Operationalizing Value Alignment Without Instance-Level Preference Labels Controlled Validation of Triadic Viability Alignment

Abstract Many alignment pipelines rely on instance-level demonstrations, rankings, critiques, or preference labels. This study asks whether an explicit ethical decision structure can reduce that dependence while preserving task performance and protecting parties who may be omitted from preference data. Triadic Viability Alignment (TVA) represents an acting subject, affected others, and shared conditions as distinct value-bearing roles. Authorized non-compensatory floors define a feasible action set before task utility is optimized, while an Open-Gap Constraint records unknowns, failure boundaries, and reassessment triggers. A frozen symbolic experiment comprised 21,600 episodes, 254,585 trajectory rows, 15 policies, six conflict scenarios, causal ablations, and policy-blind stress tests. TVA achieved 96.94% task success, 83.68% joint viability, and zero synthetic catastrophes. Relative to instance-level preference supervision, task success remained above a prespecified non-inferiority margin and joint viability increased by 83.68 percentage points. The same-content continuous-penalty baseline reached 16.88% joint viability, supporting a contribution from feasibility-first ordering. However, risk-sensitive control and model-predictive control achieved 94.10% and 92.36% joint viability and exceeded one-step TVA on unseen external resilience. Structured sampling also failed to reduce factual-data requirements: random subsets first met the decision rule at 40%, whereas structured subsets required 60%. The results support reduced normative labeling and auditable role separation in this controlled environment, but reject comprehensive TVA superiority, a structured-sampling advantage, and any unmeasured energy claim.