摘要
本文提出 ISA(内部结构化对齐)路径协议的第一版规范。我们论证:当 AI 从回答系统走向行动系统,安全证据必须从"输出是否合规"推进到"行动形成过程是否可被指认、归因、审计与前置约束"。协议以 PAAC 四要件为核心,给出运行时事件模型、约束注册与准入流程,以及第三方可独立复核的审计通道设计。
Abstract
We present the first specification of the ISA (Internal Structured Alignment) Path Protocol. We argue that as AI moves from answering to acting, safety evidence must shift from output compliance to whether the action-formation process is identifiable, attributable, auditable, and pre-emptively constrainable. The protocol centers on the PAAC four elements and defines a runtime event model, constraint registration and admission flow, and an audit channel verifiable by independent third parties.
BibTeX
@techreport{trantorlabs2026isa,
title = {The ISA Path Protocol v1},
author = {TRANTOR LABS Research Group},
year = {2026},
institution = {TRANTOR LABS}
}