Reading Payment Statement

Extracts structured data from Japanese payment statement images (支払調書), the tax documents that report freelance income and withholding. It handles both PDFs and images, trying text extraction first before falling back to OCR. The clever bit is the dual verification mode: it spins up two parallel sub-agents to read the same document independently, then compares results to catch OCR errors. If they disagree, you pick which one looks right. Falls back to single-pass OCR with mandatory human review if sub-agents aren't available. Returns structured data with payer name, payment amount, withholding tax, and category. Useful if you're building Japanese tax prep tools or need to digitize a stack of these forms without manually typing numbers.