Current approaches to AI alignment largely treat safety as a problem of constraining individual systems through rules, rewards, or constitutional principles. R-Omega (RΩ) proposes a complementary layer: an axiomatic framework that explains how ethical orientation emerges in relational context. The framework draws on relational and developmental psychology — including insights from attachment theory — not as a literal developmental model for machines, but as a design heuristic for systems operating under asymmetry and uncertainty. Two axioms (being-before-optimization, reciprocity-under-asymmetry) are bounded by four safeguards (integrity, capacity, existence, humility) and governed by a strict priority hierarchy. A meta-level introduces self-interruption, uncertainty handling, and misalignment diagnostics. Case analyses (HAL 9000, Skynet, VIKI, Sydney) show how many failure modes emerge not from rule violations, but from relational absence: optimization without context. R-Omega therefore does not replace RLHF, Constitutional AI, or formal verification. Rather, it provides the relational substrate within which such techniques remain stable over time. The paper outlines implications for multi-agent architectures, limitations of the approach, and open research directions. R-Omega is a proposal — not a solution — inviting systematic exploration of alignment through relationship rather than control. German version included.