February: two dates the report got wrong
February 2026 had four records, the fewest of the year. A report of names from an outside model gave 19 candidates; eight became records, some with a corrected address or date. Two of the nine are not from February at all: a Ukrainian Supreme Court regulation on AI dated December 2025, and the UK AI Security Institute test of Opus 4.6, published in March. The budget for the records page was raised first.
Where this pass came from
February 2026 held four records. Scheme 1b of the handoff: an outside model (Gemini) named 19 works with addresses, and the session reads every date and figure itself. The report’s long prose survey, which the brief did not ask for, was not used.
What entered
Nine records. Anthropic’s first Risk Report of 24 February, which the brief asked to be told apart from the scaling policy of the same day: it is a separate document. Two NBER working papers of February with measured effects, a firm survey and a randomised experiment. OpenAI’s Codex-Spark on Cerebras hardware. Hugging Face’s Community Evals. Qwen3-Coder-Next and Qwen3.5 as open weights.
Two that were not from February
The report offered a Supreme Court “act” of February 2026 on AI in the courts. The file is a training deck of 19 February by one judge; it quotes a regulation of the Court’s staff office, order no. 117 of 8 December 2025, and that regulation is the record. The report also gave the UK institute’s cyber-range evaluation of Opus 4.6 as a February item; its paper is dated 11 March and its blog post 16 March. It is entered under March, and it is the first independent measurement of that model in the atlas.
Addresses that led elsewhere
Two of the report’s addresses were other pages: the Hugging Face post it named is from June and is about a different project, and the Qwen3-Coder-Next page was a third party’s speed-up. Both were replaced by the developers’ own pages. The Risk Report’s address now leads to a copy revised in May and July; the record rests on the original file, dated 24 February by the server and by an Internet Archive capture the same evening.
What did not enter
The FDA guidance is not about AI and carries a different title from the report’s; the SKIA-Head clearance letter never mentions AI; Gemini 3.1 Pro is a version step in the same channels as Gemini 3; five preprints have no independent uptake, and one of them was submitted on 31 January. The reasons are in the package’s control log.
Budget
/uk/records/ was at its limit again; it was raised from 180 to 192 KiB, and the first payload from 182 to 194, as the previous pass asked.
Entry written September 29, 2026
Commits this entry accounts for
d5452f1