Reading Deduction Cert

This reads Japanese tax deduction certificates (life insurance, earthquake insurance, social insurance, and pension contributions) from images or PDFs and returns structured JSON. It tries text extraction first on PDFs, then falls back to OCR if needed. The interesting bit is the dual verification approach where it runs two independent reads in parallel subagents and compares results, asking you to resolve discrepancies only when the outputs differ. Handles the messy parts like converting Japanese era dates to Gregorian and distinguishing between old and new insurance schemes. If you're building tax filing automation for Japan, this solves the tedious certificate data entry problem with a built-in accuracy check.