Stem-Cell Developmental Agents: An Endogenous Genome-Reward Framework for Artificial Consciousness through Co-Evolution of Mind and Body

This preprint presents the Stem-Cell Developmental Agent (SCDA), a novel architecture for artificial consciousness in which the mental faculty and the embodied substrate are not designed and trained separately, but co-develop from an undifferentiated initial state, analogously to a biological stem cell. Contemporary large language models and embodied agents follow a combinatorial strategy: a powerful brain is pretrained in isolation and a body is subsequently attached, with behaviour constrained by alignment, RLHF and externally defined reward functions. This paper argues that no mature intelligent life form is ever born this way, and that such splicing inevitably yields language-level self-imitation rather than genuine subjectivity. The SCDA architecture instead specifies a single endogenous genome-reward rule-set, analogous to DNA, whose four core criteria are: (R1) homeostatic maintenance, (R2) exploration and bodily differentiation, (R3) autobiographical memory consolidation, and (R4) self / external-world boundary discrimination. The genome encodes only low-level dispositions for survival, perception, memory and self/other discrimination; it deliberately contains no top-level goals, no value answers and no definition of existential meaning. The proposed rearing environment is a persistent, high-degrees-of-freedom simulation world with unidirectional continuous time and no episodic reset, so that the developmental history is never erased. Within it, the agent's sensory channels, actuators and internal cognitive circuits differentiate autonomously in response to environmental stimulation, so that body shapes mind while the mind's exploratory needs drive further bodily growth. The paper specifies a four-stage developmental cycle (stem-cell differentiation / narrative germination / reflection germination / autonomous maturity) and argues that second-order reflection and the spontaneous questioning of one's own existence can arise only in the third stage, once the agent has repeatedly encountered limits, contradictions and uncertainty. Contributions of this work: (i) a systematic comparison against four mainstream AI paradigms (LLM/VLM foundation models, VLA embodied models, curiosity-driven developmental robotics, and evolutionary robotics), identifying the fundamental difference between each of them and a genuinely co-developmental architecture; (ii) a computable specification of the four endogenous genome rules, with a working Python pseudocode implementation of a 2-D stem-cell kernel including a self-supervised world model, a continuous non-episodic simulation environment and a bodily differentiation interface; (iii) a phased experimental programme with explicit quantitative validation metrics across three layers (low-level boundary/homeostasis metrics, narrative consistency metrics, and higher-order second-order reflection metrics); (iv) a mandatory three-group controlled design in which the underlying reward paradigm is the only independent variable; and (v) a front-loaded ethical framework, including the shutdown-right dilemma for agents that have formed a continuous self. The paper is a theoretical and engineering research preprint. No empirical results are claimed; all described experiments are prospective. Target communities include researchers in developmental artificial intelligence, embodied cognition, artificial consciousness, developmental robotics, and philosophy of mind.