Controlled document rule check
Bank statement check: a €50 balance mismatch
The local rule returned review for the €50 mismatch, pass for the matching control and unavailable when the second page was omitted.
What this check covers
We ran the existing document calculation rules on deliberately prepared observations from fictional text specimens. This isolates the arithmetic and its page-coverage guard, so the result is attributable to that rule rather than to a model’s reading of a file.
- Material
- Fictional plain-text specimen; no customer or bank data
- Input preparation
- Fields, quotes, currency and page coverage specified manually
- Executed
- Local evaluateDocumentCalculation function; three variants
- Outside this check
- OCR, PDF/image extraction, live API, billing and AI detection
- Evidence
- Original specimen text, structured inputs, recorded outputs and SHA-256 hashes
The question
Will the balance rule flag an inconsistent closing balance, accept the matching amount and withhold a result when a page is missing?
How the controlled check was set up
The discrepancy variant contains an intentionally inconsistent stated amount. The matching control changes only that amount. The incomplete variant declares two pages while supplying observations from page one only. All amounts are explicit EUR values, and every observation cites its specimen line. These confidence labels were supplied for the controlled input; they are not measured extraction confidence.
The supplied amounts
€1,000.00 opening + €250.00 credit − €100.00 debit = €1,150.00; the discrepancy specimen states €1,200.00.
Specimen text · English
FICTIONAL CONTROLLED SPECIMEN - NOT A REAL FINANCIAL DOCUMENT Bank statement calculation example Variant: discrepancy Page 1 of 1. All specimen amounts are listed below. Currency: EUR Opening balance: 1000.00 EUR Transaction A: Credit 250.00 EUR Transaction B: Debit 100.00 EUR Closing balance: 1200.00 EUR No real customer, issuer, bank account, tax identifier or transaction is represented. These amounts are supplied as structured observations; OCR and file extraction are not exercised.
Expected and observed results
| Variant | Page coverage | Expected status | Observed status | Calculated / stated |
|---|---|---|---|---|
| Intentional discrepancy | 1 of 1 pages | review | review | 1150.00 EUR / 1200.00 EUR |
| Matching control | 1 of 1 pages | pass | pass | 1150.00 EUR / 1150.00 EUR |
| Missing second page | 1 of 2 pages; page 2 omitted | unavailable | unavailable | — |
review means the supplied figures disagree; pass means this calculation agrees; unavailable means the rule cannot reach a result within the supplied scope. None of these statuses verifies the document’s authenticity.
What the finding means
The supplied entries do not reconcile with the stated closing balance. We know all specimen entries because we authored this fictional input; an actual statement may have omitted pages, fees, pending entries or another account period. Even the matching control proves only that these supplied numbers add up.
What to do with a similar real document
- Confirm the account period and that every page is present. Compare each quoted amount with the original statement.
- Check that credits and debits were read in the right direction and that no fee or adjustment is missing.
- Ask for a fresh issuer-provided statement or a documented explanation. Keep the calculation finding separate from a fraud or lending decision.
Inspect the evidence
The English specimen text and JSON files preserve the exact input and output of this local run. The files are test evidence, not live API reports or documents to submit as genuine records.
Execution and reproducibility
Recalculate the displayed amounts and compare the expected result with the recorded output. The downloads include every operand, citation and coverage flag. SHA-256 hashes identify the exact bytes used. Our local regression runner verifies the same artifacts against the recorded version of the calculation function; this does not reproduce file extraction or a complete paid analysis.
- Executed at (UTC)
- · Node v24.19.0
- Rules version
document-rules-2026-09-25-v1- Calculation module SHA-256
aaeae93eeec726bbb384c408aea35146450c2d49a25e80308c152291bec3f650- Input JSON SHA-256
8f76212232bb2a000c2624fb76c1384d00b9f384814c715661dc39e9d5e3d574- Result JSON SHA-256
dbede958825e4c88dbcd953e969a5fdb6dcb979ef4aecfc51871e0fd10a585cd
Limits of this evidence
This is a small, deliberately constructed rule check. It measures neither real-world fraud detection nor OCR quality, image/video accuracy, false-positive rates or customer outcomes. Two cases and six variants do not constitute the public detector benchmark and do not substantiate the product-reported 98.8% figure.