Line-item parsing, checked against the bank.

Pathway first reads every statement in the deal together and resolves the business, its owners, and each distinct bank account. Each account gets a fixed ID, and extraction can only assign transactions to those IDs.

That context is used later. A transfer to ****4521 is tagged internal because ****4521 belongs to the business. A Zelle payment to a name on the application is tagged as an owner transaction.

One extraction agent runs per document, all at the same time. Amounts are stored as positive numbers with a separate credit or debit direction, which removes the sign errors vision models make when credits and debits share a column. Duplicate and overlapping statements are removed before extraction.

The starting balance plus every signed transaction has to equal the printed ending balance within $0.05. The check runs per account per statement, so a transaction placed in the wrong account shows up as two ledgers off by the same amount in opposite directions.

A ledger that fails goes to a correction agent with the source PDF, the extracted transactions, and the size and direction of the gap. It can flip a transaction’s direction, remove a row, or add one it finds in the document. It can’t edit descriptions or invent amounts. After three attempts, or when the bank’s own figures don’t add up, the ledger moves forward flagged with an explanation.

In a typical deal, about 800 transactions compress into about 120 description groups. Three classifiers tag them in parallel: deterministic rules for checks, wires, P2P, stop payments, NSFs, and overdrafts; a core model that knows the business’s accounts and owners; and a loan model with your funder registry in its prompt.

True revenue, debt-to-income, and every other metric are computed from the tags. Retag a group of transactions or assign them to a position, and every number, sheet, and screening verdict updates without a reparse.

Each PDF’s structure is checked in parallel with tagging: how many times it was saved, creator and producer mismatches, creation and modification timestamps, and the fonts used. The signals are weighed together with the reconciliation result, because a statement that balances but was edited carries a different risk from one that fails and shows edits.

The tampering score, from 0 to 7, is available to screening rules.

Each document type has its own parser, and every result lands in the same book.