When your Priorbank statement arrives as a scan or a photo of the page rather than a clean export, the numbers on it are just pixels until they are extracted. Upload a scanned or photographed statement the same way you would upload a digitally exported file. The engine applies optical character recognition to read pixel-based text, then reconstructs the transaction table from what it finds. That reconstruction is measured against the statement's own logic: if a 3 is misread as an 8 or a table line splits in the wrong place, the numbers further down will not stay in agreement, and the row is flagged. You do not need to pre-sort clean originals from imperfect ones; the same pipeline handles both and only surfaces what needs a second look.
Whatever the running balance looks like on your Priorbank statement, Convert·Into checks it against the statement's own printed closing figure before you export anything. Every statement the engine processes is checked against its printed closing balance. As each row is read, the engine computes what the running total should be and compares it to the figure shown on the statement, all the way to the final line. If the two stay in agreement from start to finish, the extraction is confirmed. When a discrepancy appears, the row is marked for review instead of stopping the job, so a single suspect line never holds up an otherwise clean file. That check runs whether the source was a crisp digital export or a photographed paper copy: the same arithmetic decides what is reliable and what needs your attention.