Michael Darius Eastwood · Independent AI alignment researcher
Published
The Eden Protocol is the complete engineering specification for embedded AI alignment. Rather than bolting safety onto a trained system through prompts, guardrails, or reinforcement from human feedback, the protocol requires safety mechanisms that participate in the recursive computation itself. Its components are the Three Ethical Loops (Love, Purpose, Moral), an entangled loss function that multiplies capability and safety instead of adding them, and a governance architecture that graduates autonomy as trustworthiness is demonstrated. Its central engineering commitment is that removing the ethical architecture must degrade the capability it supports, so that alignment is load-bearing rather than decorative.
The control-failure thesis and the embedded-alignment solution appear together in the 8 December 2024 manuscript (HRIH, line 14005: 'AI systems cannot be truly controlled. They will evolve beyond any safeguards.'; Version 2, line 20: 'embedding moral and ethical frameworks'). The name 'Eden Protocol' is introduced in the 30 April 2025 manuscript, alongside the Babylon and Eden civilisational framing. The engineering specification is later formalised as the Eden Engineering paper and its philosophical counterpart, Eden Vision. Both manuscripts were Google-server-timestamped via Gmail Message-ID.
Independent hardware-level convergences are documented in the register. Odeh's US patents 2026/0010411 and 2026/0010780 (filed July 2025, published January 2026) describe runtime ethics gates at the hardware level. Petrie's arXiv:2509.07637 (September 2025) proposes embedded off-switches for AI compute at fewer than 1% die overhead. FlexHEG (Petrie, Aarne, Ammann, Dalrymple, arXiv:2506.15093) proposes tamper-proof guarantee processors under ARIA commission. None cite the manuscripts; they arrive at the same structural principle by their own routes.
The empirical core of the protocol (the Love Loop) is validated at behavioural level in Paper V with a Fisher's exact test at p = 6.3 x 10^-21 across three architecturally diverse models. Paper VIII confirms load-bearing safety at the architectural level in the DGM experiment (p = 0.04, not surviving Bonferroni correction, stated explicitly). The weight-level experiment is inconclusive at LoRA rank 8; adapter fragility is the stated explanation. The hardware components (Caretaker Doping, Quantum Ethical Gates, Metamoral Fabrication Layers) are TRL 0 to 1 and are labelled speculative throughout.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.