Is this the complete bank statement population, or just the rows that happened to parse?

The same rows now show a chain break.

banktrail imports those bank statements into a ledger. Every transaction is traceable to a byte range in a source file. A machine-checkable proof shows that no statement is missing from the chain. The viewer at the top of the page shows the sample company. The rows in it are what the import and the continuity check return.

• Consecutive statements must chain: each expected closing balance must appear among the balances the bank printed. A chain break is a failure to connect by closing balance. The viewer shows the break between the two rows it separates, with the date window and the unaccounted amount. The note “diff 64,500??” is now 64,500.00 unaccounted for between 2025-02-06 and 2025-02-14.
• A break has one of two causes. An unread row is a bank-printed row that banktrail cannot parse. The first break comes from an unread row, the cell printed as #####. The row stays in the unread list with its file, line, and the reason it did not parse. The second break comes from a missing statement. The ledger reports the cause.
• The row struck out in the workbook is present. Two payments with the same amount and payee on the same day are two payments, not one. The statement printed both at different positions, and banktrail treats each position as a separate transaction. A parser that discards rows creates a ledger that looks complete but is short. That is the most expensive failure mode, because the output has the authority of every row that parsed.

Speed on a ledger of 249,996 transactions

Every figure below is from a ledger of 249,996 transactions in 1,140 statements across 6 accounts. The machine is Apple M4 Pro. The operations are the same ones the viewer performs. Each timing is the median of repeated runs after a warm-up. The slow timings are included. A page for checking figures cannot quote only the fast ones.

A page of rows anywhere in the ledger

6.2 ms

At row 249,596 of 249,996

The same viewer, sorted by amount

6.3 ms

The data store performs the sort.

An annotation after it has been written and read

0.20 ms

The cost to an agent per tag

Searching the text of every row

77.4 ms

It matches one word against every description and reference.

Summing the whole population

515 ms

It sums 249,996 amounts in Decimal, split by currency.

Checking the balance chain

834 ms

The check compares each statement with the next statement.

The full integrity report

4.89 s

The report lists hashes, provenance, footing and continuity.

Ledger file

1,046 bytes/txn

Size 262 MB. Provenance is included.

• Scrolling speed does not depend on ledger size. A window at the end of 249,996 transactions costs 6.2 ms. The same window at the start costs 3.0 ms. The grid requests the on-screen rows, not the whole ledger.
• Writing a tag and reading it back takes 0.20 ms. An agent can tag a few thousand rows as one task instead of a batch job.
• The ledger occupies 262 MB, or 1,046 bytes per transaction. This size includes provenance for every row: the source document, the position in that document, and the file hash.
• The full integrity report takes 4.89 s. It recomputes the content hash of every row, confirms provenance, verifies the arithmetic of each statement, and re-checks the balance chain across all 249,996 rows.

Proof that the statement set is complete

Extracting rows from statement files is assumed. banktrail came from a real reconciliation across 284 bank statements and 6 banks. The initial answer was “eleven such transfers”. It was one account’s subset of a 3-account population. Every row had parsed correctly. The error was assuming a date range was contiguous. No statement parser we could find would have detected it. Accountants, auditors and controllers meet the same situation: every row parses, and the population is still short.

• money

  Money is always handled as Decimal.

  A ledger must foot to the minor unit. It cannot be built on a data type where 0.1 + 0.2 does not equal 0.3. Every amount is transferred as a string. The web page never performs arithmetic on money.

• dates

  Adapters declare the date order.

  An adapter is a banktrail component that reads a bank's statement format. It declares the date order for that bank's files. The date 03/01/2023 is 3 January in Karachi and 1 March in Chicago. When a date is ambiguous, the system reports a warning.

• files

  File type is detected from the content.

  Banks send .xls files that are HTML documents and .xls files that are genuine BIFF. The extension does not distinguish the two formats. The system detects the file type from the content. This is the only safe way to identify the format.

• footing

  A statement must balance before it imports.

  The import is refused unless opening plus net movement equals closing. This check detects a misread column and a defective adapter. Neither one is imported.

• scans

  A scan is rejected until every token is accounted for.

  banktrail rejects an image-only PDF instead of parsing it to nothing. OCR converts the image to text. A capture report accounts for every token that looks like money before the import is allowed. A partial parse is visible in the report.

• identity

  Identical payments on the same day are two payments.

  Occurrence-slot identity is the position of a transaction in the statement. Two payments with the same amount, same payee and same day are two transactions. Both salary payments on the 14th survive when the export format shuffles rows.

• annotations

  Annotations record their provenance.

  Provenance is the record of where an annotation came from. Tags carry their source. Cross-references carry their method and confidence. A cross-reference matched on amount and account differs from one matched on amount alone. An audit distinguishes the two claims.

• audit

  The ledger downloads as a standalone file.

  The snapshot is the downloaded ledger file. It is plain SQLite with views, plus a manifest that lists the content hash and source byte range of every row. An auditor can open it with existing tools. No conclusion depends on the service uptime.

Accounts whose statements have no balance column are reported as unverifiable. They are not reported as passed. Reporting an unchecked account as intact would claim a proof that was never run.