Personalized Semantic-Trajectory Safety (PST-Safety): Differential Safety Intervention via Definitional and Behavioral Trajectory Estimation

This paper reframes personalized AI safety from a surface-risk filtering problem into a definition-aware, trajectory-sensitive personal-effect estimation problem. Conventional safety systems often focus on prohibited-content categories, refusal policies, surface-risk signals, and single-turn outcomes. These tools are essential, but they do not by themselves determine what a user means, whether an unusual expression is healthy or harmful for that specific user over time, or whether confidence is being inflated by long but autocorrelated interaction histories. The central contribution is Personalized Semantic-Trajectory Safety (PST-Safety), an AI safety architecture and validation protocol that separates definitional estimation, social semantic calibration, safety-trajectory tracking, and differential policy routing. The framework distinguishes Ndef, the evidence count for what the user means, from Neff, the autocorrelation-corrected effective sample size for what that meaning does to the user over time. It routes cases through safe, uncertain, and harmful zones rather than a single binary threshold, while reporting the uncertainty burden created by clarification. PST-Safety is not a clinical, diagnostic, or deployment-ready system. It does not infer mental disorders from text and does not treat trajectory monitoring as acceptable without consent. Current evidence supports architecture- and protocol-level plausibility: analytic results, synthetic simulation, illustrative Claude/GPT feasibility checks, and one conditional external benchmark test on Spiral-Bench. The central empirical bottleneck is structure-proxy validity, which is treated as a falsifiable condition rather than an assumed truth. The accompanying reproducibility package provides analysis scripts, input and output data, figure-generation code, evidence-status labels, and manifest checks for auditing the reported results.