Hallucinated Agents: A Cognitive Mechanism for Resolving Ambiguous Threats

Why do people "see" ghosts, monsters, or hidden adversaries in ambiguous shadows or noises? Why do panic attacks erupt from seemingly nowhere? Hallucinated agents under threat are not mere cognitive glitches, they represent an evolved safety mechanism to prevent physiological and cognitive catastrophe. This theoretical working paper argues that when fear, urgency, and ambiguity converge rapidly, the human defensive system cannot afford to linger in an unresolved state. High-arousal ambiguity is metabolically costly, cognitively destabilizing, and physiologically hazardous: sustained catecholamine surges, autonomic dysregulation, reduced heart-rate variability, and—in extremes—arrhythmia or stress-induced cardiomyopathy (Takotsubo) can emerge from prolonged freeze or overload. To enforce bounded time-to-decision under asymmetric survival costs (missing a real threat is far worse than a false positive), the brain deploys a dual-threshold controller architecture embedded in the defensive cascade. When ambiguity persists beyond person-specific thresholds, the system inserts a Minimal Viable Agent (MVA)—a fictional but computationally minimal adversary (with inferred intent, location, trajectory)—that re-anchors cognition to familiar external threat vectors. This reflexive circuit-breaker restores directional policy selection (orient, avoid, flee, inspect), breaking the non-terminating evaluation loop and minimizing dwell in dangerous high-arousal states. "Hallucinated" agency thus functions as an adaptive off-ramp, not a malfunction. For internal ambiguity (e.g., panic loops driven by catastrophic misinterpretation of bodily sensations as signs of imminent catastrophe—heart attack, loss of control, going insane), the model proposes a parallel but distinct primitive: the Internal Minimal Viable Cause (IMVC). Here, the insertion is a benign reinterpretation (e.g., "this is just adrenaline surge" or "this is temporary arousal") that collapses catastrophic interoception, halts escalation, and restores regulatory control. Unlike external MVA insertion (reflexive and pre-deliberative), IMVC resolution often requires more effortful, deliberative re-anchoring—explaining why panic attacks feel more acute and resistant without cognitive intervention. The framework unifies disparate literatures under a single dual-threshold stop rule: freeze/fight neural dynamics (amygdala–PAG–frontal transitions), hyperactive agency detection (HADD/HAD reframed as implementation), error management theory (fail-fast bias), embodied predictive processing (priors favoring agentive/cause completions to minimize prediction error), catastrophic interoception models of panic (Clark, 1986), intolerance of uncertainty, and compensatory control/meaning maintenance. We formalize the controller with state variables for fear intensity, urgency (time compression), and ambiguity level; define dual termination thresholds (τ* for external, λ* for internal); derive a composite risk function R(t) capturing escalating physiological/cognitive costs; and provide full pseudocode for the off-ramp mechanism. Testable predictions span modalities (e.g., auditory accumulation vs. visual commitment in agent attribution), cross-modal congruence, arousal modulation, individual differences (Type M: internal/spontaneous generators with shorter freeze but higher false positives; Type P: external scaffolds with greater reliance on rituals/rules under control threat), and intolerance of uncertainty scaling.Appendix B details VR-ready experimental paradigms for validation: low-light visual occlusion tasks, patterned auditory ambiguity buildup, cross-modal conflict, interoceptive perturbation (e.g., CO₂ challenges or somatic misinterpretation induction), and participant-specific threshold mapping. This working paper advances integrative models of threat cognition, panic maintenance, freeze dynamics, and ambiguity-driven hallucination. It offers a computational, physiological, and evolutionary lens for researchers in cognitive science, psychophysiology, neuroscience, computational psychiatry, VR-based threat simulation, and AI alignment (real-time bounded decision-making under uncertainty).