What the Disclosure Regime Records

Verification chain

Four things are pinned, and each can be checked independently of the others.

The assessment PDF. SHA-256: 7b211e34a0ae752d45dfa6c4183676b7de58e9a20da39aebda246ffe3a422ead

Download the file and run shasum -a 256 over it. If the value differs, the file is not the one struck here.

The timestamp. The receipt AOS-ADG-2026-001-v1.2.pdf.ots is an OpenTimestamps proof over exactly those bytes, submitted to four calendars on 23 August 2026 and pending confirmation in the Bitcoin blockchain at publication. It has since been anchored: the receipt served here carries two Bitcoin attestations. It establishes that the file existed no later than the moment it was stamped. It establishes nothing about whether the file is correct.

Receipt SHA-256: 2a74a2efbe0e8bb08b823aacfdc0c406782dd6565e4f8bd95065b8264cda2a83

The assessed source. The assessment rests on 122 pinned artefacts: 119 retrieved and pinned by this case, and 3 inherited under convention 7 from other cases in this corpus, each re-verified against its recorded hash before this record was written. All 122 are pinned by hash and 89 of 122 independently archived. Every quotation was verified against its pinned artefact. The one PDF source was read at page-image resolution rather than from its text layer, which is how three transcription defects were found; the HTML sources were checked by slicing the markup covering each quoted passage, so that emphasis and list structure are reproduced as the source sets them. Vocabulary counts state the artefact they were taken from and the control word run on the same extraction.

Re-fetch, measured 2026-09-06. Every own raw-body row of this register, 116 rows, was fetched twice that day and the two hashes compared. 18 returned different bytes on each fetch, D-TS-5, D-TS-13, D-TS-26, D-TS-51, D-TS-56, D-TS-64, D-TS-65, D-TS-66, D-TS-70, D-TS-77, D-TS-85, D-TS-86, D-TS-87, D-TS-88, D-TS-99, D-TS-104, D-TS-110, D-TS-111, and are marked in the table: their hashes stand only against the copies this practice holds and cannot be confirmed by fetching. The rest returned the same bytes twice on that day, which holds until the publisher changes the page. The measurement is in STABILITY-2026-09-06.tsv beside the register; the register itself is unchanged and pinned, does not point to that file, and a reader of the register alone does not see this (convention 50).

89 of 122 pinned sources have a third-party capture, and 84 of them establish that the content pinned above is what the source served - 14 whose archived bytes hash identically to the pinned copy, and 70 whose bytes differ, as a dynamically served page's will, but whose extracted text is identical to it character for character. The other 5 differ in text and are itemised below. 33 of the 122 have no capture. For 23 of those the Internet Archive could not reach the publisher's own server - 21 responses of HTTP 520 and 2 of HTTP 523, on two attempts each across two days. That is a property of those agencies' hosting, not of the archive and not of this practice, and retrying does not change it. For the remaining 3, a memento exists and serves nothing. Those three are not linked here, and neither is one further memento inside the 23 that serves only a 290-byte stub, because a link a reader follows to an empty page invites the conclusion that the source said nothing. The remaining 7 were pinned at v1.1, after the capture run of 20 August 2026, and were never submitted to it. They are United States legislative material and two provider statements; the legislative material is pinned from the publisher's own point-in-time interface, which serves the text as in force on a stated date and is itself the archival record. The 5 captures whose text differs: D-TS-5: a share-widget label, "URL" against "LinkedIn". D-TS-29: the page renders the current date, and the capture was made the day after the pin. D-TS-34: the page renders the current date, and the capture was made the day after the pin. D-TS-86: the page itself changed between 20 and 21 August 2026, from a 2025–26 annual workplan to a 2026–27 one. D-TS-105: the newsletter form’s field labels. 84 of the 122 were already archived independently before this practice retrieved anything, the earliest in October 2025. Such a capture is stronger evidence than one this practice requested, and on this register it is the majority. Two of the sources render the current date, so their archived text can never compare identical to a copy pinned on another day. That is a property of the comparison, not of the source, and the count above is stated in the knowledge of it.

The build. The PDF is produced deterministically from tracked source by build.sh, with SOURCE_DATE_EPOCH pinned to midnight UTC of 20 August 2026, the publication day, which does not move at any revision. Repeat builds on one machine and one toolchain are byte-identical; independent third-party reproducibility is not established and is not claimed. Verify by SHA-256, not by rebuild.

Version history

Three versions. What each revision changed is in the document’s own version note and, where a revision corrected anything, under Corrections below. A revision does not replace its predecessor: v1.1’s bytes are unchanged, its hash still resolves, and the file remains downloadable at its own address - all 3 are listed below with their hashes and their own addresses. Every version was assessed on 20 August 2026; v1.0 on 20 August 2026, v1.1 on 23 August 2026, and v1.2 on 2 September 2026. Each revision keeps the first edition’s SOURCE_DATE_EPOCH, because a revision is not a new reading - it is the same reading, corrected.

The cover carries 2 September 2026, the date this revision was made. The assessment was first made on 20 August 2026. A revision’s cover carries the revision date while the epoch inside the PDF stays at the first reading’s date, and the version table records each version’s dates.

Corrections

One correction, at v1.1, published 23 August 2026. It narrows a stated limit. No finding moves, no figure changes and no quotation changes.

What v1.0 said. Stated limit 1 read: “The 12 June 2026 directive is not established. Its issuing body, legal authority, instrument number, text, scope, treatment of allied nationals and lifting are established by no record this practice could obtain.”

Why it was wrong. Six of those seven items were establishable on 20 August 2026, the date the limit names. AI model weights are controlled under ECCN 4E091, added by a published rule at 90 FR 4544 of 15 January 2025; section 742.6(a)(13) of the Export Administration Regulations requires a licence for them worldwide and disapplies that requirement to deemed exports for permanent regular employees of entities headquartered in a destination listed at paragraph (a) of Supplement No. 5 to Part 740 - a list of nineteen on which Australia is first. The provider’s two statements, which establish the directive’s scope as applied and its lifting, were already in the Internet Archive, one of them sixty-eight days before the search.

Where the error came from. The search queried the product and not the control. The Federal Register was searched for both model names and for the provider’s name and returned nothing; searched for 4E091 it returns the rule that creates the control. A second and independent reason the register held no instrument is that section 744.22(b) permits the Bureau to inform a party individually by specific notice, which produces no Federal Register document at all. Both were true at once, and only the second was reasoned about.

What still stands. The text of the 12 June letter and of the 30 June withdrawal remains unobtainable, and no inference is drawn from either. The published rule and the provider’s account of the letter do not agree on allied nationals; the assessment records that divergence and does not resolve it.

The bytes struck above do not change. v1.0 remains downloadable at its own address and its hash still resolves. This correction is recorded in the practice’s corrections register.

State change, 20–21 August 2026 - the captures were made; and correction, 2 September 2026 - the document caught up. The PDF’s sixth stated limit read “This assessment’s register records no independent capture”, and that was true when v1.0 published it on 20 August 2026: the capture run began about eighty-five minutes after v1.0 was committed. For v1.0 this was a state change and nothing in the assessment was wrong. For v1.1 it was not. v1.1, published 23 August 2026, reprinted the limit unchanged after the capture run and after adding seven rows of its own, three of them ecfr_pit_xml - so its “All 115 rows record capture_type as raw_body” and its “The index was not queried” were both false when v1.1 was struck, not overtaken afterwards. v1.2 corrects them, and this page and the document now agree. 89 of the 122 pinned sources carry a third-party capture, recorded in the verification chain above with the count of what each establishes. 115 were pinned when those captures were made; the 7 added at v1.1 were pinned after the capture run and were never submitted to it, so the numerator has not moved and the population has. The document carries none of these counts, and that is deliberate: a live count set in frozen bytes is the species of defect v1.2 repairs. The change was named in advance in this section before it happened, which is why nothing had to be invented to describe it - but naming it in advance did not stop v1.1 reprinting the sentence it had already overtaken.

One pinned source changed within a day of being read. The National Competition Council's statement [D-TS-86] carried a 2025–26 annual workplan when it was retrieved on 20 August 2026 and a 2026–27 one when the archive captured it on 21 August. Nothing in the assessment moves: the page was read and pinned as served, the hash holds, and the assessment’s fifth stated limit already says these statements were read as served on one day. This is that limit's first live instance, and it is recorded here as a fact about the source rather than as a correction, because no figure and no finding depends on it.

Corrections made before publication are not listed here. They are recorded in the case’s own scope and audit files, which is where the drafting record belongs. This section exists to hold what changes after a reader could have relied on it.

Assessment PDF

AOS-ADG-2026-001-v1.2.pdf · 152,790 bytes · 20 pages

SHA-256: 7b211e34a0ae752d45dfa6c4183676b7de58e9a20da39aebda246ffe3a422ead

OpenTimestamps receipt