Accuracy

Each certified layout is benchmarked through the complete production pipeline. A seeded generator produces statement models with known-true rows. Those models are rendered as real PDF files with a text layer, exactly the way the adapters will meet them. The same extraction, parsing, and reconciliation code that runs in your browser then reads the bytes back, and every recovered row is compared with the truth.

The corpus covers single and multi-page statements up to 250 transactions, month and year boundaries, leap days, overdrawn balances, zero balances, wrapped multi-line descriptions, and same-day rows without repeated dates. On top of the clean corpus, corrupted variants are generated with a running balance shifted by one penny, a row silently removed, a debit swapped into the credit column, and an unreadable date. A corrupted fixture that comes out labelled verified counts as a false verified, and the release gate requires that number to stay at zero.

Fixture provenance: the corpus is currently synthetic, built to replicate each bank’s digital statement structure. Synthetic fixtures exercise the arithmetic and layout logic thoroughly, but they are not scans of real statements, and we say so here rather than imply otherwise. As authorised, sanitised real-world fixtures are contributed, they will be added to the corpus and these numbers will be regenerated.

Thresholds for enabling a layout: at least 98% exact row recall, 99.5% date and amount accuracy, 99.5% direction accuracy, zero false verified results, at least 85% of fixtures completing without manual repair, and a median under five seconds for a 25-page statement.