About This Site
How it was made
Section titled “How it was made”This site was drafted quickly in early 2026 by Ozzie Gooen (QURI) with substantial assistance from large language models, primarily Claude. Read it with that in mind.
Revisions
Section titled “Revisions”June 2026: hardening.
- A nine-agent review covered the whole corpus, followed by a fix pass against its findings.
- Flagged fabricated or misattributed citations were removed or corrected. Real-history claims (dates, court cases, nuclear and aviation figures) were checked against secondary sources, mostly Wikipedia plus a few official pages such as nrc.gov. Few were checked against primary documents.
- The propagation math was re-derived, and one canonical worked example replaced several inconsistent ones.
- Hypothetical scenarios were labelled, and a ~25,000-word fiction section was removed.
- Vocabulary was consolidated against a written terminology canon.
October 2026: audit and contraction. Model-written audits found that problems remained: retired terms still in use, formulas that contradicted the core one, invented numbers presented as data, citation errors, and examples frozen in 2023. The site was cut from about 300,000 words to roughly 90,000 . Deleted pages redirect to their nearest surviving page, and the surviving core was corrected. The audit reports, with quoted evidence, are in the source repository under audits/2026-10/.
Book edition withdrawn. The v1.0.0 PDF predates the June 2026 and October 2026 revisions and has been withdrawn. Until a new edition ships, the website is the authoritative version; watch GitHub releases.
Limitations
Section titled “Limitations”- Little expert review. Most pages have not been reviewed by domain experts.
- Illustrative numbers. Most numbers are illustrative and labelled as such. Where a number is cited, check the source before relying on it; errors of the kind the audits found may remain.
- Untested. The framework has not been tested on deployed AI systems. The recommendations are reasoned, not validated.
- Labelled hypotheticals. Worked examples and scenarios that are hypothetical are labelled as such.
- One speculative section. The Power Dynamics page is speculative and marked so.
The parts that seem most robust are borrowed: expected-loss accounting, common-cause failure analysis, least privilege, and AI Control’s adversarial methodology. What the site adds is the combination, applied to agent harnesses. Whether that combination earns its keep is an open question.
How to help
Section titled “How to help”Open an issue on GitHub. The content is licensed CC BY 4.0. The most useful contributions:
- factual and citation corrections;
- logical inconsistencies between pages;
- real incidents, positive or negative;
- measurements of error correlation between checks, which would test the site’s central number.
Acknowledgments
Section titled “Acknowledgments”The framework draws on Eric Drexler’s CAIS, Redwood Research’s AI Control work, the economics of agency and insurance, and decades of nuclear and financial risk practice. Errors are ours.