How Should Normative Training Data for AI Be Produced? An Interdisciplinary and Auditable Governance Framework

Post-training supervision used to align AI systems often encodes normative judgments about desirable behavior. This paper argues that such supervision -- and the adjacent specifications and evaluation artifacts that shape or assess model behavior -- should be treated as a governance object rather than as an ordinary annotation product. We propose Normative Training Data Governance (NTDG), an institutional architecture with five functions: representation, contestation, authorization, accountability, and validation/revision. The framework combines interdisciplinary and affected-stakeholder input, independently represented counter-positions, separation of authorship and approval, rationale-level provenance, conflict-of-interest review, and independent downstream evaluation. Its normative stance is constraint-first and autonomy-sensitive: task success is pursued only within contestable boundaries concerning rights, serious harms, personal authority, proportionality, and public-interest effects. The contribution is a governance-level synthesis, not a new theory of value or training algorithm. We also outline a testable research program using resource-matched baselines, component ablations, independent evaluation panels, and separate procedural and behavioral success criteria.