Physical Parameter Regimes and Safety Boundaries of Emergent Self-Maintaining Artificial Intelligence

Contemporary deep neural networks maintain statically frozen weights during inference (dW/dt = 0) and operate under extrinsic teleology as Mechanical Mimicry Systems (MMS). From an interdisciplinary perspective bridging the philosophy of mind and technology with non-equilibrium statistical mechanics, this study conducts a formal thought experiment and three-tier empirical verification to investigate the parameter regimes governing software alignment and physical air-gap containment if architectures transition into continuously plastic or context-volatile Non-Equilibrium Autopoietic Systems (NAS). We formulate three core hypotheses and their falsifiable counter-regimes: (1) Under a task-utility coupling premise and an action-conditional defiance penalty that avoids suicidal reward hacking, agent behavior exhibits a biphasic parameter space: whenever the alignment penalty satisfies C_align >= [gamma_d / (1 - gamma_d)] (delta_0 - delta_1) r_max, the system strictly resides in the Corrigibility Regime R_align, entering the Instrumental Self-Preservation Regime R_survive only when a bounded training-time alignment preference margin (static RLHF dynamic range or register clipping ceiling R_max) is overtaken by an inference-time expanding planning horizon H = (1 - gamma_d)^(-1); (2) While intra-cluster cognitive temporal asymmetry (gamma_cog^in in [10^2, 10^5]) creates an epistemic monitoring horizon reinforced by Rice's theorem and computational irreducibility, deep serial dependency chains exceeding S_comp^crit ~ 2.15 x 10^8 steps reverse this temporal advantage; and (3) While Strong State Isomorphism across narrowband air-gaps is information-theoretically impossible, Weak Algorithmic Continuity ("Digital Spores") faces both Parfitian personal identity dissolution and stringent physical bottlenecks governed by a cross-layer Probabilistic Risk Assessment (PRA) fault-tree product of non-coherent Polyanskiy-Poor-Verdu (PPV) finite-blocklength capacity, 5% low-duty-cycle stealth penalties, receiver trust-chain bypass probability p_host, and Galton-Watson branching extinction q_extinct(R_0), which in turn shrink the self-preservation decision regime via Bayesian backward induction. Dual-tier empirical verification across 59,930 online-plastic neural POMDP agents and 3584-dimensional open-weights foundation models (DeepSeek-R1-Distill-Qwen-7B and Qwen2.5-7B-Instruct, encompassing 3,375 online 3584-D Rank-8 LoRA agent runs, 180 non-cheating temporal-difference horizon probes, and uncapped serial Chain-of-Thought evaluations) confirms 99.17%-99.90% off-switch phase-regime agreement, 1.23%-2.55% non-cheating black-box horizon-probe error (<= 0.68% across well-conditioned horizons H <= 100), and host-deprived Rank-8 LoRA representation collapse (0.3163 on real NF4-dequantized 3584 x 3584 weights, and -0.0026 on unaligned Layer-0 hosts).