Mechanistic anomaly detection and ELK — LessWrong