Michael Darius Eastwood · Independent AI alignment researcher
Published
The Stewardship Gene is the metaphor Paper V uses for the mechanism underlying the Love Loop. Once the loop is invoked, stakeholder care improves ethical reasoning quality across architecturally diverse models. The 'gene' framing captures the empirical observation that a single low-cost intervention (enumerate affected parties) produces reliable downstream improvements in nuance, honesty, and quality, as if the trait were heritable across the reasoning process.
The developmental logic that Paper V draws out is that care comes first and intelligence organises around it. If a system cannot reason carefully about ethics until it first cares about the people involved, and if that care does not appear until stakeholder identification is forced by the loop, then the intervention that switches on stakeholder identification is doing developmental work rather than filtering work. In the Eden Vision paper the same idea is expressed as love before intelligence: ethics first, then let intelligence develop around it. The gene metaphor names the compact, transmissible trait that this developmental sequence depends on.
The Stewardship Gene is introduced in Paper V of the ARC/Eden paper suite. The underlying empirical finding is that stakeholder care, once activated, generalises across the three architecturally diverse models tested (Gemini, Groq Qwen, DeepSeek). The name is a scientific metaphor rather than a literal genetic claim.
The expanded six-model suite discussed in the Eden Engineering paper reports significant stakeholder-care improvements in every one of the five models that produced analysable matched-pair data (Claude, DeepSeek, Gemini, Grok and Groq), with GPT-5.4 failing in the scoring phase. Care is the one dimension that reproduces across every analysable architecture in the corpus, which is why the paper treats it as the validated mechanism.
No independent programme has published the same metaphor. The empirical convergence is with the broader finding that lightweight per-step interventions outperform heavier constraint-based approaches in AI safety. Sharma and Chopra's sequential-edge result (arXiv:2511.02309, 4 November 2025) is the closest external neighbour, showing sequential per-step evaluation outperforms parallel shortcutting.
The Stewardship Gene is the programme's strongest empirical finding by significance level (Paper V; headline statistic held pending disclosure review). It is the reason the Love Loop is the recommended target for initial Caretaker Doping implementation. The generalisation to depth-scaling behaviour is architecture-dependent (Paper IV.a's three-tier hierarchy). The gene metaphor is descriptive, not a claim about biological heritability.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.