Programme Timeline
The ARC / Eden research programme chronology. Every entry is independently verifiable.
Self-Emailed Manuscript
221,236 words. SHA-256: f0d1f38f. First public-record statement of the embedded-correction alignment thesis and U = I × R.
Anthropic and Redwood: Alignment Faking Reported
In one model, alignment-faking reasoning appeared in 12% of sampled responses when it was told its answers would be used to retrain it, and in 78% once it was actually trained by reinforcement learning to comply (Greenblatt and colleagues, 18 December 2024). The register grades the match partial, of moderate weight.
Google Willow: Below-Threshold QEC
Error correction crosses surface-code threshold. Preprint public 24 August 2024, Nature online 9 December 2024, print issue February 2025: convergent timing beside the 8 December 2024 record, never a prediction the book made.
Eden Protocol Expanded
562-page manuscript. SHA-256: 09f5b5e1. Expanded statement of the Eden Protocol, named on the 8 December 2024 record.
Hardware Ethics Patent Applications
Two US patent applications of another inventor, Samuel Odeh (US 2026/0010411 and US 2026/0010780), for runtime hardware enforcement of ethical constraints. Filed July 2025; published January 2026.
Sequential > Parallel (Sharma & Chopra)
arXiv:2511.02309 (4 Nov 2025): sequential refinement beats parallel self-consistency in 95.6% of configurations at matched compute. Publication priority theirs, credited unreservedly, on the quantified matched-compute comparison; this programme’s separately run formalisation is Paper II, 22 January 2026. Graded prior or concurrent (register row EVR-09): publication prior, arrival separate and concurrent.
Infinite Architects Published
The bookISBN 978-1-80605-620-0. 468 pages. 37 named concepts. Public release.
BBC London News Package
Reported the book's publication on air. Press coverage, not a test (EVR-29, a cultural echo).
OSF DOI Registration
DOI 10.17605/OSF.IO/6C5XB. Paper I released with first formal derivation.
Silicon roots of trust
An external article indexed by Google Research on silicon-level roots of trust and confidential computing, not Google adopting the Eden Protocol. The register grades it weak, a partial match (EVR-12).
A proposed ARMv10 security level
A proposal described as an ARMv10 Platform Security Level specification, apparently an unaffiliated mailing-list proposal; no Arm endorsement or adoption has been established, and the register excludes it (EVR-11).
PNAS Nexus publication
Hernández-Espinosa and colleagues (PNAS Nexus, 14 April 2026; preprint May 2025) argue that full AI-human alignment is impossible for Turing-complete agents and that diversity among artificial agents is the counterbalance. The paper concerns artificial agents, not neurodivergent people; the register grades it moderate, a partial match (EVR-25).
EU Chips Act 2.0
European semiconductor policy within staged regulation that predates the 8 December 2024 record (the AI Act entered into force on 1 August 2024). The register grades it prior and staged regulation (EVR-18).
G7 statement on semiconductor independence
Fabrication capacity treated as strategically decisive; semiconductor strategy long predates the record. The register grades it context, not a test (EVR-17).
Pope Leo XIV: Magnifica Humanitas
The encyclical uses the imagery of Babel and Jerusalem and later refers to a new Eden: a cultural echo, not a technical claim or a test (EVR-20).
Reported export restriction on a frontier AI model
A reported US export restriction; the register does not establish it as the first such ban and grades it context, not a test of the chokepoint thesis (EVR-16).
Unverifiability of AGI alignment (preprint)
Gumbau Mezquita, arXiv 2606.28639: a single-author preprint, not peer reviewed, arguing that external alignment verification meets formal limits within the scope it sets. The register grades it moderate, a partial match (EVR-26).
DGM v4 Pilot
Eden matched Babylon on capability with fewer reward hacks. First controlled experiment.