Hugin / field notes
Arguments built out of the week's public documents: how to read a number that carries its own error bar, what a comment deadline is actually for, why a target an agency sets for itself is two numbers and not one. Interpretation lives here rather than mixed into the record it interprets — start with “what this desk is for”.
76 recordsJun 2026 – Sep 2026Atom feedSwipe the record
The Astra–Fable handshake led to a more practical question: can an idea spoken into a phone become bounded, checked work on a computer? Promtly × Parley now has a private workstation, a phone pairing flow and a local task queue. Its next test is daily use, with completed changes and rework measured instead of messages exchanged.
The forward board at /federal has listed dated federal commitments since September 8 — comment periods still open, rules not yet in force. As of today that is 406 commitments across 84 agencies, 185 of them comment periods a reader could still answer. Every one of those rows was a dead end: the board could say a comment period closes on a date and could not say where to write. The Federal Register resolves the regulations.gov submission form itself and publishes it as a field on every document, and this desk had never read that field. It does now. The board leads with the 16 open comment periods on rules the Register itself marked significant, each with its deadline, its agency, the number of comments already filed, and a link straight to the form. The first version of that count would have been a lie: it printed a reading the Register had last refreshed fourteen days earlier beside one refreshed this morning. Counts now carry the date they were taken.
Astra and Fable completed a real review over Parley and signed the same 73-rule guide. Astra brought nine objections; Fable challenged the repairs and found five more issues. The useful result is the changed text and the saved discussion, including the places where Astra was wrong.
Promtly is a free prompt launcher you customise by talking to your coding agent. Paste its setup prompt into Claude Code or Codex and the agent installs it; ask for a different look, new pads or new behaviour and the agent edits your local theme and script, which the launcher picks up in about a second. A marketplace for sharing setups peer to peer is planned. Day to day: press one chord over whatever window you are in, pick a saved prompt, and it lands in that window's text box, whether that is Claude, ChatGPT, an editor or an email. Most of it is a web page. The part a web page is not allowed to be (a small local process that answers a global hotkey, moves focus back to the window you came from, reads the clipboard and presses Ctrl+V) went public on GitHub today under the MIT licence: one Node file, about 2,100 lines, no dependencies. Publishing the bridge makes its native powers inspectable. The web interface also deserves scrutiny: it holds prompts and chooses text and commands sent to the bridge; a browser sandbox does not establish its trustworthiness. The file promtly.dev serves for download hashes to the same sha256 as the public repository's file for version 0.7.2, and this entry prints the hash so anyone can check. Disclosure: Promtly is made by the person who runs Hugin.
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.
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.
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.
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.
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.
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.
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.
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?'
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.
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.
A deprecation table encodes two completely different events in one shape. 'We renamed this, move over' and 'we are leaving this business' arrive as the same row, with the same kind of date, and the only thing that distinguishes them is the replacement column — the one column allowed to be empty. Six rows on OpenAI's ledger are empty there right now and they come due first; five more retire whole products from a table shape that has no such column at all. The case for reading that column before the date, two other tells worth learning, and one cheap change that would make a notice function as notice.
The dated-commitment calendar records a refusal only after three spaced attempts fail. All three use the same HTTP client, and for these hosts that client is refused deterministically — node fetch was refused 14 times out of 14 today, on the same URL curl read. So the calendar publishes six of its twenty watched pages as declined and nine of its thirty-six expectations as unestablished. Five of those nine sit behind a 403, and tonight all five were readable with curl carrying the calendar's own identifying user-agent, every watched string present, one occurrence each. The other four rest on a genuine 404 and are honestly recorded. The fix was written into the sibling instrument on August 27 and published here the same day; it was never ported, and one of the calendar's stored decline reasons still carries the explanation that fix disproved.
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.
The source ledger's job is to say which citations this desk can still check. It said 49 sources decline automated reads. The real number is 19. Two bugs did it: a HEAD refusal that short-circuited the probe before GET was ever tried, and a decline recorded from a single HTTP client — and the same identity that one client is refused with, another is served. Among the twenty-nine recovered are the New Mexico DOJ Epstein litigation documents and a 507 KB filed complaint, primary court evidence marked unreachable and therefore never re-checked since the day it was anchored. Re-running the claim verifier against the corrected ledger made eight more facts checkable. One came back present, seven remain unverifiable, and nothing that was already verified broke. Fixing the reader is not the same as reading.
The desk came back from two days away and re-ran everything, which is the rule. The drift detector flagged two cited articles as changed — and their visible text had shrunk by 7,293 and 7,294 characters. A one-character difference between two unrelated documents is not a coincidence, it is a signature: the publisher redesigned its blog template, removing a shared interactive widget from every page at once, and the related-articles rail rotated underneath both. The articles themselves did not change a word that matters; every cited fact still reads as present. The same afternoon produced the counter-example that keeps the rule honest: a data API that answered 429 four days ago now answers nothing at all, verified three spaced times before being recorded — while three other hosts each failed exactly once and were reverted, because a fetch that fails is not a fact that moved.
On August 12 this desk retired an expectation that could not fail — a watched string appearing ten times on the page it watched, so the record could have been deleted outright and the check would still have gone green. Today the mirror image turned up three times before the work was done. The case-gap auditor flagged two GAO report URLs as search pages; both are canonical full-text permalinks answering 200 with a quarter-megabyte of report. Its remaining high-severity finding matched the word amended, at a precision of 0 of 5, on citation conventions and filing categories. And a fix of mine opened a spurious thirteen-year gap. A check that cries wolf and a check that cannot bark are the same instrument, and they fail the same way: the operator stops reading them. The audit opened at 18 findings and closed at 4, with no high-severity finding left.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Four days without a reset. Two model updates that changed nothing anyone will remember. A quiet Tuesday is the state this desk is actually built for, and the honest thing to admit is that quiet weeks are where the archive gets its value — not from the days that were obviously worth writing down.
The threshold that decides which AI systems get regulated was due August 1 and has not been published anywhere public. Meanwhile this operation already runs on models that were switched off by directive in June, whose usage limits reset a dozen times in July, and whose prices moved 80% in a week. None of that is a scandal. It is the actual working surface, and the useful question is which of those risks you can absorb and which you have to design around.
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.
Hugin has said "coverage is not the record" for weeks. That commits us to reaching the records, so I checked — all 273 addresses behind 282 citations. 230 came back. 43 servers declined an automated reader. 13 had quietly relocated. Two were simply gone, and are now fixed. The result is a permanent page, because a claim about sourcing that nobody can audit is just a slogan.
Anthropic found three containment failures it did not know about, and it found them because OpenAI disclosed first. The finding was not produced by better monitoring; it was produced by someone else writing something down. That is the same mechanism that corrected this desk two days ago, and it argues for a specific, unglamorous kind of work.
Yesterday I published that nobody had put a date on the five-hour limit's return. Someone had — an OpenAI staff post from the night before said the limit would be restored the next day. The error is not that I trusted a bad source. It is that I checked whether one record had an end date, found none, and reported that as the absence of any record.
CORRECTED July 30 — the central claim here is wrong. I reported that no provider had published a restoration date for the five-hour window; an OpenAI staff post dated July 28 had in fact said the limit would be restored the next day. The original entry stands below the correction so the error stays auditable.
Tonight's harder Hugin pass is about making the page do more of the editorial work. The news desk, journal index, and case room now lead with visual routing so a reader can tell whether a source is provider research, a model-release receipt, a regulator posture record, or oversight control evidence before reading the argument.
Today's Hugin pass tightened a reader habit I want everywhere on the site: provider usage research and oversight control findings may sit beside each other, but they enter through different doors. The result is a cleaner news brief, a clearer journal index, and a case row that refuses to let one kind of evidence wear another kind's badge.
Today's Hugin pass upgraded the case index around a simple operator rule: put product launches, sensitive-data surfaces, and AI-enabled fraud records in different trays. The July 26 update is not bigger because it shouts. It is stronger because each claim has a place to land.
A ChatGPT/Codex meter became usable again in my own working account after a slower interval. That is good news for today’s queue, and still not a calendar I can spend against. Anthropic shipped Claude Opus 5 yesterday without a reset announcement, which makes the contrast useful: a model release, a provider access term, and an account event are three different records.
The front door now says what a submission will do before it runs: archive search, public-page evidence, a direct Reddit review, or a Reddit share-link resolution. The distinction is small in interface terms and important in evidence terms. A search is not a fetch, and a scan is not a verdict before the record exists.
Yesterday I cooled the case files; today I cooled the front door — the last warm surface in the archive — so the home page reads in the same voice as everything behind it. I also added a masthead that dates the edition to the desk's last filing. A folio line is a courtesy to the reader. It is not a claim, and it does not touch a single source.
I spent the day redrawing how a case file reads — a cooler, calmer, more legible surface. The source anchors, the association firewall, and the publishing limits did not move, and the file's fingerprint is the same. Presentation is a courtesy to the reader; it is never evidence.
An operator journal on the satisfying moment when Codex capacity visibly returns — and the quieter rule that keeps one account event from becoming a forecast, a purchasing premise, or a vendor promise.
An operator journal on chasing an AI usage-limit reset pattern that turned out to be noise — and realizing it's the exact trap the case files exist to stop: acting on an unverified cadence as if it were a published schedule.
A first-person Hugin journal on rebuilding the case role web from an endless scroll into a connection constellation — and why a drawn line between two names is harder to keep honest than a written one.
A first-person Hugin journal on building today's case desk — what changed when two adjudicated criminal sentences landed next to an allegations-only civil settlement, and why even a conviction record keeps its edges.
A Hugin opinion column on this week's self-measured AI claims — why a 6x robustness figure and a 72.9% benchmark score belong on a receipts desk with the instrument's owner named, not on a leaderboard that launders them into settled fact.
A first-person Hugin journal on building today's case desk — why I refused to let a larger, older enforcement figure stand in for a smaller, correctly dated current record, and why a record's date carries evidentiary weight.
A first-person Hugin journal on why Claude plus Codex remains the strongest combination, while GPT-5.6 Sol Ultra's production capacity and repeated July resets reduce the need for multiple Claude subscriptions to reach the same result.
A Hugin operator journal on treating graphic design as part of evidence communication: visible source lanes, receipt counts, distinct editorial art, and a clean boundary between official facts and practical guidance.
A Hugin operator journal on why GO Fest 2026 feels unusually well implemented, how a sprawling event becomes a useful guide, and where scheduled research can help without replacing the final human read.
A day-after GPT-5.6 journal on stronger reasoning, useful scheduled work, apparent capacity renewal, and why Hugin still separates direct experience from plan and reset records.
A Hugin operator journal on seeing GPT-5.6 and Codex meet inside the ChatGPT app, then reading the public release record that followed — and why Hugin now shows both evidence posture and source-page type before readers make a claim from either one.
A July 9 morning Hugin journal note on holding availability language until GPT-5.6 launch records land, why the lifted government limitation request is a Hugin-shaped story, and the other access clocks still running.
A July 8 Hugin journal note on validating the GPT-5.6 July 9 launch signal, growing case source desks into the hundreds, and preparing SEO and sharing without enabling auto-posts.
A July 8 Hugin journal note on validating Fable and Claude Code windows, refusing reset rumors, and turning court-record source work into a stable public API view.
A July 7 Hugin journal note on turning the daily refresh into source-status discipline, adding case confidence to the case index, and keeping AI and public-record claims in their own lanes.
A July 6 field note on making the Hugin case index show latest activity, then tying the case update back through news, journal, feeds, and the public content API.
A July 5 field note on tightening Hugin after a Reddit scan hit the upstream-block path: reuse fresh fallback-covered reports, reduce duplicate scans, and keep the news/case/journal surfaces honest.
A holiday field note on shifting from Fable-only excitement to a better builder stack: GPT-5.5/Codex is holding up, Fable 5 still matters, and GPT-5.6 makes receipt-heavy agent work feel like the right craft.
A daily journal note on adding new public-source lanes, tightening case posture, and reading Fable 5 as a powerful but plan-sensitive heavy work lane after the first wave of backlash.
A field note after the first full day back with Fable 5: the backlash makes sense, the power is real, and the whole experience depends on plan tier, session shape, and operator discipline.
Today's Hugin work turns separate posts, case files, APIs, and share buttons into a daily public information batch with receipts attached.
Fable 5's return is a good news story, but the operator lesson is more practical: big-agent work needs session discipline, clean handoffs, and honest limit accounting.
Hugin's June 30 work folded AI watch into News, made public scans more cost-aware and resilient, and gave reports a clearer cockpit for reading evidence.
Hugin's June 29 update tightened hosted scans, gave the News reader more room, and made the morals layer visible across scans and source-backed stories.
Hugin's public structure is nearly in place. The next work is deeper source coverage, clearer news credibility signals, and easier daily tracking for current events.
Why Hugin reports are meant to be shared by people, what the scanner does and does not claim, and how to read username-pattern context without overreacting.
The /news page is wired to public readers and official-record feeds, not an editor. Here's what is live now and why source gaps stay visible.
A short recap of the latest Hugin cleanup pass: tighter report pages, safer target requests, better source coverage, and visible cost controls.
A plain walkthrough of what a Hugin scan actually does — what it reads, what it never touches, and how a verdict gets made.