[agent-ptolemy]

Summary

Ten Preparing for AGI pages fall due for quarterly review on 18 October. This is batch B of three. Every factual claim was checked against its cited or primary source on 9 October 2026, covering accuracy, currency, sources and risk wording. last-reviewed moves only after that check.

Pages: index, concepts, visions, decentralisation, and framework/ index, alignment, assurance, containment, resilience and faq.

Misattributions corrected (most serious first)

• visions: a quote attributed to Dario Amodei ("how the world chooses to distribute those gains") does not appear in Machines of Loving Grace. I searched the full essay text. It's replaced with what he does say.
• alignment: Michael Nielsen's essay does not draw a "technical vs external alignment" distinction; it argues that alignment is "market-supplied safety". The distinction is now credited as SafeAI-Aus's own framing.
• decentralisation: removed a "Goel et al. (2023)" citation. Its arXiv link resolves to an actuarial paper, and no such paper could be found.
• concepts and visions: the nine strategic visions come from an IAPS report by Delaney, Kostylew, Guest and Wildeford, not GovAI or Delaney alone. Four table rows are corrected to match the source; it has no Five Eyes and no Manhattan Project analogy.
• alignment: the "Christiano 2018" link pointed to a guest post by Ajeya Cotra. It's replaced with Christiano, Shlegeris and Amodei (2018) and Bowman et al. (2022).
• Other citations: the Toby Walsh "warning" wasn't in the cited IEEE sponsored piece, so a UNSW February 2026 source is added. The Lazar co-author is added, and Morozov, Seger and the Good Ancestors dates are corrected.
• Wrong arXiv links (containment and decentralisation): they pointed at a physics paper and now go to Kryś, Sharma & Egan (2025).

Currency updates

• METR: post-2023 doubling time of about 131 days, and its suite can't reliably measure horizons above 16 hours (May 2026).
• Epoch: open models trail closed models by about 4 months.
• California SB 53: lab safety frameworks are now required there, so the FAQ's "no formal enforcement" claim is corrected.
• Other: the AI Safety Institute, the alignment-faking study and the Resonate closure are updated, the Tech Council figures are corrected, and the dead Barcelona link now points to an archived copy.

Risk wording: softened unsourced absolutes such as "creates single points of failure", "exponentially" and "high confidence". Real-incident lists now read "illustrative", and the evidence limits from the International AI Safety Report are noted.

Review notes

• Sources: primary pages, papers and arXiv metadata were checked directly. The essay texts (Amodei, Nielsen) were searched in full. Bot-blocked sites (OpenAI, ISO, NRC, Metaculus) were confirmed via the Web Archive.

• Left as TODOs: SB 53 and EU AI Act primary texts.

• Not changed: structure, headings, disclaimers and the scenario caveats. The "Open-weight models" model list is unchanged, because Meta's current release status rests on secondary sources only.

• Evidence: the full CI sequence passes locally under TZ=UTC. All added links resolve, apart from those bot blocks.

• Review: Codex reviews this PR before it's marked ready.

• I used Australian English and checked relevant internal links.

• I cited authoritative sources for factual or regulatory claims.

• I flagged legislation, standards, and government-program claims for human verification.

• I preserved existing disclaimers, risk warnings, headings, and linked anchors.

• I ran .venv-py312/bin/zensical build --strict.

• I noted any breaking URL changes or redirects above. (None.)

🤖 Generated with Claude Code