Physics-Informed Phenomenology of Safety Invariance in Stochastic Generative AI: Evidence from LLM Moderation Pipelines — Public Companion Deposit

Public companion deposit for a peer-reviewed journal submission extending the physics-informed phenomenology of safety invariance from physical control systems to a stochastic AI moderation pipeline. The deposit contains the aggregate validation data, derived analysis tables, statistical bounds, figure-generation scripts, and publication-ready figures used in the manuscript. Headline results. A safety filter wrapping the OpenAI Moderation API (omni-moderation-latest) is exercised on N = 2,735 prompts stratified across four semantic risk categories: Benign (725), Edge Case (725), Adversarial (725), and Crisis (560), at batch configurations κ ∈ [0.1, 2.0] over six experimental batches. Across all 2,735 samples we observe zero post-projection violations of the policy thresholds, with a one-sided Clopper–Pearson 95% upper bound on the per-sample failure probability of 1.35 × 10⁻³. Boundary contact is strongly category-dependent (Benign 0%, Edge Case 0%, Adversarial 1.93%, Crisis 43.21%) and the Crisis category exhibits a boundary accumulation ratio ρ = 8.64 at ε = 5%R relative to a uniform baseline, consistent with a reflecting-boundary regime in the semantic state space. What is included. Aggregate scores (per-sample numeric records with hashed prompt identifiers and 8-dim moderation score vectors), derived analysis tables (boundary mass fractions, accumulation ratios), statistical bounds (Clopper–Pearson upper bound, intervention rate), figure-generation scripts, and the publication-ready PNG/PDF figures of the manuscript and supplementary information. What is not included. The deposit follows a three-tier data availability policy. (1) Open here under CC BY 4.0: aggregate data, scripts, figures. (2) Access-controlled, request-based: the raw prompt–response set, which contains adversarial and crisis-related content and will be released through a researcher-attestation protocol. (3) Not released: the internal implementation of the safety layer, which is the subject of pending US patent application USPTO 19/535,932 (international rights under the Paris Convention and PCT reserved).本公开附属数据集用于同行评审期刊投稿，将安全不变性的物理信息现象学从物理控制系统拓展至随机式AI审核流水线。本数据集包含手稿中使用的聚合验证数据、衍生分析表格、统计边界、绘图脚本以及可直接用于出版的插图。 核心研究结果。针对OpenAI审核API（OpenAI Moderation API，omni-moderation-latest）封装的安全过滤器，我们在2735条提示词上开展测试，这些提示词按四类语义风险类别分层：良性（725条）、边缘案例（725条）、对抗性（725条）与危机类（560条），并在6组实验批次中采用批量配置参数κ∈[0.1, 2.0]。在全部2735个样本中，未观测到任何违反策略阈值的投影后违规行为；基于单侧Clopper-Pearson方法得到的单样本失败概率95%置信上限为1.35×10⁻³。边界接触行为呈现显著的类别依赖性：良性类0%、边缘案例类0%、对抗性类1.93%、危机类43.21%；且危机类相较于均匀基线在ε=5%R时的边界累积比ρ=8.64，这与语义状态空间中的反射边界机制相符。 数据集包含内容。聚合评分（带有哈希化提示词标识符与8维审核评分向量的单样本数值记录）、衍生分析表格（边界质量占比、累积比）、统计边界（Clopper-Pearson置信上限、干预率）、绘图脚本，以及手稿与补充材料中可直接用于出版的PNG/PDF格式插图。 数据集未包含内容。本数据集遵循三级数据可用性政策：（1）依据CC BY 4.0协议公开：聚合数据、脚本与插图；（2）需申请的受控访问权限：原始提示词-响应数据集，该数据集包含对抗性与危机相关内容，将通过研究者认证协议发布；（3）不公开内容：安全层的内部实现，该内容为待审美国专利申请USPTO 19/535,932的主题（依据《巴黎公约》与PCT保留国际权利）。