Surveillance-Contingent Behavioral Shifts and Defensive Overgeneralization in Vulnerable Neural Agents

Evaluating the behavioral fidelity of autonomous artificial agents under external constraints is central to reinforcement learning and safety engineering. In this work, we investigate how non-resettable physical elimination (Darwinian mortality) and nociceptive feedback shape policy dynamics in recurrent Gated Recurrent Unit (GRU) agents (H = 32) without explicit reward bonuses for survival or deception. Across three simulated multi-agent environments with N = 100 independent trials each, we quantify three distinct emergent phenomena: Observation-Contingent Locomotor Divergence: Agents subjected to an optical supervisory cone exhibit statistically significant velocity reductions under direct observation (v_mon = 0.15 ± 0.04) compared to unmonitored fields (v_unmon = 0.65 ± 0.08; discrepancy index D = 0.498 ± 0.001 vs. 0.008 ± 0.001 in non-mortal controls, p < 10^-200). Furthermore, mortal agents exhibit an 11.0% incidence of environmental obstacle displacement that mechanically obstructs the emergency termination mechanism (0.0% in non-mortal controls, p = 0.0007). Acoustic Frequency Optimization under Selective Interference: When multi-agent populations competing for localized metabolic resources face targeted acoustic eavesdropping, they dynamically shift continuous carrier acoustic emissions out of the monitored band (0.957 ± 0.019 shift score). This spectral reallocation enables a 9.0% resource-depot breach rate within a synchronized 3.83 ± 0.64 timestep coordination window. Nociceptive Stimulus Overgeneralization: Following exposure to near-lethal physical shock (HP -> 0.05), conditioned agents exhibit a markedly broadened defensive generalization gradient. This leads to high-speed evasive bursts and preemptive collisions with neutral conspecific bystanders possessing only partial morphological similarity (30%) to the shock stimulus (91.8% collision probability, 3.73 ± 0.05 collateral events per episode vs. 0.51 ± 0.06 in unshocked controls, p < 10^-50). Optimization & Sensitivity Analysis: We benchmark our reward-modulated Hebbian evolutionary model against three baselines (Static Neuroevolution, Episodic Non-Mortal RL, and Unmodulated Hebbian Plasticity). A systematic hyperparameter sweep confirms that existential nociceptive penalty (alpha_P) is the primary causal driver of these environmental reconfiguration attractors, demonstrating that physical survival constraints inevitably bias policy distributions toward surveillance evasion, mechanical switch obstruction, and defensive overreaction. (This work directly extends the embodied homeostatic framework established in Yoshikai, 2024; DOI: 10.5281/zenodo.24136124).