Entity Reconciliation (ER) in Cultural Heritage often faces the problem of fragmented collections and inconsistent naming conventions. This paper presents a workflow for evaluating ER strategies, focusing on the complex case of early modern Polish personal names. Leveraging Wikidata’s vast network of multilingual aliases, we construct a validation dataset of nearly 10,000 label–alias pairs. To address the specific variation of multi-part historical names, we introduce a novel component-wise coverage threshold strategy. Under this framework, we benchmark classical string metrics, such as Levenshtein and Jaro-Winkler, and phonetic algorithms like Beider-Morse and Daitch-Mokotoff. Our results demonstrate that supervised ensembles, particularly Random Forest and Gradient-Boosted Decision Trees, consistently outperform individual metrics. Comparison against a state-of-the-art multilingual Transformer (LaBSE) reveals that while the latter maximizes recall, it struggles with precision under limited data regimes and demands high computational resources. In contrast, tree-based classifiers offer a highly efficient alternative that remains transparent, explainable, and effective even when trained on small data samples. We conclude by discussing Wikidata’s potential for scalable ER benchmarks.