Separate document capture acceptance from comparison completeness

A readable selected record proves capture usability, not that every relevant record agrees with a reference. When one policy spans several coverage types, joining selected identity back to a unique row can reject valid evidence; comparing only that row can hide another coverage's discrepancy. Use an explicit row index for capture, then independently reconcile every supported populated row. Preserve incomplete entered rows as unknown, but exclude blank preprinted form sections and unsupported extensions.

Evaluate prompt and comparison changes against frozen reference snapshots, and visually review changes in both directions. More matches can reflect harmless formatting normalization; more mismatches can reflect newly covered evidence. Neither is automatically an accuracy improvement. Keep policy characters and calendar dates strict; only normalize bounded representation differences whose meaning is established.

Evidence/example — 2026-09-14: local regression tests exercise shared policy identities, matching liability with conflicting cargo, malformed evidence, blank-template/extension prompt boundaries, and bounded insurer normalization with negative cases. Development replay and machine visual review exposed these failure mechanisms, but do not establish human-adjudicated accuracy or generalization. Remaining OCR errors must remain visible rather than be repaired from reference values.