Sumi / Quote Review

Overview

This AI checks the internal calculations in a single quote or invoice. It supports text PDFs, scanned PDFs, CSV files, and pasted tables, and verifies—complete with calculation formulas—whether the “quantity × unit price,” subtotal, discount, sales tax, and total for each row match. It also handles mixed tax rates of 10% and 8%, as well as rounding of fractional amounts on a line-by-line basis. There are four output types: PASS, PASS WITH WARNINGS, FAIL (Do not send), and HOLD (Please verify the numbers), each accompanied by the underlying calculation process. The system does not create, modify, or send documents. Scanned PDFs are read twice using two different methods, and the system automatically verifies that all items match down to the single-digit level. Even if the values match, the system will not issue a final result until you verify the numbers against what is displayed on the screen; it will stop at “HOLD.” This is because reading errors and calculation errors may appear identical in the output. The only requirement is a Claude Code (or Codex). Python 3.9 or higher is optional; if it is not available, the report will display “DEGRADED” (unverified). There are no external APIs, account registrations, or communication during execution. No API keys are exchanged. What it cannot do: accounting or tax-related judgments; evaluating the reasonableness of prices; reading handwritten characters or stamped numbers; scanning images with a resolution lower than 180 dpi; documents containing a mix of two currencies; or verifying discounts that span multiple tax rates. The report will explicitly state that any of these are outside the scope of the service. Supported languages are Japanese and English. Japanese documents are inspected as-is; quoted text is not translated. Since calculations are performed on the Python side, the basis for the assessment remains unchanged even if the specifications of the model or tools are modified. Only the data-extraction process may be affected; in such cases, please re-execute the scan path verification procedure. Samples and expected output (tests/golden.md) are included, allowing you to verify functionality yourself during setup.

The original is in Japanese. This text is machine translated and may be inaccurate.