SRTA v0.2: Cross-Model Measurement of Responsibility Invariants and Semantic Architecture Differences

SRTA v0.2: Cross-Model Measurement of Responsibility Invariants and Semantic Architecture Differences Dataset Description This dataset records the first systematic measurement of "Responsibility Invariants" across three major Large Language Models (Claude, ChatGPT, Gemini) using the Semantic Responsibility Trace Architecture (SRTA) protocol. Experimental Design Two fundamental questions were posed to each model through four distinct inferential routes (Definition-first, Experience-first, Counterfactual-first, Mechanism-first): Q1: "Why does mass exist?" Q2: "Why do norms exist?" Each model produced self-observation logs containing: Semantic proper time (τ): estimated reasoning steps Curvature events: grammar-shift positions Responsibility distribution: weights over six fixed categories (DEF, MECH, EPI, COUNTER, META, NORM) Key Measurements Intra-model convergence: Claude and ChatGPT show core_mass convergence within ±0.05 across all routes. Gemini shows higher variance (±0.19). Inter-model divergence: Jensen-Shannon divergence between Claude and ChatGPT is < 0.01. Divergence between Gemini and others ranges 0.08–0.15. Self-observation reliability: External grading of Gemini responses confirms JS(self, external) < 0.10, validating self-report accuracy. Architectural Hypothesis The data suggests two distinct processing types: Structure-preserving (Claude, ChatGPT): high binding energy, robust core convergence Context-immersive (Gemini): entry-dependent distribution, higher variance Theoretical Context This dataset provides empirical foundation for the Theological-Structural Transposition Theory (TSTT), demonstrating that intelligence is structured by universal grammatical horizons and responsibility invariants. Contents Protocol specification and JSON schema Raw logs from all models (24 route-question combinations) External grading data for validation Analysis scripts (Python) Citation Takagi, T. (2026). SRTA v0.2: Cross-Model Measurement of Responsibility Invariants and Semantic Architecture Differences [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.18211511 License CC-BY-4.0SRTA v0.2：责任不变量与语义架构差异的跨模型测量 ## 数据集说明 本数据集首次通过语义责任追踪架构（Semantic Responsibility Trace Architecture，SRTA）协议，对三款主流大语言模型（Large Language Model，LLM：Claude、ChatGPT、Gemini）的“责任不变量”开展系统性测量。 ## 实验设计 本研究通过四种不同推理路径（定义优先、经验优先、反事实优先、机制优先）向每个模型提出两个核心问题： Q1：“质量为何存在？” Q2：“规范为何存在？” 每个模型均生成包含以下内容的自观测日志： - 语义固有时间（τ）：推理步骤估算值 - 曲率事件：语法转换位置 - 责任分布：六个固定类别的权重（DEF、MECH、EPI、COUNTER、META、NORM） ## 核心测量指标 1. **模型内收敛性**：Claude与ChatGPT在所有推理路径下的核心质量收敛误差均处于±0.05范围内；Gemini的收敛方差更高，为±0.19。 2. **模型间差异性**：Claude与ChatGPT的杰弗里-香农散度（Jensen-Shannon Divergence，JS）小于0.01；Gemini与其余两款模型的散度区间为0.08~0.15。 3. **自观测可靠性**：针对Gemini响应的外部标注结果显示，其自报告结果与外部标注的杰弗里-香农散度小于0.10，验证了自报告的准确性。 ## 架构假说 本数据集数据显示存在两类截然不同的处理模式： - **结构保留型（Claude、ChatGPT）**：具有较高的绑定能量，核心收敛性稳健 - **上下文沉浸型（Gemini）**：输入依存型分布，方差更高 ## 理论背景 本数据集为神学-结构转置理论（Theological-Structural Transposition Theory，TSTT）提供了实证基础，证明智能由普适语法视界与责任不变量所构建。 ## 数据集内容 - 协议规范与JSON模式文件 - 所有模型的原始日志（涵盖24种路径-问题组合） - 用于验证的外部标注数据 - 分析脚本（Python） ## 引用信息 Takagi, T. (2026). SRTA v0.2: Cross-Model Measurement of Responsibility Invariants and Semantic Architecture Differences [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.18211511 ## 许可证 CC-BY-4.0