June: six of twenty-one, and an address that led to another paper
June 2026 had thirteen records and none on Ukraine, courts, open weights or the EU's marking rules. An outside report offered twenty-one candidates; six entered: a Ukrainian dataroom for military AI, a Texas order sanctioning invented citations, a West Virginia rule on AI and criminal discovery, two open-weight releases (Cohere and Z.ai) and the EU code for marking AI-generated content. The main finding: two of the report's 'independent confirmations' were not, and one address led to a different paper.
Where this pass came from
June had thirteen records, seven of them from the first seed on a manufacturer’s page alone. Empty lines: Ukraine, courts, open weights, the EU’s rules on marking generated content. The brief followed scheme 1b: the outside report (Gemini) gave names and addresses for twenty-one candidates, and the session read every date and figure itself.
What entered
Six records out of twenty-one candidates.
Ukraine, 11 June. The Ministry of Defence reported that more than 100 companies had access to Brave1 Dataroom, where they train military AI on real datasets. The figure is the ministry’s own, so the record says so and carries medium confidence.
Two courts, 11 and 16 June. A judge in the Eastern District of Texas sanctioned a lawyer for a brief with an invented case, invented quotations and unsupported propositions: public reprimand, USD 5,000, a review of every filing she had signed since January. Five days later the Northern District of West Virginia forbade the defence to put sensitive criminal-case materials into any AI tool without the government’s written consent.
Open weights, 9 and 16 June. Cohere released North Mini Code (30 billion parameters, 3 billion active, Apache 2.0) and Z.ai released GLM-5.2 (a million-token context, MIT licence).
EU, 10 June. The Commission published the final, voluntary Code of Practice on marking and labelling AI-generated content, ahead of the duty that applied from 2 August.
The main finding: confirmations that did not confirm
The brief asks for independent confirmation of existing records. The report offered two: a Ukrainian news site for the export control on Fable 5 and Mythos 5, and Fierce Biotech for GPT-Rosalind. The first cites only Anthropic’s own announcement and a letter; the second is an article of 17 April. Neither is independent, and neither entered.
The addresses failed in three other ways. The case number given for the Texas order was wrong (the order is 4:25-cv-773). The address for a June preprint about logical reasoning opened a different paper, on music identification. And the document offered as the Justice Department’s complaint against Colorado’s AI law is a complaint in intervention filed on 24 April, in a case brought by X.AI.
What did not enter
Fifteen candidates: a ministry explainer of its own strike systems’ capabilities with no independent measure; a live Waymo statistics hub; a hardware blog dated in May; and eight preprints that exist with the right titles but have no replication, review or adoption yet.
Budgets
Six records cost the size budgets little; no limit was raised.
Entry written September 29, 2026
Commits this entry accounts for
fcca114