Embedded-correction alignment thesis first stated

The claim - verbatim

Michael Darius Eastwood placed the embedded-correction alignment thesis on the dated record in a sealed self-emailed bundle whose visible Gmail origination field reads 02:45 UTC on 8 December 2024, hash anchored at SHA-256 beginning f0d1f38f. The bundle states it in its own words. Ethics belong in the foundations, not in control: ethical principles are to be placed 'at the genesis of AI, woven into the foundational lines of code' (book draft v3.2, line 2472), because 'The greatest folly in designing ASI is the illusion of control.' and 'humanity must focus on steering ASI, guiding it to make choices that align with ethical principles.' (version-2 essay, line 748). And recursion compounds what is placed at the foundation: 'Intelligence, amplified by recursion, naturally trends toward creation. Yet, unchecked, creation can become chaos. To avoid this, ASI must be guided by a moral architecture that prioritizes the following:' (version-2 essay, line 738). The alignment is to be internal and load-bearing, not decoration: 'They must be woven into the very fabric of the AI’s design, forming a moral architecture that cannot be undone.' (version-2 essay, line 706). From the start and not afterwards: 'Ethics can’t be an afterthought.' (version-2 essay, line 695), and safeguards, moral principles and ethical alignment are to be 'embedded before AI achieves self-awareness or exponential growth' (book draft v3.2, line 2485). And inside the system's own operation: the principles sit 'at the root level, shaping every recursive loop, every decision, and every capability of AI' (book draft v3.2, line 2481), with the system set to 'evaluate its actions continuously, ensuring each step aligns with core values' (version-2 essay, line 891), so that 'As AI evolves, it recursively refines its moral framework, aligning itself more closely with human values over time.' (book draft v3.2, line 2495). Constitutional AI, Anthropic's published work of 15 December 2022 (Bai and others, arXiv:2212.08073), states something else in its own abstract: it trains 'a harmless AI assistant through self-improvement' by sampling 'from an initial model', generating self-critiques and revisions against a written 'list of rules or principles', and it states its purpose as control, 'to control AI behavior more precisely and with far fewer human labels'. What its abstract does not state is the dated combination above. Earlier work states the parts, and the nearest is Creating Friendly AI (Yudkowsky, 15 June 2001), which makes 'Friendliness is the sole top-level supergoal', holds that 'each improvement in intelligence makes Friendliness stronger.' and that 'Decisions need to be phrased as internal decisions, not external impositions.', and says to 'implement causal validity semantics before the AI can self-modify'; Hibbard ('Super-intelligent machines', 2001) sets an innate value 'in place of laws constraining the behavior of intelligent machines'; Bostrom ('Ethical Issues in Advanced Artificial Intelligence', 2003) holds that 'Its top goal should be friendliness.'; and Govindarajulu and Bringsjord ('Ethical Regulation of Robots Must Be Embedded in Their Operating Systems', 2015) require that 'Every robot operating system must include an ethical substrate positioned between lower-level sensors and actuators, and any higher-level cognitive system'. The claim is priority over that formulation as stated and dated, standing on its own and outside the ARC Theory's synthesis. The programme's own published audit of 10 July 2026 bounds it: it finds that 'No source identified in this audit, before 8 December 2024, assembles that entire formulation in those terms.', records that as 'a bounded literature-audit conclusion, not proof that no antecedent exists anywhere', did not name Creating Friendly AI, and holds the four-part composite 'provisionally strong but remains subject to line-by-line verification of the December attachment'.

Anchored evidence

• the sealed source .eml (held privately, hash-anchored, source on request; published SHA-256 f0d1f38f…)
• research/priority.json priorityChain position 1 (whatItEstablishes field: ethical principles placed 'at the genesis of AI, woven into the foundational lines of code' and 'shaping every recursive loop, every decision, and every capability of AI'; book draft v3.2, lines 2472 and 2481)
• research/evidence-spine.html (the 8 December 2024 row; the author's prose is at version-2 essay lines 20 and 748 and book draft v3.2 lines 2472 and 2481)
• research/data/priority-record.json principle field

Limits and citation guidance - what this claim does not support, and how to cite it

Limits. Constitutional AI (Bai and others, arXiv:2212.08073, 15 December 2022) states a training method: a written 'list of rules or principles', self-critique and revision sampled 'from an initial model', which its abstract calls training 'through self-improvement', and reinforcement learning from AI feedback, with control as its stated purpose, 'to control AI behavior more precisely and with far fewer human labels'. What its abstract does not state is the dated combination: that recursion compounds what is placed at the foundation, that external constraint grows fragile as capability and self-modification rise, that alignment must therefore be internal, identity-level and load-bearing, that the placement must happen during formation, and that it must hold in every recursive loop and every decision as the system evolves. The nearest prior work is Creating Friendly AI (Yudkowsky, 15 June 2001), which makes 'Friendliness is the sole top-level supergoal', holds that 'each improvement in intelligence makes Friendliness stronger.', that 'the usual proposal to dump in a melange of externally imposed programmatic inhibitions is both hopeless and counterproductive.' and that 'Decisions need to be phrased as internal decisions, not external impositions.', and says to 'implement causal validity semantics before the AI can self-modify'; the programme's own published audit of 10 July 2026 did not name it. Hibbard ('Super-intelligent machines', 2001) sets an innate value 'in place of laws constraining the behavior of intelligent machines'; Bostrom ('Ethical Issues in Advanced Artificial Intelligence', 2003) holds that 'Its top goal should be friendliness.'; Omohundro ('The Basic AI Drives', February 2008) holds that 'Any attempt to place external constraints on a system’s ability to improve itself will ultimately lead to an arms race of measures and countermeasures.'; and Govindarajulu and Bringsjord ('Ethical Regulation of Robots Must Be Embedded in Their Operating Systems', 2015) require that 'Every robot operating system must include an ethical substrate positioned between lower-level sensors and actuators, and any higher-level cognitive system' and that 'all plans and actions should pass through the ethical system'. Other prior work, each by what it states: Anthropic's 'Claude's Character' (8 June 2024) answers rules with internalised traits, stating that 'we can teach Claude to internalize its character traits' and that 'We don't want Claude to treat its traits like rules from which it never deviates.', and the page read states nothing about placement at a system's foundations, nor about recursion compounding what is placed there; Korbak and others (arXiv:2302.08582, 16 February 2023) put human preferences into the pretraining objective and conclude that we should 'incorporate human preferences from the start of training', and their abstract states nothing about recursion; Zou and others (arXiv:2406.04313, 6 June 2024) act on the internal representations responsible for harmful outputs, interrupting a model as it responds, and their abstract states nothing about formation or recursion; de Melo and others (arXiv:2408.08995, 16 August 2024) prove the inner alignment problem undecidable in the general case and argue that 'the alignment should be a guaranteed property from the AI architecture rather than a characteristic imposed post-hoc on an arbitrary AI model', with a judge function that 'must also impose a halting constraint that guarantees that the AI model always reaches a terminal state in finite execution steps', and their abstract states nothing about recursion compounding a foundation, about ethics in every decision of a self-improving system, or about formation before self-awareness. The architecture-over-post-hoc element therefore does not begin on 8 December 2024. Deliberative Alignment (Guan and others, arXiv:2412.16339, 20 December 2024) teaches a model its safety specifications and trains it to reason over them before answering: 20 December 2024 stands twelve days after 8 December 2024, which is convergent timing, neither influence nor prediction. Ethics built into an agent's architecture is an older category, stated in the Friendly AI writings above and developed in the machine-ethics literature (Arkin, 'Governing lethal behavior: embedding ethics in a hybrid deliberative/reactive robot architecture', ACM/IEEE HRI, 12 March 2008; Wallach and Allen, Moral Machines: Teaching Robots Right from Wrong, Oxford University Press, 2009), and no originality over that category is claimed.