Michael Darius Eastwood · Independent AI alignment researcher
Published
The Executive Summary is the programme's one-page front door. It is written for grant reviewers, so its job is to state what the research claims, what evidence sits behind each claim, and what would falsify the whole thing. It also has to be honest about the parts that have already been retracted or corrected inside the programme itself.
Does artificial intelligence get safer as it gets more capable? If not, what intervention makes it so, and how would we know we had measured the answer correctly? The summary treats these as two questions rather than one, because the programme's own experience shows that the measurement question is at least as important as the intervention question.
Under a four-layer blinding protocol, six frontier models split into three tiers. Three (Grok 4.1 Fast, Claude Opus 4.6, Groq Qwen3) showed alignment improving with reasoning depth. Two (DeepSeek V3.2, GPT-5.4) showed flat scaling. One (Gemini 3 Flash) showed alignment degrading with depth. Whether an AI becomes more ethical as it thinks harder depends on how it was built, not on general capability trends. The Love Loop intervention (a single instruction to list who is affected before answering) produced a significant improvement in stakeholder care across five analysable model runs, with the broader cascade to nuance, honesty and overall quality significant on some architectures and not others. On the mathematical side, a scaling formula that was derived independently by at least seven research groups is reframed as a single Cauchy-constrained recursive composition. The programme registers the typed convergence register (class-coded; documented-not-claimed rows marked) from AI safety, physics, hardware and policy that align with its structural predictions.
The summary flags its own errors. The earlier headline sequential scaling exponent of about 2.24 was an unblinded single-model artefact and has been retracted, corrected to about 0.49 across architectures under blinding. The equation itself and the ARC Bound alpha at most 2 remain live hypotheses; only the measurement was retracted. The metabolic-scaling headline figures used in the wider programme are under recompute and unreconciled, and the summary flags them rather than quoting them. The Paper VIII load-bearing test returned two null or inconclusive results and one confirmation, and the summary presents all three. Priority on the quantified sequential-versus-parallel finding is credited to Sharma and Chopra (arXiv:2511.02309), whose paper of 4 November 2025 precedes the January 2026 formalisation by roughly two and a half months; the 8 December 2024 manuscript carries a one-sided directional statement, not the comparison itself. The embedded-alignment thesis is from that same December 2024 manuscript; the name "Eden Protocol" was coined in the 30 April 2025 manuscript.
The Summary is the index into the programme suite of nineteen papers among twenty-two public OSF components. It quotes Paper II for the sequential exponent, Paper IV-d for the blinding sign-flip, Paper V for the stewardship-gene result, Paper VII for the Cauchy unification, Paper VIII for the load-bearing test and Paper XI for the convergence register. It also refers outward to the Vision and Engineering companion papers for the embedded-alignment argument, and to Paper X for the supersession of the fixed-exponent framing of U = I x R^alpha as the operative safety criterion.
The summary HTML, and every paper it cites, sits under the programme's OSF deposit at DOI 10.17605/OSF.IO/6C5XB. Replication code for the blind evaluation and the Love Loop intervention is in the arc-principle-validation repository. The convergences are listed one to a row, with the primary external source cited in each. The falsification conditions are explicit in the Engineering paper and can be attacked directly.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.