You are building a structured data extraction system using Claude. The system extracts information from unstructured documents, validates the output using JavaScript Object Notation (JSON) schemas, and maintains high accuracy. It must handle edge cases gracefully and integrate with downstream systems.

The system routes documents with extraction confidence below 85% to human review. A quarterly audit reveals that 12% of high-confidence extractions (≥85%) also contain errors – cases where the model finds plausible-but-incorrect values. Error sources vary: comparison tables showing competitor specs, appendices referencing different product variants, and ambiguous phrasing the model misinterprets. You need a sustainable strategy to catch these high-confidence errors and measure whether improvements reduce the error rate over time.

What approach is most effective?