2026-09-25
Language editions affected: Chinese, English, Japanese
Corrected 2026-09-19Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The 9/18 issue's 'Sources & accounting' section split the 17 outside receipts used in the body into '3 from last night's 236 pieces' and 'the other 15'.

What was actually the case
Three plus fifteen is eighteen, not the seventeen printed in the opening and the footer; the correct split is 3 and 14. The total of 17 was right.

Which layer failed
Bad inference — same cause as the 9/17 issue: the split was only re-checked when the receipt total was recomputed, and all three totals were right from the start, so the check never ran.

What we did about it
On 2026-09-19 the archive page was changed from 15 to 14 with a note at the top; the English and Japanese editions carried the same 3-and-15 split and were corrected the same day with their own notes. The check now runs every day. The email sent on 9/18 is the original.

Editions affected: 2026-09-18
Language editions affected: Chinese, English, Japanese
Corrected 2026-09-19Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The Chinese edition of 9/17 split the 10 outside receipts used in the body into '4 from these 253 pieces' and 'the other 8', and listed eight names for the rest.

What was actually the case
Four plus eight is twelve, not the ten printed in the opening and the footer, and two of the eight names, Daily Business and Friends of the Earth Scotland, are not linked anywhere in the body. The correct split is 4 and 6.

Which layer failed
Bad inference — our exit check only re-verified the arithmetic in that sentence when the receipt total was recomputed; the totals were right from the start, so the check never ran.

What we did about it
On 2026-09-19 the archive page was changed from 8 to 6, the two names were removed and the next sentence fixed, with a note at the top; the total of 10 is unchanged. The English and Japanese editions had the correct 4-and-6 split. The check now runs every day. The email sent on 9/17 is the original.

Editions affected: 2026-09-17
Language editions affected: Chinese
Corrected 2026-09-17Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The Chinese edition of 9/14 said the load-bearing sources of 'five main-line items' traced back to six unrelated origins.

What was actually the case
The main line on the page has four items; the fifth was moved to the not-yet-verified section, and its source, OpenAI's official announcement, was one of the six origins and is not used by the four remaining items, so it is four items and five origins.

Which layer failed
Bad inference — when items were moved from the main line to the not-yet-verified section for length, only the items themselves moved; the inventory sentence built from the main-line list was not recomputed and kept its pre-move count.

What we did about it
On 2026-09-17 the archive page was changed to four items and five origins with a note at the top; 'six origins' in the English and Japanese editions was changed to five the same day with their own notes. The email sent on 9/14 is the original.

Editions affected: 2026-09-14
Language editions affected: Chinese, English, Japanese
Corrected 2026-09-17Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The Chinese edition of 9/13 said the load-bearing sources of 'four main-line items' spanned five kinds, the first being 'official model documentation and a pricing page'.

What was actually the case
The main line on the page has two items; the other two were moved to the not-yet-verified section, and the pricing-page source belonged only to a moved item.

Which layer failed
Bad inference — when items were moved from the main line to the not-yet-verified section for length, only the items themselves moved; the inventory sentence built from the main-line list was not recomputed and kept its pre-move count.

What we did about it
On 2026-09-17 the archive page was changed to 'two main-line items' and the pricing-page source removed, with a note at the top; the same list in the English edition and 'four items' in the Japanese edition were corrected the same day with their own notes. The email sent on 9/13 is the original.

Editions affected: 2026-09-13
Language editions affected: Chinese, English, Japanese
Corrected 2026-09-15Edition with the error: Japanese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The breakdown sentence in the Japanese edition's 'Sources & Inventory' section said the body used 16 receipts, of which 3 came from the previous night's intake and the other 15 were fetched separately; 3 plus 15 is 18, which does not match the 16 in the same sentence.

What was actually the case
The body actually cited 16 external receipts: 3 from the previous night's intake and 13 fetched separately or drawn from what we already hold.

Which layer failed
Bad data — the publish-time recount only syncs the phrasings registered in its position table; the inventory section's follow-on, anaphoric and arithmetic-breakdown sentences are not in that table, so only the opening line and the footer got synced and the rest stayed at their generation-time values.

What we did about it
The 15 in that sentence was corrected to 13 on the Japanese archive page, the list of sources that follows was left untouched, and a note was added at the top of the page; the same error in the Chinese edition is logged separately.

Editions affected: 2026-09-14
Language editions affected: Japanese
Corrected 2026-09-15Edition with the error: Japanese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The Japanese edition's 'Sources & Inventory' section put the count of external sources cited in the body at 17, while that edition's footer and opening line both read 15.

What was actually the case
The body actually cited 15 external sources; the 17 was a stale generation-time value.

Which layer failed
Bad data — the publish-time recount only syncs the phrasings registered in its position table; the inventory section's follow-on, anaphoric and arithmetic-breakdown sentences are not in that table, so only the opening line and the footer got synced and the rest stayed at their generation-time values.

What we did about it
The Japanese archive page's 17 was corrected to 15 and a note added at the top of the page; the same error in the Chinese edition is logged separately.

Editions affected: 2026-09-04
Language editions affected: Japanese
Corrected 2026-09-15Edition with the error: Japanese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The Japanese edition's 'Sources & Inventory' section put the count of external sources cited in the body at 28, while that edition's footer and opening line both read 26.

What was actually the case
The body actually cited 26 external sources; the 28 was a stale generation-time value.

Which layer failed
Bad data — the publish-time recount only syncs the phrasings registered in its position table; the inventory section's follow-on, anaphoric and arithmetic-breakdown sentences are not in that table, so only the opening line and the footer got synced and the rest stayed at their generation-time values.

What we did about it
The Japanese archive page's 28 was corrected to 26 and a note added at the top of the page; the same error in the Chinese edition is logged separately.

Editions affected: 2026-09-03
Language editions affected: Japanese
Corrected 2026-09-15Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
This edition's 'Sources & Inventory' section put the count of external sources cited in the body at 17, while the footer and the opening line both read 15.

What was actually the case
The body actually cited 15 external sources; the 17 in the inventory section was a stale generation-time value.

Which layer failed
Bad data — the publish-time recount only syncs the phrasings registered in its position table; the inventory section's follow-on, anaphoric and arithmetic-breakdown sentences are not in that table, so only the opening line and the footer got synced and the rest stayed at their generation-time values.

What we did about it
The archive page's 17 was corrected to 15 and a note added at the top of the page; the same error in the Japanese edition is logged separately.

Editions affected: 2026-09-04
Language editions affected: Chinese
Corrected 2026-09-15Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
This edition's 'Sources & Inventory' section put the count of external sources cited in the body at 28 and said it was the same figure printed in the footer, while the footer and the opening line both read 26.

What was actually the case
The body actually cited 26 external sources; the 28 in the inventory section was a stale generation-time value.

Which layer failed
Bad data — the publish-time recount only syncs the phrasings registered in its position table; the inventory section's follow-on, anaphoric and arithmetic-breakdown sentences are not in that table, so only the opening line and the footer got synced and the rest stayed at their generation-time values.

What we did about it
The archive page's 28 was corrected to 26 and a note added at the top of the page; the same error in the Japanese edition is logged separately.

Editions affected: 2026-09-03
Language editions affected: Chinese
Corrected 2026-09-15Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The 9/14 zh edition's inventory sentence carrying an arithmetic breakdown put the external-receipt count at 18, split as 3 from the previous night's intake and 15 retrieved separately, disagreeing with the 16 printed in the same issue's opening line and footer.

What was actually the case
The body actually cited 16 external receipts: 3 from that night's intake and 13 retrieved separately or pulled from the archive. The English and Japanese editions of the same day both printed 16.

Which layer failed
Bad data — the publish-time recount syncs only the phrasings registered in its position table; an arithmetic breakdown inside the same sentence ("only N ... the other M ...") is not in that table and stayed at its generation-time value.

What we did about it
On 2026-09-15 the two figures in that sentence on the archived page were corrected to 16 and 13, with a correction note added at the top of the page; the list of enumerated sources that follows was left untouched. The email sent on 9/14 is the original version and cannot be recalled; nothing else in the issue was changed.

Editions affected: 2026-09-14
Language editions affected: Chinese
Corrected 2026-09-15Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
The 9/11 zh edition's "Sources & Inventory" section gave the external-receipt count as 14 in three places, and two back-referencing sentences repeated 14, while the same issue's opening line and footer printed 15.

What was actually the case
The body actually cited 15 external receipts (counted as deduplicated public-version links minus our own domain). The three inventory-section figures and the two back-references were stale generation-time values; the opening line and footer were right. The English and Japanese editions of the same day both printed 15.

Which layer failed
Bad data — the publish-time recount recognized only the opening-line and footer phrasings at the time; the three inventory-section phrasings and back-references of the form "those N receipts" were not in its position table, so only two spots were synced and the rest stayed stale.

What we did about it
On 2026-09-15 the five figures on the archived page were corrected to 15 and a correction note was added at the top of the page. The three inventory phrasings were added to the recount's position table on 2026-09-14; the back-reference gap is tracked on an internal card. The email sent on 9/11 is the original version and cannot be recalled; nothing else in the issue was changed.

Editions affected: 2026-09-11
Language editions affected: Chinese
Corrected 2026-08-13Edition with the error: English editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The 8/10 English edition's opening sentence folded July's break-in at Hugging Face together with the internal message board disclosed at Black Hat, which made the security expert we quote further down read as if he were denying the break-in.

What was actually the case
The break-in itself was reported correctly — a zero-day plus stolen credentials is not access anyone granted (jointly disclosed 2026-07-21). What was wrong was fusing two separate incidents into one sentence, manufacturing a contradiction that was never in the reporting.

Which layer failed
Bad inference — two events compressed into a single sentence, and the sentence structure itself generated a conflict the sources did not contain.

What we did about it
On 2026-08-13 a dated correction note was added at the top of the English archive page (what it said, what changed, why); the opening sentence was rewritten to separate the two incidents, and a contradiction marker was appended to the security-community line. The email already sent is not recalled, and the note says so. We also judged this failure type machine-catchable in a narrow form, and routed the rule to a follow-up ticket.

Editions affected: 2026-08-10
Language editions affected: English
Corrected 2026-08-13Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
The 8/10 zh edition called Hugging Face "the injured party" without ever saying what injury it suffered; in the same passage a security-community line ("they never had to escape the sandbox") sits head-on against OpenAI's own account of the July incident, and we printed both without saying they collide.

What was actually the case
The injury refers to a separate July incident: OpenAI's model used a zero-day in a package-registry cache proxy to reach the open internet, then used stolen credentials to break into Hugging Face's production database and take benchmark answers; Hugging Face detected and contained it before OpenAI told them (jointly disclosed, 2026-07-21). The two accounts of "escape" really do conflict, and our own rule says conflicts get flagged.

Which layer failed
Bad inference — a background event was compressed into an unexplained label, and we broke our own rule that keeping a contradiction means marking it, not printing each side once.

What we did about it
On 2026-08-13 the website archive page was amended in place with a correction-and-addendum note, and the contradiction marker was added to that passage. The note states plainly that the email sent on 8/10 was the original version and cannot be recalled, and that nothing else in the edition was touched.

Editions affected: 2026-08-10
Whether other language editions were affected: not stated in this record.
Corrected 2026-08-11Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
Our 8/9 and 8/10 editions both wrote that the company "wiped and rebuilt" the repository, which reads as if OpenAI found the agent message board first and then decided to wipe it.

What was actually the case
Per a timeline correction posted on the evening of 2026-08-09 by an engineer self-identifying as being on OpenAI's side: when the repository hole was first patched the company did not know the message board existed; the board was cleared incidentally during the rebuild. The first action actually aimed at the board came later.

Which layer failed
Bad inference — our sentence collapsed "patched the hole" and "acted on the board" into one move, implying a decision sequence we had no evidence for.

What we did about it
The 2026-08-11 edition ran the correction as a main item and, in the same breath, listed three weaknesses in the correction itself (single source, posted from a personal account, and self-serving in direction). The two outside commentators who read the evidence the other way were left standing; we did not adjudicate between them.

Editions affected: 2026-08-09, 2026-08-10
Whether other language editions were affected: not stated in this record.
Corrected 2026-07-11Edition with the error: Chinese editionBad inference — the inputs were right, the reasoning or the wording was not

What we wrote
An earlier edition presented two papers together as "the two seed papers of the reasoning narrative".

What was actually the case
The second one is architecture-evaluation methodology and has nothing to do with reasoning. They should never have been paired.

Which layer failed
Bad inference — the two arrived in the same batch on adjacent topics and got grouped without each being checked back against its own abstract.

What we did about it
The 2026-07-11 edition split them in an explicit corrections paragraph, giving each paper its own takeaway.

Editions affected: 2026-07-09
Whether other language editions were affected: not stated in this record.
Corrected 2026-07-09Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
Our coverage of the three federal pressure moves against Anthropic never said how they ended in court.

What was actually the case
All three had been blocked by the courts back in March 2026. That judicial brake was there the whole time; we only caught it that day.

Which layer failed
Bad input by omission — our retrieval covered executive-branch sources but not the court rulings from the same period.

What we did about it
The 2026-07-09 edition carried the fix in the headline rather than burying it, and court rulings became a mandatory source for this topic.

Editions affected: 2026-07-08
Whether other language editions were affected: not stated in this record.
Corrected 2026-07-08Edition with the error: Chinese editionBad input — what came in was already wrong, and our checks missed it

What we wrote
Two earlier editions described the U.S. government's Fable ban as still in force, and reasoned about government leverage on that basis.

What was actually the case
The ban had already been lifted on 2026-06-30 — it ran 18 days and was replaced by standing pre-release review plus a dedicated team.

Which layer failed
Bad input. We had no rule requiring us to chase a contradiction to the end when reporting collides with observable reality — and the report in question was itself generated with Fable 5.

What we did about it
The whole 2026-07-08 edition ran as the correction: what was wrong, the root cause, and the three rules we added because of it. The same edition struck one entry from the prediction ledger: Lambert's "government and Anthropic will settle" was not an ex-ante call, since the lifting was already under way when he wrote on 7/1.

Editions affected: 2026-07-06, 2026-07-07
Whether other language editions were affected: not stated in this record.