Beyond Probabilistic Safety: A Deterministic Regime-Geometry Theory of AI Collapse, Hallucination, and GA-Time Dynamics

This work proposes a unified geometric foundation for understanding modern AI failures, showing that hallucination, contradiction, drift, adversarial fragility, reward hacking, policy flips, and multimodal collapse arise not from probabilistic noise but from deterministic transitions in admissible geometry. Tracing fifteen years of AI development—from deep learning and representation learning to transformers, reinforcement learning, and multimodal agentic systems—the paper reveals a consistent pattern: systems operate coherently until their underlying admissible region becomes unstable, at which point behavior collapses abruptly into a different structural regime. Building on evolving-domain geometry and inspired by geometric mechanics, the work introduces Geometry-Attached Time (τ̃) and the Time Stack as core mathematical structures explaining why modern AI systems exhibit discontinuous transitions and why probabilistic safety frameworks fail to capture them. This deterministic perspective provides a common mechanism for the most persistent and widely observed failures in large AI systems.