Epistemological Governance of External Memory in a Longitudinal Human-AI Sociotechnical System: Architecture, Evidence, and Open Challenges from the Dufour 24 Project

This paper presents the architecture and formalized contributions of Dufour 24 — Derrotero Técnico, a longitudinal human-AI sociotechnical system developed over multiple AI succession cycles for the technical reconstruction of a classic sailboat. The system emerged not from theoretical design but from solving real problems of continuity, authority, memory loss, and error management in sustained human-AI collaboration. We identify and formalize three original contributions: (1) epistemological governance of external memory, a regime that determines not only where information resides but what its identity, competence, temporal cut, authority, and permitted operations are; (2) a quadruple separation of capability, authorization, autonomy, and constitutive authority as independent dimensions governing AI conduct; and (3) a multidimensional state model with fourteen independent dimensions for document and object lifecycle management. We present internally accredited evidence from successive AI cycles, including succession protocols, pre- and post-commissioning tests, and preserved failure events. The system is classified as an advanced field prototype with demonstrated internal capabilities in continuity, memory governance, and error containment, pending comparative validation, independent replication, and sustainability assessment. We discuss convergences with recent work on governed memory for LLM agents, institutional AI reasoning, autonomy level frameworks, and meaningful human control, and we identify the boundaries of what can and cannot yet be claimed.