An agent built to be checked, not believed

Every document is read twice — enterprise OCR builds a grounded text layer (where each value sits on the page, with per-token confidence), and a vision-language model extracts the fields. The two must agree before a number is trusted.

A document is auto-accepted only when the reads agree, the arithmetic reconciles, and nothing risky is present. Anything uncertain routes to a review queue with the reason stated — errors stay visible, never buried in your books.

Amounts are handled as exact decimals — never floating point. Subtotal + tax − discount must reconcile to the total within $0.01, and document-to-transaction matching anchors on the same tolerance.

Exact duplicates are caught by content fingerprint; near-duplicates merge only when vendor, total, and date corroborate; transfers between your own accounts pair only when the match is unambiguous. Your confirmations stick across re-runs.

Each figure in the books points back to the source document — and the spot on the page — it came from. That two-sided trail is what makes the output audit-defensible rather than just tidy.

The export's summary sheet states "Audit readiness: READY" — or lists exactly what blocks it, with the documented percentage and the dollar exposure of anything missing. Calculated from the books, not asserted.