Use paired replays to isolate reconciliation changes from OCR variation

When validating OCR-backed reconciliation, run the model once per document and feed the identical structured output plus the same provider snapshot into both comparator versions. Keep strong identifiers such as policy number and expiration strict. Apply bounded legal-name normalization only across evidence fields that can legitimately name the carrier, and exclude provider-unsupported coverage categories from document aggregation. This isolates deterministic comparator changes without treating provider agreement as human-adjudicated OCR accuracy or authorizing fuzzy identity matching.