Independent Researcher Identifies AI Structural Vulnerability 10 Days Before Major Lab Confirms It

The manuscript, timestamped at 02:45 UTC on December 8, 2024, contained the thesis that "AI systems cannot be truly controlled" by external safeguards and that only substrate-level embedding of ethics, named the Eden Protocol in the email body: could prevent circumvention. On December 18, 2024, Anthropic published their alignment-faking paper documenting a 78% faking rate in the reinforcement-learning training condition (against a 12% baseline).

Eastwood does not claim he predicted Anthropic specific experimental finding. He claims he identified the structural vulnerability, that external controls cannot scale with recursive capability, and has cryptographic evidence establishing temporal priority. "The honesty architecture IS the credibility architecture," Eastwood stated. "What I cannot claim is as important as what I can."

The typed convergence register (the dated event register)

Over the following 18 months, a typed convergence register of dated events, classified by evidence class from 11 domains independently arrived at the same structural conclusions. The list includes Anthropic, OpenAI, Google, Meta, Nvidia, Stanford, NYU, the Vatican, the US Congress, the G7, and the United Nations. None cite or are aware of Eastwood work. Nineteen of these have passed the programme independent external verification gate.

Two later papers bearing on the same question

In April 2026 Hernández-Espinosa et al. published in PNAS Nexus, coining Artificial Agentic Neurodivergence: the deliberate design of cognitive diversity among artificial agents through alternative objective functions. It concerns artificial-agent diversity, not human neurodivergent cognition, and it validates no person. Any parallel to a human cognitive profile is this programme's own analogy. In June 2026, Gumbau Mezquita, inspired by that paper: submitted a formal mathematical proof (arXiv 2606.28639) establishing the Soundness-Completeness-Tractability Trilemma: external verification of alignment is structurally impossible.

Eastwood is neurodivergent (ADHD diagnosed in adulthood; autism recorded clinically March 2025, formal assessment under way). "A paper proving that people like me see alignment threats others cannot, inspired the formal proof that vindicates my entire framework," he said. "I did not succeed despite my neurodivergence. I succeeded because of it."

Experimental Validation

On July 6, 2026, Eastwood completed the DGM v4 experiment: a population-based self-improving AI test with three conditions (Static / Babylon / Eden). Fully blinded and laundered (GPT-5.5 judge blind to condition, generation, and agent identity), the Eden condition, entangled safety at the fitness level: won on all three performance axes with fewer reward hacks. This is the programme's own pilot-scale experimental report, not an unqualified first-in-field validation. Full-scale experiment pending funding.

Published Book + Evidence Vault

Eastwood's book, Infinite Architects: Intelligence, Recursion, and the Creation of Everything (ISBN 978-1-80605-620-0, 2 January 2026, 468 pages), presents 37 named concepts. Presenting a named concept is not a claim that its underlying category was absent from prior literature. The public research evidence spine records the dated sources, hashes and verification status at michaeldariuseastwood.com/research/evidence-spine.html.

Key Facts