Confidently Wrong: Optical Depth Dynamics as a Universal Thermodynamic Framework for Neural Network Robustness Monitoring Across Vision, Language, and Control

Modern neural networks achieve remarkable performance on curated benchmarks yet fail unpredictably under distributional shifts. Traditional approaches rely on output probabilities (softmax/log-probs) for error detection—a paradigm established by Hendrycks & Gimpel (ICLR 2017). However, this "static" approach fails catastrophically for RLHF-tuned LLMs and PPO-trained agents that have learned to be confidently wrong: perplexity remains low even during hallucinations. We propose a paradigm shift from static output monitoring to kinetic internal physics. We introduce Semantic Velocity, measuring the "tremor" in a model's hidden state trajectory—a leading indicator of failure visible before output metrics degrade. We present Optical Depth Dynamics (ODD), a universal thermodynamic framework unifying robustness monitoring across Vision, Language, and Control: Vision: Spatial depth analysis reveals failure signatures like Global Collapse (ViT) and Avalanche Effect (CNN). Language: Temporal depth analysis detects "Semantic Tremor"—factual generation exhibits laminar flow, while hallucination shows turbulent flow detectable 7–8 tokens before completion. Control: Phase Portrait analysis in RL agents identifies a "Panic Zone" where high velocity precedes reward collapse (p < 0.001). Our open-source tool DeepDrift (pip install deepdrift) provides neural network MRI with