LLM EKG is an open-source tool for monitoring the behavioral health of Large Language Models in production. It treats LLM outputs as numerical time series and applies signal processing techniques to detect quality degradation, hallucination patterns, and behavioral drift — using pure mathematics, not NLP. The system extracts 16 domain-agnostic numerical features per response, feeds them into a proprietary behavioral state engine, and produces anomaly scores, drift measurements, and multi-scale persistence analysis. All diagnostics are self-referential: zero hardcoded thresholds. Requires only numpy and matplotlib.