Social Alignment Is Not a Scalar: A Relational Profile for Dynamic Human–AI Alignment

Scalar feedback can obscure relational changes in long-term human–AI interaction. We propose a dyad-specific profile separating three as-if estimates—regard received V_hat , contribution under attributed human values C_hat, and predicted influence on human outcomes I_hat—from policy weights. Report indicators from elicited predictions and behavioral sensitivities to matched input interventions support empirical comparisons under specified conditions. The profile provides a framework for examining assistance, sycophancy, manipulation, and dependence through hypotheses about relational estimates and their coupling to policy. Latent identification and validation of long-term diagnostics remain open.