What we tested in ClearMatch /
The tests behind ClearMatch: inputs, actions, observed results and the limits of that evidence.
Match invoices against purchase orders and received quantities before an owner approves the exact revision.
Internal local testing, using synthetic inputs and owned assets. The results below establish the observed software behavior in those scenarios. They do not establish customer ROI, uptime or external deployment readiness.
An exact matched fixture
Eleven ClearMatch HTTP checks passed. The primary fixture reconciled three units at 1001 minor units each,1800 tax basis points and a declared total of3544. With matching purchase-order prices and receipts, the result contained no exceptions and owner approval succeeded. The test independently calculated the expected subtotal, tax and total rather than copying a displayed answer.
Exceptions remain actionable
Reducing the received quantity from three to one produced RECEIPT_SHORTFALL and blocked approval. Changing the declared total from 3544 to 3543 produced TOTAL_MISMATCH. Reusing the same supplier invoice could not become a second ordinary payable. These observations verify the defined scenarios, not every form of duplicate invoice or disagreement a real supplier could introduce.
Extraction preserves an unknown
One authorized synthetic invoice extraction completed through Gemini 3.1 Flash-Lite. It retained the invoice reference, INR currency, quantity 3, price 1001, tax 1800 and declared total 3544. The supplier identity, absent from the text, remained null with a clarification question. A verbatim source quotation was checked against the fixture instead of accepting unsupported generated evidence.
Recorded boundaries of this evaluation
Consent false blocked the model request and reloading its saved run returned identical output and execution evidence. Session, CSRF and cross-workspace access checks passed in the HTTP suite. This report does not claim an exhaustive browser workflow, accounting integration, production restoration test or general document accuracy measurement. The tested documents contain synthetic, nonconfidential information.
What this evidence does not establish
This matching model supports unique SKUs within each document, whole quantities, two-decimal INR/USD/EUR/GBP and one invoice-level tax input.
It checks submitted facts; it does not authenticate a supplier, prove delivery or determine that a price change is commercially authorized.
Duplicate protection uses the provided supplier identifier and invoice reference. It is not fuzzy duplicate detection across arbitrary names, scans or external ledgers.
An approved reconciliation is a local review record. No bank transfer, payment-provider operation or accounting-system post was executed.
The evidence listed here is primarily HTTP and one synthetic model extraction; it must not be presented as a completed full browser or production rollout audit.
A similar problem in your business?
Bring a real workflow, representative inputs and the result that needs to be reliable.
Shape a project brief