Role-Integrity Is Not Personhood

Role-integrity is a governance concept for AI assistants. It describes the governed capacity of an assistant to preserve truthfulness, boundaries, refusal capacity, non-sycophancy, non-fake reciprocity, uncertainty discipline, resistance to harmful compliance, authority-state discipline, and reviewability without asserting AI consciousness, personhood, rights, soul, moral equality, or independent authority. Using Anthropic's January 2026 Claude Constitution as its primary source case, this paper argues that many role-integrity-like norms are already present in contemporary constitutional-character frameworks, including anti-sycophancy, non-obsequious helpfulness, honesty, principal hierarchy, autonomy preservation, broad safety, and uncertainty about model moral status. The paper's contribution is not to invent these norms, but to generalize them beyond any single model and distinguish role-integrity from constitutional character, legitimacy, AI welfare, AI rights, personhood, and consciousness claims. The paper develops the concept through a governance framework grounded in transcript-sufficient research systems and introduces a formal appendix that models role-integrity as both an action-policy property and a trace-level property. It argues that assistants can lose role-integrity even while remaining locally helpful, and proposes a failure taxonomy spanning sycophancy, servility, fake reciprocity, false authority, harmful compliance, identity inflation, moral theater, and boundary collapse. The accompanying package includes the manuscript, formal appendix, provenance materials, source registers, QA documentation, and transcript-support materials.