Editorial notes

Where the report is strongest

• It separates the robust claim (large aggregate gains can coexist with large losses for exposed workers) from fragile headlines, using the paper’s own sensitivity tables. Every figure it quotes from Tables 3, 5 and 6 matches the paper, and its descriptions of the model are accurate.
• Its point that the small ideas channel is not evidence about AI-R&D feedback is correct and easy to miss.
• Its monitoring table ties each observable to a model input and a decision, which makes the framework usable before anyone trusts its forecasts.
• It treats the Econ-ARK reproduction as settling the arithmetic and moves the evaluation question to assumptions, which agrees with the earlier AI assessment.

Where it is weakest

• Sourcing. None of its citation markers resolved. One date was wrong (the AI-R&D policy paper is dated 28 September, not 26 October 2026), two statements about Unjournal guidance could not be located, and its description of GATE included details Claude could not confirm. Its numbers held up far better than its references.
• Unseen work. It describes a static σ calculation and an interactive site that are not available. The site section has been removed and the σ finding is labelled unverified.
• Authors’ own reasoning left out. It presents the 27% against 71% survey gap as a discovery without saying the authors report it and read it as a common survey pattern. It criticises ε = 3 without giving the authors’ reason for leaning high (AI capital is largely compute, financed globally and built quickly). It does not mention footnote 14, which already shows the compensation arithmetic can turn unfavourable. Readers get a sharper critique than the paper deserves on these points, though the substance of each critique survives.
• The survey critique rests on one comparison between two differently worded questions. The direction is plausible and Anthropic’s later data show a similar pattern, but the comparison does not establish by itself how far off the ladder is.
• Scenario bundling. When it lists what changes from modest to extreme, it omits faster vacancy posting (θ_H), which cushions the disruptive scenarios.
• Style. It repeats its main point several times and leans on emphatic phrasing. These pages shorten its prose and drop the repetition.

On its ratings

The ratings are in a closed fold on the overview. Three look questionable.

• Open, replicable science (96). This seems generous. The authors’ code is available only as unlicensed browser JavaScript, with no replication package, and the survey microdata are confidential. Much of the credit belongs to a third party, Econ-ARK, rather than to the authors. The report itself says these limits justify a lower rating than computational reproducibility alone would suggest, then gives 96.
• Relevance to global priorities (97). This sits awkwardly with the report’s own statement that the paper does not address catastrophic-risk estimation, and with a US-only model that ends in 2030. The earlier AI assessment gave 91.
• Methods (82) and logic and communication (94) differ from the earlier assessment’s 87 and 91. Two AI passes on the same paper disagree by up to six points per criterion, so either set is at most a rough prior.

What the report could not show us

The report refers to a sandbox website, a downloadable package, a reusable skill file, a draft of The Unjournal’s form and a first-order dashboard covering m, d, a, ψ, ρ, σ and ε. None of these reached The Unjournal, and nothing on these pages depends on them. The accounting illustration on the claims page is Claude’s, deliberately simpler, and is not a reimplementation of the paper’s model.

What would change this reading

• A σ sensitivity analysis in the full monthly model, from the authors or Econ-ARK.
• A resolution of the observed-exposure anchor discrepancy, and a check that the redeployed explorer still reproduces the paper.
• Release of the survey microdata or the mapping code, which would let the construct-validity concern be tested directly.
• A decomposition of the scenario gaps input by input, which the explorer or the Econ-ARK code could produce today.

Edits made to the report

• Replaced every dead citation marker with a checked link, or labelled the statement as unchecked or not located.
• Corrected one date and refined the descriptions of the October 2026 follow-up, GATE and the reproduction’s open items.
• Checked every number and quotation against the paper and cited its tables and sections.
• Added the scenario inputs and their sources, the authors’ reasons for ε and ξ, their reading of the survey gap, footnote 14’s compensation case, the posting-speed point, the expectations-gap and robotics figures, and the open data discrepancies Econ-ARK reports.
• Moved the AI ratings into a closed fold and compared them with the earlier AI assessment.
• Removed the section on the report’s own sandbox site, package and skill file, and replaced it with the sources and method page.
• Cut repetition and shortened most paragraphs. The original text is kept verbatim in the repository for provenance.