For journalists and checkers
Press kit
Everything in one place, and everything checkable. Quotes carry their sources, the failure record is a first-class link, and nothing here asks to be believed.
Bios and photo
Three lengths, ready to lift, on the press page, with the author photo. The one-sentence form: Michael Darius Eastwood is an independent researcher and author in London, whose ARC / Eden research programme publishes its failures beside its results.
The claim, in its ruled form
The ARC Bound is the proposed stable limit: the most a recursively self-correcting system can grow and stay correctable. A scaling limit, not a speed limit, and proposed, not measured. The measured exponent on frozen systems is 0.49 under blinding; an earlier 2.24 was retracted in public.
Quotables, with sources
- “The honesty architecture IS the credibility architecture.” · the release
- “What I cannot claim is as important as what I can.” · the release
- “I work in public, on the record, and I ask to be checked rather than believed.” · about
- “The cryptographic evidence is published. You do not need to trust. You only need to verify.” · the release
The documents
- The whole story () · the programme end to end
- The evaluation-blinding methods report ()
- The funders brief ()
- The corrections log · every retraction and withdrawal, in public
- Replicate · the exact published commands, verbatim
Pitch drafts · not sent, awaiting the author
Draft · operator approval required before any use
The methods pitch (specialist press). An independent researcher ran the same AI-alignment measurement twice, once normally and once blind, and two of four models reversed direction. The finding forced him to retract his own headline number in public, and the retraction is the story: a one-person programme running the honesty machinery most labs promise. The four-layer blinding protocol, the sign-flip data and the retraction are all on the record, with commands to reproduce them.
Draft · operator approval required before any use
The record pitch (mainstream). On 8 December 2024 a London author sent himself a manuscript arguing AI cannot be permanently controlled, ten days before Anthropic published evidence pointing the same way. Eighteen months later his dated record runs through patents, papers and policy he never touched. He claims priority for nothing beyond what the timestamps show, publishes every failure beside every fit, and asks to be checked rather than believed.
Contact and photo: the press page. Nothing is ever sent from this site; the drafts above leave only by the author’s hand.