Alignment Is Relational: Auditing Safety Guards Across Human Constituencies

Open guard models mediate what deployed language-model systems treat as "harmful," emitting a single safe/unsafe verdict, yet human judgments diverge sharply: on DICES-350, racial group majorities directly oppose each other on 77 of 350 conversations (22\%). We audit four open guards against disaggregated rather than collapsed human labels and find three patterns. First, every guard shows substantial alignment differences across racial-group majority targets (full-set AUC gaps .03--.14, widening to .14--.23 on group-conflicted items; all bootstrap CIs exclude zero). Second, guard confidence tracks rater disagreement in guard-specific ways, from WildGuard ($\rho = -.44$) to ShieldGemma-9B ($\rho = +.12$, uncorrected $p = .02$). Third, apparent differences in which group a guard best tracks are descriptive only (non-significant at $K=5$) and should not be read as distinct ``bias profiles.'' An achievable-alignment analysis shows DICES-350's disagreement structure does not itself preclude substantially more balanced binary agreement, giving a model-free reference for interpreting guard disparities. We release the audit harness and argue guard model cards should report perspectival alignment profiles alongside aggregate performance.