From Sample Safety to Path Legitimacy: T0 Admission Gating for Unauthorized Reasoning Transitions in LLM-Integrated Systems從樣本安全到路徑合法性：LLM 整合系統中未授權推理轉移的 T0 准入閘門

The paper came from the system. The system did not come from the paper.

HF2007 was written at a specific point in the development and implementation of my system. I extracted the smallest layer that could be reviewed and reproduced externally without disclosing the private core: T0 admission, an auditable eight-relation representation, a fixed 51-transition experiment, source provenance, hashes and deterministic results.

The formal review and acceptance preserve evidence that path legitimacy had already moved beyond assertion. The mechanism could be compiled, executed, audited, replayed and tested before an unauthorized transition entered ordinary reasoning.

My position is not that AI safety might optionally use this framework. When models, tools, external information and action form a reasoning path, sample-level filtering alone cannot close the safety problem. The path must establish its subject, evidence, causality, boundary, authority and responsibility before it gains the right to act.