Large language models frequently generate plausible but factually incorrect information (hallucinations), limiting their deployment in high-stakes domains. Existing detection methods rely on expensive LLM-based judges or supervised neural models that lack interpretability. We propose a constraint-based approach for factual question answering that provides explainable, high-recall detection by extracting named entity constraints, filtering by semantic relevance, and verifying presence using length-adaptive thresholds. On the HaluEval QA benchmark, our method achieves 83.7% recall with 54.6% precision at 57.1% accuracy. While a simple entity overlap baseline achieves higher overall accuracy (65.3%), our method provides actionable constraint-level diagnostics showing which specific facts were violated, running 7-15× faster than GPT-4 with zero API costs. Ablation studies show semantic filtering contributes 9.3% accuracy by preventing over-flagging. Our high-recall, explainable approach is suitable for applications where catching hallucinations is prioritized over minimizing false alarms, such as content flagging for human review. Multi-domain evaluation reveals task-specificity: effective for factual QA but not conversational dialogue.
Jyotsna Bulchandani (Thu,) studied this question.