Semantic Governance and Global Reporting in the LLM Era: Semantic Relativity Theory as an Epistemological Framework for International Interpretive Stability

This paper investigates the semantic reliability of Large Language Models (LLMs) when processing, paraphrasing, and interpreting global sustainability and climate-related reporting frameworks. While industry narratives increasingly claim that AI systems can “harmonize” complex regulatory standards—such as IFRS S1/S2, TCFD, and the CSRD—no empirical evidence supports such assertions. To address this gap, we apply the Paraphrastic Resistance Index (IRP), a structured methodology designed to quantify semantic drift, structural degradation, and conceptual substitution in model-generated outputs. Four controlled experiments were conducted using five advanced LLMs (GPT-5.1, Grok 4.1, Copilot GPT-5, Claude 4.5 Sonnet, and Gemini 3), each tasked with paraphrasing key regulatory passages under identical prompting conditions. Results show a clear divergence in semantic stability: GPT-5.1, Grok, and Copilot consistently achieved high IRP scores, preserving both structural and referential integrity. Claude exhibited conservative but stable behaviour. Gemini 3 demonstrated systematic collapse, losing conceptual anchors in normative and legally dense contexts. These findings reveal that LLMs do not maintain intra-framework semantic fidelity reliably, and therefore cannot be used to generate inter-framework equivalence between ISSB, TCFD, and CSRD provisions. Theoretical analysis is grounded in Semantic Relativity Theory (TRS v2), which explains observed failure patterns through constructs such as Unsupervised Observer Collapse (CNO), Distortion of the Semantic Field (DCS), and Semantic Gravity. Overall, the study concludes that automated regulatory harmonization using LLMs is methodologically invalid at present and proposes IRP-driven semantic governance as a necessary safeguard in sustainability reporting, auditing, and compliance applications.