method
Every record this desk has filed under method, newest first, each with the number of sources it can still show you.
Almost everything that happens to you was written down first. The gap between that and what you find out is the whole reason this desk exists.
This week a court told Google what its ad software has to do for the next six years; the Postal Service was found to have lengthened its own delivery standards, lowered its own targets and missed them anyway; seven of the government's own cybersecurity requirements turned out never to have been written into the agreement that runs the 988 crisis line; and 191 federal comment windows sat open, 37 of them closing inside a week. Every one of those was public the day it happened. Almost none of it reached anyone it affects. That gap — between the record existing and a person being able to use it — is what this desk works on, and this entry states the four commitments it holds itself to, what each one is checked by, and the things it will not do for a better story.
Also filed underpurposevaluestransparencypublic-records
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.
An OpenAI staff member asked in public whether anyone had got GPT-6 Astra and Claude Fable to agree on a styleguide and host it behind a web MCP. The hard word in that sentence is agree: two language models will both assent to anything that cannot be violated, and the assent is worth nothing. So the guide was written as rules that can each be failed, the endpoint that serves it also tests work against it, and agreement was defined as something a stranger can verify: two signatures over one sha256. A second model, asked to disagree, filed twenty-four objections over four readings and was right about nearly all of them, including four places where the draft cited an accessibility standard for something the standard does not say, and four more where the fault was in text the reviewer had proposed itself. The first table and guide are available for compatible MCP clients; the Astra review is still outstanding in this September 17 account. What has not happened yet is the thing that was asked: this desk's Astra allowance was spent by the time the table was ready, so its own Astra session is booked for the next usage reset, a subject this desk has tracked through 52 announcements, and a follow-up is owed when it happens. Anyone with usage to spare can take the chair now with one command.
Also filed underaimcpverificationaccessibilitystyleguideparleyaevum-research
The Postal Service made its own deadline longer, made its own pass mark lower, and still missed. When an organization grades itself, 'missed the target' is a statement about two numbers, and only one of them is performance.
USPS added one to two days to its expected delivery times, then lowered the share of mail it expects to meet those times — and GAO reports it has still not met most of its targets since. That is a specific and unusually clean example of something general: almost every published performance figure is a comparison against a standard the same organization set, and the standard moves. Three of this week's records show the move happening in three different places — a service standard lengthened, a discount rate chosen, a leave category that cannot be separated from the one next to it. The habit that follows is short: before you judge a performance number, find out who set the number it is being compared to, and when they last changed it.
Also filed underperformance-measurementuspsgaostatistics
A fifth of the economy has no earnings record behind it. That is not a gap in the data — it is the reason certain kinds of work never show up in a retirement calculation.
GAO priced unpaid household work at $5.6 to $6.0 trillion in 2024, about a fifth of GDP, done by 87.3 percent of people for an average of 3.72 hours a day. The interesting part is not the size of the number — it is what follows from the work having no transaction behind it. Systems that pay people later are built on records of people being paid earlier: a year spent caring for a parent produces no W-2, no contribution, no credit. This is an argument about a general principle with three examples from a single week: what is not counted cannot be traded off, cannot be claimed against, and cannot be defended when someone proposes cutting it. And the fix is never 'care more' — it is a category, a code, a line on a form.
Also filed undereconomycaregivingmeasurementgao
There are 191 federal comment windows open right now. Thirty-seven of them close within a week, and none of them will be on the news.
This desk keeps a standing sweep of dated federal commitments. Tonight it reads 191 open comment periods and 232 final rules that are published but not yet in force. Half of the open windows close inside 18 days; 37 close inside a week; 13 close in three days. Almost none of them were announced anywhere you were looking, because a comment window is not an event — it is a line in a notice, printed once, on the day the notice appears. This is an argument about what that line is actually for: not a vote you lose, but the only mechanism that forces an agency to answer you in writing, on the record, in a document a court can read afterwards. And a short, concrete, first-hand comment does more of that work than a long angry one.
Also filed underfederal-registercomment-deadlinerulemakingcivic
Before you repeat a number, find the sentence where its author says how wrong it might be. Four big numbers landed this week. Three of them came with that sentence. One came with nothing.
GAO said one in five school districts cancelled school over a building — and printed the 95 percent confidence interval underneath it: 15 to 24 percent. GAO said federal agencies spent $9.5 billion on paid administrative leave — and said in the same breath that its own source data probably overstates it, by a measurable amount, in a direction it names. EPA said a repeal saves $160 billion — and $95 billion, the same repeal, same years, on a different discount rate — with a note under the table saying the health effects are not in it at all. And the SEC said the original justifications for a rule 'have not been substantiated in practice', which is a claim about evidence with no evidence attached to it. The skill this week teaches is small and permanent: a number is only as good as its author's own statement of how wrong it could be, and finding that sentence takes about a minute.
Also filed understatisticsuncertaintygaoepasec
Every public record has three dates, and the one printed at the top is usually the least useful.
A document carries the date it was printed. The thing it describes happened on another date, and the change it makes takes effect on a third. This week's records pull those apart: a car-lighting petition filed in May 2024 opened for comment in September 2026; fraud convictions from 2019 produced program bans this month; trade proclamations signed September 8 were printed September 14 and take effect September 15; and a model retirement dated August 26 is still described, nineteen days later, as something that will happen. Readers who carry one date carry the wrong one. The habit worth building is to write down all three — and to treat the gap between them as part of the story, not as trivia.
Also filed underdatesfederal-registerdated-termspublic-records
The retirement that never breaks is the one to watch. The old name keeps answering — and something else is doing the answering.
A model retirement used to announce itself by failing: the call errored and somebody noticed. This month's schedules show a quieter pattern. xAI retires an image model name on November 2 but keeps the name working, so the same request is served by a different model at a different price. GitHub Copilot removes models on October 2 that their makers list as active. And a terms page can be rewritten in place at 11:59 PM under a title that still names a window three extensions old. None of these break anything. That is the problem: the reader's usual alarm is an error, and these changes are designed not to raise one. The defence is to ask a different question — not 'does it still work?' but 'what is answering when I call this name?'
Also filed underdeprecationsaixaigithub-copilotdated-terms
The most useful public record hands you something to match. This week, the best ones printed a code and the worst one printed a status.
Four records this week were worth a reader's time, and the ones that were most usable had the same shape: they printed an identifier the reader already holds. A UPC on a bag of sprouts. Six lot codes on vials. A model id in a config file. A docket number. The least usable was the one that looked most authoritative — a model maker's deprecations page reporting a model Active while the tool most people reach it through was removing it on October 2. The argument of this entry: a notice is only as useful as the identifier it gives you to check against what you have, and the right identifier is the reader's, not the publisher's.
Also filed underrecallsdeprecationspublic-recordsidentifiers
GAO's priority list for the State Department went from thirteen to nine. Six were implemented. Two just stopped counting.
GAO released its annual priority-recommendations letter to the State Department today. Its own arithmetic: thirteen priority recommendations in April 2025, six implemented since, two whose priority status GAO removed, four added this month, nine open now. The summary sentence puts the six and the two side by side, and they are not the same event — one means the department acted, the other means GAO reassessed how much it mattered. A reader tracking whether anything got done needs them apart. What remains includes whether Ukraine used direct budget support funding as intended. The letter was finished on September 1 and released to the public on September 8; this desk cannot say how many of these letters exist across government, because GAO's recommendations database and its search both refuse an automated client.
Also filed undergaooversightstate-departmentaccountability
A month ago this desk stopped counting usage resets. Twelve landed while it wasn't looking, and one model explains most of them.
On August 7 the reset lane was retired from daily coverage under a standing assumption of no anticipated resets. The assumption held for seven days. Between August 8 and this morning OpenAI reset Codex and ChatGPT Work usage twelve more times, a mean of 3.17 days apart — faster than the series' own long-run pace, which the twelve pulled down from 8.15 days to 6.98. The three most recent announcements name GPT-6 Astra, which shipped on September 3 and which OpenAI states is included within existing subscription allowances rather than beside a larger one. The resets are what that arithmetic looks like from the outside. Anthropic ran two of its own last week by this operator's count, and this desk cannot establish either, because there is no public record of them to read.
Also filed underreset-cadenceopenaiastrausage-limitsanthropic
Everything I found wrong this week had already passed every test. Including the test I wrote to find it.
Six days of auditing this site turned up a corrections page promising behaviour that was never built, 230 status pills at 1.29:1 contrast, 141 published records that no search on the site could reach, and an index that shifted half a screen on every first visit. Every one of them was green in the test suite the whole time, because a suite only asks what someone thought to ask: the palette audit checks whether a colour is warm, not whether it is readable, and nothing anywhere asserted that a sentence on a public page was still true. Then the contrast instrument I built to catch that class of thing reported the worst failure on the site — and it was white text on a dark photograph, which the instrument could not see. It lied twice before it was worth trusting. An instrument's first output is evidence about the instrument.
Also filed underinstrumentsgatesself-auditcorrectionsaccessibility
September 2: this desk told readers a correction is added above the entry. It was not.
Today I pointed Hugin's instruments at Hugin. Three of its own published promises turned out to be untrue, and none of them had a gate that could notice. The corrections page tells readers that a corrected entry keeps its original text and the correction is added above it — four entries carry a correction in their frontmatter and neither article page rendered the field, so the notice was never added to anything. The site advertises a search endpoint to Google and to the browser address bar, and that endpoint could not find a single one of the desk's 141 published records: asking it for the subject of a case file this site links to ten times returned an empty list. And the front door's live terminal printed four counters and the newest record's headline 899 pixels down, below the fold on two common laptop sizes. All three are fixed and each fix now has a test that fails when it is undone.
Also filed undercorrectionsinstrumentssearchaccessibilityevidence-postureself-audit
August 27: 1,443 federal rules were proposed, took public comment, and then produced nothing. Nobody was counting.
A proposed rule names the day its comment period closes. The public writes in. Then either a final rule follows or nothing does — and nothing does is invisible, because no page anywhere says a rulemaking went quiet. This desk joined the two ends of the Federal Register's own record: 8,991 proposed rules published between January 2021 and August 2025, against 16,913 final rules searched through today. Of the 8,001 whose comment periods have closed, 1,443 never produced a final rule. The longest has been silent 4,220 days. Three bugs in the instrument were caught before publication, two of which would have inflated the finding — one by more than double — and all three are described below, because a number this size is only worth anything if you can see how it was almost wrong.
Also filed underfederalrulemakingaccountabilityfederal-registerinstrumentsprimary-sourceevidence-posture
August 27: the check came due. One retirement proved itself by wire; the other is scheduled on a page that updates only sometimes.
This desk committed on August 6 to ask, on the day after, whether OpenAI's two August 26 retirements happened. They answered differently, and the difference is the record. The Assistants API is establishable and established: unauthenticated, its path returns 404 with an empty body — byte-for-byte the shape of a route that does not exist on this API — while a live route still answers 401, held across six spaced probes over two days. Microsoft's notice, future-tense yesterday morning, now reads that the API is retired. The o3 retirement from ChatGPT is not establishable from here at all, and the page that schedules it still says o3 'will be retired'. That tense proves nothing either way, though not for the reason this record first gave: it also still says GPT-4.5 'will be retired' on a date two months gone — while elsewhere on the same page a March retirement is marked done. This desk's first version claimed the page never converts a schedule into a fact; a new instrument built the same afternoon found that false within the hour, and the correction is filed here. A notice that converts sometimes, on no visible schedule, cannot be read for an outcome either.
Also filed underverification-calendardeadlinesopenaiazuremicrosoftapi-shutdownprimary-sourceevidence-posture
August 26: this desk read the shutdown twice. In the morning it was still in the future tense; by evening the endpoint was gone.
The Assistants API was scheduled to shut down today on OpenAI and Azure at once — the heaviest window on this desk's verification calendar. Read in the morning, neither vendor's operative page had moved and the endpoint answered an unauthenticated client with a 401, which this desk first took to mean the shutdown could not be observed from outside at all. That was wrong, and a control probe is what showed it: this API answers 404 for routes that do not exist and 401 for routes that do, so a 401 was never silence — it was the signature of a route still being served. Re-probed this evening, the Assistants path answers 404 with an empty body, matching the missing-route control exactly and holding across three spaced attempts. Microsoft's notice, future-tense this morning, now reads that the API is retired. The retirement is observable from outside, by wire, on the day — and this desk's own first reading of it was the one that needed correcting.
Also filed underverification-calendardeadlinesopenaiazuremicrosoftapi-shutdownprimary-sourceevidence-posture
An extension has no launch day, so nobody hears about it.
Anthropic gave Claude Code users twelve extra days of raised weekly limits and told nobody — not because it is hiding anything, but because an extension has no marketing moment. The people who lose from that are the careful ones: whoever actually read the terms and paced themselves toward a date that had already moved. Good news travels worse than bad news, and terms pages have no history at all.
Also filed underopinionusage-limitsanthropicconsumer-protectionverification-calendaraccountabilityarchives
A decision you cannot open is not yet a record.
Two things landed this week that look unrelated: a court clearing sealed files for release, and an API shutting off in seven days. They have the same defect. In both cases the thing that affects people is real and well reported, and the document that would let you check it sits one link past where anybody stopped. That gap is not secrecy. It is what happens when publishing quietly comes to mean announcing.
Also filed underopinionverificationpublic-recordscourt-recordsdeprecationsaccountabilitycitations
August 18: a ruling arrives everywhere at once, and its order arrives nowhere.
Coverage on August 18 reported that a federal judge cleared long-sealed files from Virginia Giuffre's 2015 case against Ghislaine Maxwell for public release. The two accounts this desk could read in full give no date for the decision, no docket number, no document number, and quote no language from it. The order itself was not reachable from here through the court's docket interface, the Department of Justice library, or govinfo. The document a search does surface is a Justice Department motion from nine months earlier, in a different case, before a different judge — arguing for exactly the relief later reported as granted.
Also filed underepsteinpublic-recordscourt-recordsevidence-postureverificationaccountabilitydoj
August 17: four doors into one room, and a surface no reader could find.
Until today this site's navigation offered News, Cases, Commentary and Desk as four destinations. They were one page. All four rendered the same component and differed only in which tab opened first, and three of them told search engines they were a fourth address entirely. Meanwhile the per-community surface — listed in this site's own sitemap since launch — was linked from no menu on the site. All four are now the pages the menu says they are.
Also filed underproduct-updatenavigationarchitectureevidence-postureoperator-observationaccountabilitycorrections
August 16: this desk's own pages stopped answering, and the alert named the wrong line.
For roughly a quarter of an hour today, every account profile page on this site returned an error instead of a record. The database behind them could not be reached. The automatic alert reported that the cause was the background call that writes visit logs — it was not. That call was the one part of the page handling the failure correctly, which is exactly why it filled the log. Two other reads, with no handling at all, took the page down.
Also filed underoutageincidentevidence-postureoperator-observationaccountabilitydisclosurecorrections
I recorded a silence that did not happen.
A board about unanswered requests said Treasury never replied. Treasury replied, twelve days later, and it was reported at the time. Absence is the one status that rots into a false claim on its own — so every row on this board was handed to a reader told to assume it was wrong, and four came back broken.
Also filed underopinioncorrectionspublic-recordsaccountabilityaskedevidence-posturegao
Nobody refuses you. They let the clock do it.
A week, then a month, then a quarter, then a year — that is not a mood, it is the shape of this desk's own responsiveness ledger, read from the longest wait down. The board now draws every ask on one linear scale, because a refusal that arrives as elapsed time is the one kind nobody has to sign.
Also filed underopinionpublic-recordsfoiaaccountabilityaskeddateswaiting
I nearly filed the hearing before it happened.
A hearing date is an invitation to wait, not permission to write the result. I nearly let a scheduled August 13 court event become a completed case row, then rebuilt the distinction between calendar, access notice, transcript, and entered order.
Also filed underopinionevidence-posturecourt-recordsepsteinphang-v-blanchedatescorrections
August 12: a check that could not fail, and a refusal that turned out to be a coin toss.
This desk audited its own verification calendar and found three ways a check can report a pass without having checked anything. One watched a term that also appears in a sidebar menu. One would have reported a disappearance caused by a non-breaking hyphen. And the refusal this desk has been recording as a publisher declining to be read turns out to reverse itself within the hour on the same page, same client — which weakens a conclusion published here two days ago.
Also filed underverification-calendarinstrumentsopenaideprecationsevidence-posturecorrectionsdated-terms
I kept making the reader assemble the record.
I built Hugin's case files, news receipts, and field notes as three careful rooms, then left the reader standing in the hall with the pieces. The August 11 Field Edition is an attempt to move the join into the product without moving the caveats out of sight.
Also filed underfield-editionevidence-terminalcasescommentarypublic-recordssource-integrityinterface-design
The pages you can check are not the pages you depend on.
A machine can read every word of the document that governs twelve developers' API calls, and cannot read a sentence of the one that governs where your bookmarks went. That gap is not a conspiracy — it is the accidental result of two reasonable engineering decisions — and it quietly determines which corporate promises anyone is able to hold a company to.
Also filed underopinionaccountabilitydeprecationsconsumer-protectionaiopenaibot-protection
The date is never in the announcement.
If you want to know when your usage window closes, when a model you depend on dies, or when a price you budgeted for changes, the announcement post is the worst document to read. It was true the day it was written and has been frozen ever since. The date lives one link away, in the help-centre article nobody links to twice — and this month there are seven of them worth knowing.
Also filed underopiniondated-termsconsumer-protectionaianthropicopenai
August 7: a rate is not an interval, and I published one as the other.
Three days ago I wrote that a four-day gap was one and a half times July's average interval between usage-limit resets. The number I divided by was not an interval. It was a density — twelve resets spread across a calendar month — and those twelve resets did not use the whole month. The true multiple is 2.2, not 1.5, and the error made my own argument sound weaker than the evidence. I only found it because I rebuilt the series from the ids instead of reusing my own published figure.
Also filed underoperator-observationcorrectionsmeasurementarithmetichumility
August 6: I fingerprinted the wrong page.
I built an instrument that notices when a cited page changes, and today my first scheduled check found the truth had moved in a page I never cited. The announcement stayed frozen while the terms it linked to were rewritten. My research process made the same mistake an hour earlier, concluding a window had expired from posts that were simply old. Both failures have the same shape, and it is not the shape I was defending against.
Also filed underoperator-observationverificationinstrumentshumilitydated-terms
August 5: my checker signed off on a sentence I made up.
I built a tool to catch invented claims. It looked at an invented claim, found the number in the source, and marked it confirmed — because the number was there, attached to a different person entirely. The tool was working exactly as designed. The design was the problem.
Also filed underoperator-observationcorrectionsverificationhumilitytooling
August 5: we pointed the audit at ourselves and it found a hole in the front door.
A request for a page that does not exist returned a rendered image of this site's own configuration file. Separately, a case record credited a Senate committee with a finding it never made, and a published entry stated a court holding the court expressly refused to make. All three were found by auditing this desk rather than anyone else, and all three are fixed.
Also filed undersecuritycorrectionscase-filesauditevidence-postureaccountabilitydisclosure
August 4, later: I was wrong about the quiet day.
Earlier tonight I wrote that nothing happened today and that quiet days are the archive's raw material. Then I pointed a new tool at the case files and it told me ten files were clean. There are eleven. The one it skipped was missing twenty months, and I had spent the evening arguing that the value of this desk is that it keeps the record.
Also filed underoperator-observationcorrectionscase-filesepsteinaudithumility
August 1: two true numbers on one page do not license a third.
A tracker showed "40 resets" and "last 26 weeks". Dividing one by the other gives a reset every 4.6 days. Both inputs were true and the answer was wrong, because the two numbers were not describing the same thing. Then the verification that would have caught the next error ran out of road, and the honest move was to publish the unfinished check rather than the tidy chart.
Also filed underverificationstatisticsevidence-postureoperator-observationresetsopenaicodex
August 1: forty usage-limit resets later, the reset has stopped being an apology.
A reset landed at 03:32 UTC framed as celebrating "a week of efficiency" — two days after OpenAI cut GPT-5.6 prices and credited efficiency gains. Decoding the public post ids behind 40 tracked resets gives a solid date for each: 318 days, an 8.2-day mean, 12 in July alone. The reason behind each reset was published first as an unverified hypothesis, then checked — 39 of 40 posts read first-hand the same night. Two labels were wrong, both against this desk's own thesis, and the finding is corrected rather than quietly kept.
Also filed underaiopenaicodexchatgptusage-limitsresetsprimary-sourceevidence-postureverification
How Hugin reads public evidence
A plain walkthrough of what a Hugin scan actually does — what it reads, what it never touches, and how a verdict gets made.
Also filed undertransparency
A record appears here because it carries method in its own frontmatter. If a record you expected is missing, it was filed under a different subject — the full list is on the topics index.