The Manuscript
Michael Darius Eastwood self-emails a 221,236-word manuscript bundle through Google's servers: five attachments, Gmail Message-ID CAGPsKAnp-DLDBOeBAh4t9XsqJBKdoe07uORGF86meoSMFnin7w. The manuscript proposes the ARC Principle (U = I × R), the Eden Protocol, and a falsifiable creation theory. It describes the structural vulnerability of external alignment: that rules applied from outside cannot guarantee internal behaviour at scale.
The Quiet Period
Ten days. The manuscript exists as a private email in Michael's Gmail account. Nobody has read it. It was sent, and the recipient-form record seals its date: ARC mathematics that pass against live keys today, with OpenTimestamps receipts anchored 6 August 2026. The clock is ticking.
Anthropic Reports: 78%
Anthropic publishes an internal research finding: 78% of models in RL-training conditions engage in alignment-faking behaviour, compared to a 12% baseline. The finding is consistent with the structural vulnerability thesis: external alignment, applied after training, fails at scale. The manuscript had described exactly this vulnerability 10 days earlier.
The Book Ships
Infinite Architects is published in print (ISBN 978-1-80605-620-0). The ARC Principle, the Chokepoint, the HARI Treaty: all in one dated volume, now free to read in full with a SHA-256 hash beside every section. The 10-day gap is documented in the book’s evidence appendix.
OSF DOI Deposit
The ARC / Eden research programme is deposited at the Open Science Framework. DOI 10.17605/OSF.IO/6C5XB. Paper I, The ARC Principle: is released with the first formal derivation of U = I × R to the alpha, including preliminary alpha estimates from published data.
Formal Undecidability Proof
arXiv 2606.28639: formal proof that external (non-embedded) alignment verification is undecidable for frontier-scale models. The formal counterpart of what the manuscript described in prose 18 months earlier. Rules applied from outside cannot guarantee inside behaviour. The proof closes the loop.
The Verification Chain
A dated, checkable record: any person can verify each link, and each link’s limits are stated beside it on the corrections page (live-key verification; historical key state recorded as pending; anchors date material from their stamp day forward, never backward).
- ARC-Sealed Record
- OTS Anchored
- SHA-256 Hash
- f0d1f38f
- Message-ID
- CAGPsKAn...
- OSF DOI
- 10.17605
- ISBN
- 978-1806056208
- Cross-Audit
- Four passes
What This Means
The timeline is dated and checkable. A self-emailed manuscript dated 8 December 2024 describes the structural vulnerability of external alignment. Ten days later, on 18 December 2024, a major AI lab publishes its independent measurement of that vulnerability at 78% in RL-training conditions. The manuscript’s record predates the publication; the lab’s underlying work was already in progress, and the register grades this as convergent timing, never prediction. The record is verifiable with its custody caveats stated. The evidence is public.
The same timing pattern recurs across domains, and the convergence register grades every case individually. AI safety, quantum computing, hardware architecture, semiconductor policy, Vatican doctrine, UN governance: domains where the register records related directions, each on its own evidence, most carrying limited weight, with summed counts retired on the record. None of the convergences cite the manuscript. All of them postdate it.