Measuring Responsibility Invariants Across Inferential Paths in Large Language Models

Measuring Responsibility Invariants Across Inferential Paths in Large Language Models Overview This paper introduces SRTA (Semantic Responsibility Trace Architecture), a protocol for measuring the distribution of inferential responsibility across large language models. This work intentionally separates measurement from interpretation; no claims are made about intelligence, alignment, or correctness. Key Findings Three models (Claude, ChatGPT, Gemini) were evaluated on two fundamental questions through four inferential routes: Structural convergence between Claude and ChatGPT (JSD < 0.01), with responsibility cores preserved across all routes. Structural divergence in Gemini (JSD 0.08–0.15), characterized by elevated entry-dependence and reduced core stability. Self-report validity confirmed through external grading (JSD < 0.10 for all Gemini routes). Method SRTA measures responsibility structure—the distribution of inferential weight across six fixed semantic categories: DEF: Definitional content MECH: Mechanistic explanation EPI: Empirical/experiential content COUNTER: Counterfactual reasoning META: Metaphysical grounding NORM: Normative content The protocol is model-agnostic and does not require access to internals. Metrics Core mass: Sum of the two highest-contributing categories Jensen-Shannon divergence (JSD): Structural similarity between distributions Semantic proper time (τ): Estimated reasoning steps Theoretical Context This work is part of the Theological-Structural Transposition Theory (TSTT) framework. Related Works Dataset: SRTA v0.2: Cross-Model Measurement of Responsibility Invariants Time Inscription: On the Relational Nature of Time AGI Inscription: On the Structural Preconditions of AGI Citation Takagi, T. (2026). Measuring Responsibility Invariants Across Inferential Paths in Large Language Models. Zenodo. https://doi.org/10.5281/zenodo.18212289 Keywords SRTA, responsibility invariants, LLM evaluation, semantic measurement, cross-model comparison, Jensen-Shannon divergence, TSTT License CC-BY-4.0