Michael Darius Eastwood · Independent AI alignment researcher
Published
The HARI Treaty is the book's name for a proposed international vehicle for AI governance at the manufacture layer. HARI stands for Hard-Alignment, Recursive-Improvement. This note describes what the book proposes and what it deliberately leaves undone.
The book uses nuclear non-proliferation as a reference frame. It is careful about the analogy: nuclear enforcement has genuine successes over decades, real limits, and specific failure modes that would translate to AI. The chokepoint analogy is that both regimes rely on a small number of chokepoint technologies whose supply is visible and controllable, and that both benefit from a small coordination surface (a handful of jurisdictions rather than a hundred).
The analogy's limit, which the book names, is that AI does not have a mushroom cloud. The signal that a violation has occurred is subtler and harder to attribute. That means the enforcement mechanism has to be at the manufacture layer, before the artifact ships, rather than at the use layer, after the artifact detonates.
The treaty as proposed would require that chip designs be submitted to an International AI Ethics Authority before manufacture. The review criterion is architectural: whether the design includes caretaker doping at the required level and whether meltdown triggers are properly integrated. The chapter does not settle what "the required level" is; that is engineering that has to be finished at the treaty-drafting stage rather than the book stage.
A named vehicle (the HARI Treaty). An analogue (nuclear non-proliferation, with named limits). A review layer (chip designs, at manufacture). A criterion frame (caretaker doping, meltdown triggers). A candidate authority (International AI Ethics Authority). A coordination surface (a small number of jurisdictions).
The statutory language. The enforcement authority within the treaty. The dispute-resolution machinery. The audit rights. The criteria for what caretaker doping "at the required level" means as an engineering specification. The transition provisions for chip designs already in fab. The relationship to existing export-control regimes.
The HARI Treaty as a named vehicle has not been adopted. The review-at-manufacture principle appears in inverted form in the EU Tech Sovereignty Package of 3 June 2026, where the concern is anti-foreign-kill-switch sovereignty rather than architectural safety review. The US Fable 5 order of 12 June 2026 is a model-layer action rather than a chip-layer action, and it operates through export control rather than through architectural review. The two policy vehicles are compatible with the HARI framing without instantiating it.
A named vehicle, an explicit analogy, a review-at-manufacture criterion, and a candid list of what the book left for the drafters. The book is honest that it is not the drafting stage. It is the stage at which the shape of the instrument becomes visible. Six months of policy discussion since publication have not adopted the vehicle by name and have not contradicted the shape.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.