An accuracy benchmark you can re-run yourself

The method, completely

• The corpus is public. Court exhibits, charity-register filings, council meeting packets, regulators' teaching material, and banks' own published specimen statements. No customer documents, ever. Each row below names its source; every document can be re-fetched from it.
• The check is arithmetic, not opinion. A conversion passes on a document when the opening balance plus every extracted transaction equals the closing balance the bank printed — row by row where the statement prints a running balance, at period level where it doesn't. Where the document prints its own transaction counts, those are checked too.
• Vendor outputs go in verbatim. A tool is measured on its raw CSV or Excel export of these same documents — no cleanup, no retries, no cherry-picking. Results are published whether the tool passes or fails, and that applies to RowSure's own column first.
• When our own reading of a document doesn't reconcile, a disagreement with another tool's output is reported as uncertain — we can't tell which of the two readings is wrong, so we say exactly that instead of deciding it.
• No adjectives. Verdicts are four words: passes, fails, uncertain, unreadable. The numbers carry the judgement.

RowSure on this corpus: 24 of 28 certified — and 4 refusals, listed first

The four documents we did not certify are the interesting ones, so they come first. All four are documents whose pages don't carry the numbers a certification needs — and delivering a "verified" spreadsheet from them would have been the lie this product exists to prevent:

And the 24 that certified, with how deep the proof goes on each:

"Row-level" means every extracted amount was proven against the running balance the bank printed beside it. "Aggregate" means the period equation closed but the document prints no per-row balance to check against — a shallower proof, and we label it as such. The four levels, explained → Total engine cost for the whole corpus: $0.58.

Other converters, measured so far — wins published first

Raw exports collected on 18–19 August 2026 from each tool's own free tier, one document at a time, compared verbatim by the same arithmetic. Nothing here is estimated: every number below can be re-derived from the tool's own export and the public document. Coverage is partial — free tiers are small — and grows as more exports are run.

CapyParse — 2 documents, 2 pass

Both wins are published exactly because the rule says so — results go up whether they beat us or not.

StatementPro.ai — 2 documents, 1 passes, 1 fails

Good-faith note: this tool's export headers weren't readable by our comparer at first. We extended the comparer rather than mark the files unreadable — the comparison measures their extraction, not our parser.

bankstatementconverter.com — 3 documents, 1 passes, 2 fail

Not yet measurable

• DocuClipper — the free-trial export is capped at 10 transactions per download (the cap is stated inside the export file itself), and we don't publish partial exports as anyone's result. The column arrives when a full export is run on a paid plan.
• Statement Extract and others — no export produced yet.

A vendor column appears on this page only when the raw export actually exists — run by us, by the vendor, or by any user. Nothing is estimated, and results go up whether they beat us or not.

• If you sell a converter: run the corpus, send the raw exports to hello@rowsure.com, and your column goes up unedited — wins included. If you believe a row above is wrong, the correction gets the same prominence as the claim.
• If you use one: put any of these documents through it and send us the export. Or check it yourself in one minute, without us: upload the PDF and the export and the same arithmetic runs on the spot, free.

Reproduce it

Every document above names its public source and can be fetched from it. The check needs no software of ours: add up the transactions in any spreadsheet and compare against the printed balances. If you want our exact corpus bundle and comparison script, write to hello@rowsure.com and we'll send it — and if you find an error on this page, we'll publish the correction with the same prominence as the claim.