Valerii Silin

Embodied Biographical Alignment (EBA) explores the possibility that some alignment-relevant capacities may depend not only on information about human experience, but also on causally structured developmental history. This version refines the experimental core of EBA after a series of toy pilots in which history-generated differences could often be reproduced through replay, terminal-state transfer, or reconstruction of update dynamics. These negative and limiting results lead to a stricter experimental formulation: the lived agent should remain in the experiment as an independent developmental reference, while synthetic baselines should be compared with it not only in terms of current behavior, but also in subsequent learning under genuinely novel demands. The paper introduces a patient-zero design, a partial clone with developmental amnesia, the working concept of Latent Developmental Reserve, and Source-level population transfer. The central question is framed as a problem of developmental equivalence: which consequences of causally lived history must be preserved for a reconstructed agent — and subsequently the Source — to remain the same future learner when the future relevance of individual fragments of history is unknown in advance. The paper does not claim the existence of machine emotions, qualia, subjective experience, or fundamentally non-copyable experience. Instead, it proposes an experimental design that does not exclude the possibility of a functional developmental remainder by assumption.

Embodied Biographical Alignment (EBA) began from the hypothesis that information about a human life may not be equivalent to the causal consequences of living through a temporally extended developmental history. Published EBA v0.2.1–v0.2.3 progressively narrowed this intuition into testable questions concerning transformative integration, developmental equivalence, and plural developmental history. This philosophical–scientific research note returns to the wider horizon of the programme. It argues that alignment should not be reduced to control and asks whether humans and machine intelligence could develop enough shared history to remain intelligible to one another as both continue to change. The paper develops the concepts of relational biography, co-developmental coupling, post-relational development, developmental inheritance, and shared developmental lineage. It further proposes that future machine intelligence might become not merely a product of human design, but a developmental descendant of a long human–machine history. The paper does not claim machine consciousness, subjective experience, or automatic moral improvement. Its purpose is to preserve a broader research question while keeping the technical hypotheses of EBA open to falsification.

Embodied Biographical Alignment (EBA) proposes that some alignment-relevant consequences of learning may depend on temporally extended, causally structured developmental histories rather than on descriptions or terminal outputs alone. EBA v0.2.1 introduced biography as a temporally connected sequence in which earlier actions alter later options and consequences, together with a requirement that integration be transformative rather than merely archival. EBA v0.2.2 then framed developmental equivalence as an operational question: if systems are similar at a comparison point, will they learn, adapt, and generalize similarly after the same subsequent experience? This note extends that question from one history to multiple histories. Its central proposal is that plural developmental history should be tested as a possible determinant of future learning. The note distinguishes multiplicity, diversity, and biography; formulates four falsifiable hypotheses; maps competing explanations including joint training, multi-teacher distillation, model merging, dynamic routing, retrieval, multiplicity, and diversity; and specifies a minimal simulated environment with matched future experience. A positive result would support only a limited claim: that multiple causally structured developmental histories can leave a reproducible and causally identifiable effect on the subsequent learning dynamics of a unified system that is not fully reproduced by simpler matched alternatives. This note operationalizes only one component of the broader EBA proposal; reproducing this endpoint with a simpler mechanism would not establish equivalence to EBA as a whole. The proposal makes no claim about machine consciousness, subjective experience, wisdom, or automatic moral improvement.

Embodied Biographical Alignment (EBA) investigates whether the integration of multiple, finite, developmentally structured biographies can be used as a source of alignment-relevant learning for an artificial system. This note treats existing studies not as tests of EBA, but as sources of experimental results, methods, and constructions that intersect with particular components of its research question. Work on knowledge distillation, representation and policy transfer, path dependence, sequential reinforcement learning, continual learning, and training-state transfer shows that artificial systems can transfer more than final outputs; that similar performance does not guarantee matching internal mechanisms or represented information; that training history can leave persistent internal differences and influence subsequent learning and generalization; and that some causally relevant consequences of history can be retained in identifiable components of training state and transferred between systems. For EBA, these results form a map of existing experimental possibilities. On this basis, the note introduces an operational question of developmental equivalence: is reproducing a system's current functional state sufficient, or does reproducing its subsequent learning, adaptation, and generalization require preserving additional causally relevant consequences of its developmental history?

Current alignment methods largely attempt to transmit human intent through preferences, rules, constitutions, oversight, demonstrations, or learned representations. This paper proposes a complementary developmental hypothesis: some human values may be difficult to acquire robustly without temporally extended, embodied, socially embedded, and finite experience. Embodied Biographical Alignment (EBA) is an experimental architecture in which multiple autonomous agents live bounded biographies, accumulate episodic and autobiographical memory, form durable relationships, encounter irreversible consequences, undergo controlled cognitive aging, and later contribute biographical experience to a more capable source system. The central requirement is transformative rather than archival integration: the target is not a source system that merely possesses records of finite lives, but one whose internal organization has been persistently altered by their integration. EBA does not assume machine consciousness, nor claim that experience automatically produces morality. It instead defines a testable question: whether biography, embodiment, finitude, and longitudinal social participation improve value generalization, calibrated restraint, respect for autonomy, corrigibility, and resistance to simplistic proxy optimization beyond text-only or preference-only baselines. We position EBA relative to embodied and enactive cognition, lived-experience approaches, lifelong-agent memory, progress alignment, machine finitude, and scalable oversight. A prior-art scan finds substantial overlap in individual components but no exact match for the complete architecture identified here. We define falsifiable hypotheses, ablations, an integration test, ethical constraints, and a limited extension for alignment to a changing human population.