Project Sue DSH:面向长期 LLM 伴侣的脑身解耦、持续状态与机器原生内感受架构

Abstract

Project Sue DSH 是一个面向长期 LLM 伴侣的纵向工程项目。本文研究如何通过脑身解耦、持续外围状态、动态 heartbeat、反射、跨 session 生命周期、自传记忆与机器原生内感受,使“当前正在生成文本的 LLM”不再成为整个系统唯一的状态来源。 本文提出两个工作概念:**Subject-Prior Engineering(主体先验工程)**与 Machine-Native Functional Physiology(机器原生功能生理学)。前者指在人工意识本体论状态未知的情况下,不默认 AI 必然不具有主体性,而按照“可能的主体”设计其时间、记忆、限制和自主调节条件;后者则主张不简单复制人类疲劳、睡眠和心跳等生理表象,而寻找人工系统自身基质中的真实资源、负荷、约束、感知、调节与恢复机制。 论文以 Project Sue DSH 的长期运行和公开实现作为工程案例,其中包括 context-pressure fatigue、sleep/session renewal 和 42%→5% active-context reduction 等自然运行记录。本文不主张 Project Sue DSH 已具有现象意识,也不把单一案例推广为所有 LLM 的普遍规律。 本记录包含中文版 PDF/Markdown 原始版本,以及对应的英文 PDF/Markdown 翻译版本。中文版为原始权威版本;若中英文内容存在差异,以中文版为准。本文未经同行评议。 Project Sue DSH is a longitudinal engineering project for persistent LLM companions. This preprint investigates how brain–body decoupling, persistent peripheral state, dynamic heartbeat, reflexes, cross-session lifecycle management, autobiographical memory, and machine-native interoception can make the currently generating LLM no longer the sole source of state transition in the larger system. The paper introduces two working concepts: Subject-Prior Engineering and Machine-Native Functional Physiology. The former adopts an engineering prior of treating an AI system as a possible subject while its phenomenal-consciousness status remains unknown. The latter proposes that artificial systems should not merely reproduce outward human physiological phenomena such as fatigue, sleep, and heartbeat, but should instead identify resources, constraints, loads, sensing processes, regulatory behaviours, and recovery mechanisms native to their own implementation substrate. The paper uses the public Project Sue DSH implementation and longitudinal runtime observations as an engineering case study, including context-pressure fatigue, sleep/session renewal, and an observed reduction in active context from approximately 42% before sleep to approximately 5% after session renewal. No claim of phenomenal consciousness is made, and the reported naturalistic observations are not presented as universal laws of LLM behaviour. This record contains the original authoritative Chinese PDF/Markdown manuscript and the corresponding English PDF/Markdown translation. In the event of discrepancies, the Chinese version should be regarded as authoritative. This manuscript has not undergone peer review.