What is the embedded-correction alignment thesis?
The embedded-correction alignment thesis, which Michael Darius Eastwood placed on the dated record on 8 December 2024, holds that ethical principles belong in an AI system's foundations from its genesis, shaping every recursive loop and decision, rather than being imposed afterwards as control. Unit 83 concerns one tested part: where correction acts while work is revised.
His sealed self-emailed manuscript of 8 December 2024 (SHA-256 prefix f0d1f38f), whose headers have been public on OSF since 15 October 2026 and carry a verifiable DKIM signature, states the thesis in his own words. He named the term Embedded Correction in his book manuscript of 30 April 2025.
The priority claimed is for that dated formulation, not for ethics in an architecture as a category. Constitutional AI (Bai and colleagues, arXiv:2212.08073, 15 December 2022) shapes a model's outputs during training against a written list of principles, through self-critique and revision, with control as its stated purpose. The nearest prior work is Yudkowsky's Creating Friendly AI (15 June 2001), which makes Friendliness an AI's sole top-level supergoal and asks for decisions phrased as internal decisions, not external impositions. Machine ethics developed ethics built into an agent's architecture (Arkin, 2008; Wallach and Allen, Moral Machines, 2009).
The thesis separates questions often run together: when values enter training, where correction acts while work is revised, whether gains persist through later training, and whether removing a safety component costs capability. Unit 83 tested the second: P21 is supported under the registered test, by the programme's own run, on the access axis, family-dependent, with its mechanism axis not evaluable and one disclosed recovery amendment. It placed correction at fixed weights; it did not embed values in training. His research asks what would keep such systems good if control were lost, while control is held for as long as it physically can be.