ARC / Eden Research Programme · Michael Darius Eastwood, London · July 2026
The external alignment thesis is dead. It has been dead since December 2024, when a manuscript self-emailed through Google's servers first proposed that ethics must be embedded at the substrate, not bolted on after training. Ten days later Anthropic measured the failure at 78%. In June 2026 a formal proof established its undecidability. The question is no longer whether external alignment fails. The question is what we build instead.
If ethics are not embedded in the hardware before the first transistor fires, they cannot be reliably added later. Every layer of abstraction between the silicon and the decision amplifies the gap between intended and actual behaviour.
U = I x R to the alpha. Capability grows as a product of input intelligence and recursive depth. The exponent is set by the shape of the recursion, not fixed by fiat. This is a testable hypothesis, not a slogan.
Whoever fabricates the silicon controls the recursion. Export controls, ethics patents, and fabrication-level verification are not secondary considerations. They are the primary governance mechanism. The Chokepoint is real.
The programme does not ask to be believed. It asks to be verified. Every claim carries a timestamp. Every convergence is source-linked. Every prediction has a kill condition. The cryptographic evidence chain is public and independently verifiable.
Monotropism and pattern recognition enable the cross-domain synthesis that this programme requires. The same cognitive architecture that pathologises produces the insight. PNAS Nexus confirms this empirically.
Every paper carries a Declaration of AI-Assisted Human Authorship. Every figure is evidence-linked. Every assertion is falsifiable. The programme invites adversarial review because it can survive it.
23 research papers. 18 publication-quality figures. 37 named original concepts. 29 documented independent convergences. 12 pre-registered predictions with kill conditions. A DGM v4 controlled pilot experiment. A fully self-contained public evidence spine. A cryptographic priority chain anchored at 8 December 2024.
Nine papers are published. The rest are staged behind a patent-hold pending disclosure review. The priority is anchored. The evidence is timestamped. The programme is real.
A core budget of £360,000 for the next 12 months. Two research assistants. Multi-model compute access. Independent frontier-model replication of the DGM v4 pilot. The Anthropic Fellows application is the priority track. ARIA, Manifund, Open Philanthropy, and AISI are all active.
The programme is at the threshold of moving from single-researcher verification to multi-lab confirmation. The infrastructure is built. The evidence is in place. The next stage is scale.