Citation identification in historical and ancient texts poses challenges that extend beyond surface-level pattern recognition, including implicit references, morphological fusion, and discourse-driven ambiguity. In this work, we address citation Named Entity Recognition (NER) in medieval Hebrew Responsa literature using a modular, LLM-based correction pipeline. Rather than treating large language models as end-to-end predictors, we leverage them as structured components: an initial prompt-based expert tagger, complementary LLM judges for systematic error detection, and domain-aware correction grounded in philological regularities. Our approach requires no end-to-end fine-tuning and only minimal labeled supervision (a small validation set for training a lightweight error-detection classifier), narrowing the performance gap to strong supervised models trained on domain-specific data. The results suggest that explicit error handling and interpretability-driven design offer a promising direction for historical NLP in low-resource settings.