SLM Meets LLM: Balancing Latency, Interpretability and Consistency in Hallucination Detection