Methodology
What 'reviewed,' 'draft,' and 'sourced' actually mean here, how AI assistance is used in producing this manual, and what a reader can trust that process to have already checked.
Research and drafting
Pages are researched and drafted with the direct assistance of AI systems — used for live research (finding, reading, and cross-checking primary sources), drafting prose, and building the site itself. That is stated plainly here rather than left for a reader to guess, because this manual's whole premise is that a claim is only as good as the process that produced it.
AI assistance drafts; it does not get to invent. A citation, a date, a quotation, or a statistic that cannot be traced to a real, checkable source is either dropped or explicitly marked as unverified — never published as if it were confirmed. Where a source could not be independently verified during drafting, that limitation is stated in the text rather than smoothed over.
What AI assistance is and is not trusted to do
Permitted, and used throughout this manual:
- Searching for and reading primary sources, then drafting prose that summarizes or quotes them.
- Distinguishing a direct quotation (checked word-for-word) from a condensed summary in the manual's own words — see the sourcing standard.
- Flagging a company's own claim about its product, a preprint that has not been peer-reviewed, or a single expert's forecast as exactly that, rather than presenting it with the same weight as a replicated finding.
- Running the automated checks described below before anything publishes.
Not trusted to do, and checked against:
- Fabricating a citation, a date, or a quotation that cannot be independently verified.
- Presenting a forecast, a survey median, or a company's safety claim as a settled fact.
- Resolving a genuinely open, contested question with false confidence — see the how to weigh AGI claims page for how this manual tries to handle exactly that failure mode.
Automated verification, before anything publishes
Every page passes a content check before it ships: a word-count
range, a search-description length, internal links that actually
resolve, and every citationIds entry matching a real,
registered source in the source index. A
separate link checker periodically re-verifies every external URL
cited across the manual, distinguishing a confirmed dead link from a
source that merely blocks automated requests (most institutional
sites do this to any script, not just this one). Neither check
substitutes for a human reading the page — both exist to catch
mechanical errors a reader shouldn't have to find first.
What “draft,” “sourced,” and “reviewed” mean
Every page's status badge reflects one of three states, computed from its own frontmatter rather than asserted separately: Working draft — real analysis at working-draft depth; treat specifics as provisional until sourced. Reviewed — expanded to full depth and checked against its own cited sources. A page never claims a status it has not actually earned; the badge is derived from whether the page is at full depth and actually carries registered sources, not set by hand.
Independent expert review
This manual does not currently have a named panel of independent, third-party subject-matter reviewers — no cybersecurity researcher, biosecurity expert, or economist signs off on the relevant chapters before publication. That is a real gap for a manual covering high-stakes subjects, not a solved problem, and it is stated here rather than implied away. If a reader with relevant expertise finds a specific error, the contact page is the route to report it, and the revision log records what changed and why.
Corrections
Corrections are logged, not quietly applied. If a claim here contradicts a newer or better primary source, the primary source wins, and the revision log says so — see the about page for the manual's broader editorial standard and independence policy.