Homeostasis of the Human, LLM, and AWM: A Process Architecture for Human-AI Collaboration

Large language models provide broad informational and generative potential, but they do not independently supply direction, responsibility, continuity of process, validation of results, or contact with real-world consequences. Human participants provide goals, experience, meaning, responsibility, and action in reality, while remaining limited by time, attention, and processing capacity. This working paper presents AWM as a process-regulation method that links the human and the LLM while preserving direction, trace, epistemic status, feedback from reality, and the possibility of local correction. The paper develops a homeostatic architecture expressed as Human → (AWM + LLM) → Human → R_d → Human. It treats error as feedback, distinguishes verified, indicated, assumed, and unknown claims, maintains parallel traces, protects human freedom and functional hierarchy, and proposes Enough+ as an operational threshold. The resulting model frames human-AI collaboration as an adaptive, verifiable, and robust process rather than a sequence of isolated answers.