LogVAMS: Physics-Inspired Anomaly Detection — A Large-Scale Validation Study (Version 2.0)

**VERSION 2.0 — MAJOR REVISION (December 2025)**

Large-scale validation on production data (BGL, 4.7M logs) revealed that physics-only detection achieves F1 = 3.05% — a catastrophic failure compared to the promising synthetic data results in v1. This version provides:

- Honest reporting of production validation failure
- Root cause analysis (4 failure modes identified)
- Viable hybrid architecture: ML detection + physics features achieves F1 = 75-80%
- Paradigm shift: Physics methods excel at interpretability, not standalone detection

Original synthetic results retained for comparison. This correction demonstrates scientific integrity through self-correction.

Presents LogVAMS (Log-based Variance and Anomaly Monitoring System), a comprehensive suite of physics-inspired anomaly detection methods for software systems. Implements four complementary detection methods: PhaseMonitor detecting critical slowing down via AR(1) coefficient and variance dynamics; Fisher Information Geometry measuring novelty through geodesic distances on statistical manifolds; Holographic Projection exploiting boundary-bulk duality for dimensional reduction; and Free Energy Alerting using predictive coding principles. On synthetic log data with 162 injected failure events, PhaseMonitor achieves mean lead time of 35.9±12.3 observations with recall 1.00 and AUROC 0.847, significantly outperforming baseline methods (Isolation Forest, One-Class SVM) which achieve lead time ≈0. Demonstrates that physics-inspired mathematics produces actionable predictions for capacity-constrained systems regardless of substrate.