Align absence semantics before scoring OCR extractors

A low aggregate OCR-extraction score can primarily reflect a contract mismatch rather than recognition failure. Before tuning models, align each field’s absence semantics: if the reference says a readable card without a marker is false but the candidate says every absent field is null, boolean fields will dominate the penalty and completeness can collapse. Also avoid calling agreement with production-generated labels accuracy, especially when the cohort was selected using those same predictions. Decompose recognition, association, normalization, and decision metrics before choosing a remedy.

Applies when: evaluating structured document extraction against production-derived labels. Keywords: OCR extraction evaluation null false silver labels selection bias