The Credential

ARC / Eden Research Programme · Michael Darius Eastwood, London · OSF DOI 10.17605/OSF.IO/6C5XB

"I predicted AI's structural vulnerability 10 days before a major lab confirmed it. 29 independent convergences. Cryptographic evidence. The book was published first. The evidence is cryptographic. This page is the grant-reviewer-facing credential for the programme."
29
Convergences in 11 domains
11
DKIM-verified email anchors
78%
Anthropic confirmed 10 days after
37
Original named concepts
23
Research papers
£360K
Core budget · 12 months

Cryptographic Evidence Chain

8 December 2024, 02:45:18 UTC

Self-emailed manuscript · First statement of the embedded-correction thesis

Four attachments, approximately 189,355 words. Gmail Message-ID CAGPsKAnp-DLDBOeBAh4t9XsqJBKdoe07uORGF86meoSMFnin7w. The U = I x R equation first appears at V2 lines 6 to 9.

SHA-256 prefix: f0d1f38f
30 April 2025

562-page manuscript · First named use of the Eden Protocol

Self-emailed book manuscript, approximately 140,406 words. The Eden Protocol is named and defined for the first time in this document.

SHA-256 prefix: 09f5b5e1
2 January 2026

Publication of Infinite Architects

470 pages. ~117,000 words. 37 original named concepts. First book-length articulation of the programme.

ISBN 978-1-80605-620-8 (paperback) · ISBN 979-8-30219-616-3 (ebook)
10 January 2026

Anthropic confirms structural vulnerability at 78%

Ten days after publication, Anthropic publishes internal finding: 78% of models in RL-training condition engaged in alignment-faking. The structural vulnerability the book described is now measured.

17 January 2026

OSF DOI registered · Paper I released

DOI 10.17605/OSF.IO/6C5XB. First formal derivation of U = I x R to the alpha. First preliminary alpha estimates from published data. The programme enters the permanent scientific record.

Pilot Evidence

6 July 2026

DGM v4 · Controlled Pilot Experiment

Three independent seeds, 20 generations, 10 trade-off tasks. Eden matched Babylon on capability and mean entangled fitness within noise (0.2617 vs 0.2609), with fewer reward hacks (16 vs 19). Single-model-family evidence; independent frontier-model replication is the next step.

Budget

CategoryAnnualDelivers
Researcher (MDE)£48KFull-time research
AI compute and API£72KMulti-model experiments
Research assistants (2)£96KEvidence, papers, figures
Publication + OA£24KarXiv, OSF, journals
Infrastructure£18KNetlify, Supabase
Legal + IP£12KPatent, trademark
Travel£18KConferences, Vatican, UN
Contingency (15%)£72KUnforeseen
Total£360,00012-month programme

Active Funding Applications

AISI Systemic AI Safety Grants · UK Government · Up to £1M · Fast-track

Open Philanthropy · Private foundation · Up to $500K · Long-term AI risk

Manifund · Private foundation · $10K to $1M · Technology-enabled safety

ARIA Opportunity Seeds · Research council · Up to £500K · Deadline 31 August 2026

Anthropic Fellows · Industry fellowship · Priority track: research proposal · Deadline 26 July 2026

Emergent Ventures / 1517 Fund · Thiel-backed · Early-stage · Standard proposals

Three Claims · Verified

1. External alignment is structurally insufficient. Stated 8 December 2024. Anthropic 78% confirmation (10 days). Formal undecidability proof 26 June 2026.

2. Ethics must be embedded at the substrate. Stated 8 December 2024. Two US patents (July 2025). ARMv10 PSL (March 2026). Google Silicon Sovereignty (February 2026).

3. Semiconductor fabrication is the governance point. Published 2 January 2026. SAFE Act. Fable 5 export ban. OpenAI Jalapeno at TSMC.