Michael Darius Eastwood · Independent AI alignment researcher
Published
The two Odeh filings are the first published US patent applications to specify a runtime ethics gate as a discrete hardware unit. The primary claim is architectural: place a small, fast, policy-carrying block between the compute path and the memory bus of an AI accelerator, and have every candidate output pass through it before being written back. The applications quote a sub-5 microsecond ethical filtering latency, low enough to sit in the hot loop of a modern inference pipeline without collapsing throughput.
Patent applications are a specific kind of evidence. They are public, dated, and adversarially examined. A speculative essay claiming "we should put ethics in the silicon" costs nothing. A patent application claiming a specific block, a specific interface, and a specific latency envelope commits its filer to a concrete technical proposition. That does not make the proposition right; it does make it non-vague, and it makes convergence claims easier to check.
In the ARC/Eden convergence register the Odeh patents sit at row 7, classified CONVERGENT, three to nine months after the April 30 2025 manuscript. The manuscript's Caretaker Doping proposal argues that ethical constraints must sit inside the substrate rather than in a filter layer above it, because filters above the substrate can be trained around. The Odeh patents implement the same architectural conclusion at the level of a discrete hardware block. The convergence claim is structural: two independent proposals, same threat model, incompatible starting points, no cross-citation.
Patents describe idealised inventions. The sub-5 microsecond number is a target, not a measured production figure; the filings do not claim a taped-out reference. Anyone building on the applications will discover that "runtime ethics" is under-specified until you name the policy family (deontic constraints? consequentialist utility bounds? refusal semantics?). None of this makes the filings less interesting; it makes them a research programme rather than a finished product.
The three hardware convergences (rows 7, 8, 9 of the register) are worth reading as a cluster. Odeh sits in the compute path. Petrie sits at the licensing layer. FlexHEG sits alongside as a guarantee coprocessor. They are three architectural bets on the same underlying claim: policy that lives only in software is trainable-around, and any credible AI governance regime has to bind, in the end, to silicon that will refuse. Whether that is desirable is a policy question. The technical case that it is possible has now been made three times, independently, in nine months.
Both applications are on the USPTO's public application database. Search by application number. The specification is where the interesting engineering claims live; the claims section is where the legal boundary sits, and it is narrower than the specification suggests.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.