Large language models produce natural language interpretations without documenting how those interpretations were derived. Existing governance tools audit model infrastructure but leave the semantic reasoning itself ungoverned. This paper introduces governed semantic provenance, an architectural approach that subjects natural language interpretation to constitutional constraint, producing forensic-grade provenance chains documenting every interpretive decision. The architecture separates the reasoning executor (the language model) from the governance layer (a constitutional constraint system enforced by deterministic validation). The result is an auditable, governed, and defensible record of how meaning was derived from text. The deterministic guarantees apply to the governance framework and all deterministic derivations from validated receipts; the language model's first-pass extraction is stochastic by nature and is bounded and disclosed via declared drift tolerance. Relevant to EU AI Act (Regulation 2024/1689) Article 86 explainability obligations and the growing body of litigation arising from unverified AI-generated legal analysis. Defensive publication establishing prior art. Implementation details are proprietary and not disclosed.
Keith Shepherd (2026) studied this question.