Agentic Data Cleaning Explored Without Clean Reference Data

This research investigates the complexities of data cleaning when a trusted "clean" reference dataset is unavailable, a common challenge where unusual values could be either errors or valid observations. The study proposes an evidence-grounded framework for agentic data cleaning that integrates various capabilities, including structured context, LLM reasoning, executable checks, controlled evidence retrieval, and conservative repair strategies. Seven distinct configurations of this framework were experimentally evaluated across financial, clinical, and environmental datasets, using both synthetic corruption and descriptive analysis of original data. The findings indicate that while a deterministic profiling baseline achieved the highest detection F1-score, no single LLM-based configuration consistently outperformed others across all criteria. The research highlights that adding more capabilities often introduces trade-offs among detection accuracy, repair effectiveness, evidence grounding, conservative behavior, reproducibility, and operational costs, rather than providing universal improvements.