Ensembling LLM-Induced Decision Trees for Explainable and Robust Error Detection