The short account of the whole project, for someone who has thirty seconds. The front page is at ../README.md; what is still open is in the worklist.
The project in one page
The problem, in two sentences
American law has a doctrine for the situation where a company’s ordinary operation puts the public in danger and no single act of wrongdoing can be pinned on anyone: the responsible corporate officer, from United States v. Dotterweich (1943) and United States v. Park (1975). The person who held the practical authority to prevent or halt the condition answers for it, whatever their job title said, and delegation is not a defense. That doctrine has never been extended to frontier artificial intelligence, and every American AI statute now in force places its duties and its penalties on the company alone.
What this project is
A public-domain model state statute that makes the extension, drafted in the open, with its audits, its corrections, and its whole drafting record published beside it. It is not a proposal paper about what a law should contain. It is the text, tagged and versioned, with section-by-section Comments, ready for a legislative counsel to conform to a state’s code the day a sponsor takes it up.
Its architecture, in one paragraph. Personal, non-delegable duties attach to natural persons with final material authority to prevent, halt, restrict or correct covered conduct. Technical work, access, advice, or implementing another person’s decision does not create liability. A covered model exceeds [10²⁶] operations or is prospectively designated frontier-equivalent; below that, records duties only. The duties are validation, factual certification after reasonable inquiry, records, and incident reporting. Consequences follow only when the statutory elements are proved.
What actually exists, counted
| The statute | 611 lines, 32 sections, tagged v3.4, with an 11,000-word section-by-section commentary |
| The repository | 67 documents, mirrored as a searchable site. Twenty-one were retired on 26 August 2026 — they were signposts saying “this moved”, and the links they redirected now point at the destination directly |
| The drafting record | more than 460 commits; every version’s redline preserved; 16 cures adopted verbatim at v3.4 and 26 more plus 4 open questions drafted and waiting for v3.5 |
| Corrections | A numbered, append-only corrections register: 61 entries, numbers reaching E75, each with the cause, what caught it, the fix, and the rule kept |
| Sources | A verification record of more than 200 rows — every instrument, its retrieval date, its grade, and, in its own section, the claims that failed verification and were withdrawn |
| The evidence shelf | 137 files held privately, including 24 primary bills and 13 congressional hearings read against their own text |
| Review | No reviewer has signed. Every open item is listed in the worklist |
| Publication | CC0 public domain, archived at CERN, DOI 10.5281/zenodo.22029795 |
Every number above is recomputed from the files by two committed tools, check_claims.py and check_links.py, which fail the build if a page and the truth disagree.
What the research found that nobody else has published
One. The provision exists, and one state has it. Illinois’s Artificial Intelligence Safety Measures Act requires an annual independent third-party audit, the designation of senior personnel, and the signature of the lead auditor. New York’s RAISE Act carried substantially the same words at § 1421(4) of print 6953-A on 3 June 2025. They came out in print 6953-B on 9 June, three days before the bill passed both houses. The provision existed for six days. Illinois then enacted it. It is the only place in American law where it is currently in force.
Two. It has been drafted four times and enacted once. The census traces the same audit-and-signature architecture through four attempts. One survivor. That is not forty legislatures independently converging on a design; it is one design surviving once.
Three. Nobody has ever had to explain why it came out. The New York Senate floor record of 12 June 2025 shows the bill called, the roll taken, and passage 58 to 1. No debate. No member laid it aside. No member asked why the audit had gone. The question is open, it is asked in public in this repository, and whatever comes back will be published as given — including “it was struck because it was wrong.”
Four. A Senate subcommittee has already stated the premise. At Too Big to Prosecute? (S. Hrg. 119-202, 16 July 2025), the chair of the Judiciary Subcommittee on Crime and Counterterrorism said of frontier developers: “have these Big Tech companies been prosecuted? No, of course not. They are getting off scot-free.” The subject of that hearing was copyright, not catastrophic risk, and nobody there proposed officer liability. What it establishes is that the enforcement gap is not this project’s inference.
Five. The evidence disappears without anyone destroying it. Two former researchers at a frontier developer told the Senate on 9 September 2025 that raw research data is deleted at ninety days as a matter of ordinary privacy policy — so removing a line from a report is enough to make the observation behind it unrecoverable. That is the argument for a records duty that attaches to the finding rather than to the data, and it is why this Act has one.
What the project asks of a reader, and what it gives back
The terms are the same for everyone. Scope agreed in writing before work begins. A complete response is three findings, verified or refuted, with reasons. Unpaid.
What is unusual is what happens to the answer. It is published entire, as written, under your name or anonymously as you choose. The maintainer may respond beside it and may not edit it. Model legislation in the United States comes from bodies that take outside views as input to a committee that then votes, and the act belongs to the committee. Here there is no committee and no vote. The nearest familiar shape is a conference paper, not peer review.
A review that refutes one of this project’s findings is worth more to it than a pass that agrees with everything. That is not politeness. One outside answer has already changed the statutory text, and the corrections register exists so that being right about a mistake is rewarded with a permanent, numbered, credited entry rather than a quiet edit.
What this project is not, stated before anyone has to ask
It is not law, not introduced, and not endorsed by anyone. Under the project’s own published rule, nobody — the maintainer included — may claim this text “survived review” until named reviewers sign, and none have yet. It is maintained by one person, unfunded and unaffiliated: not a company, a party, a government office, or an advocacy organization. The drafting is AI-assisted and says so on every relevant page; source selection, corrections and publication are the maintainer’s responsibility. The maintainer writes pseudonymously in public and is identifiable privately to reviewers before they sign.
And the known defects are published first, not last. An in-house adversarial pre-review returned seven defects going to the operative provisions, four of them in the tagged text — including that SEC. 6(a), the individual-liability offense this Act exists to create, cannot be pleaded as drafted. Those findings are on the front page, above the argument.
Where to start
- The argument end to end: the case
- The statute in plain English: the statute translated
- The strongest objections, with answers: known objections
- Every frontier AI bill in America, read: the census
- What has gone wrong here and how: the corrections register
- What is still open, and how to take a piece of it: the worklist
Corrections and questions: FrontierAIAccountabilityProject@proton.me