The AI reads. Code decides.

Inspect an invoice.

Pick one of five invoices and replay how it went through every check. This is a replay of saved evidence, not a live run: each step shows output the code really produced (what the reader found in the file, the AI's signed answer, the contract's own reason codes), and the page looks up the live receipt on NEAR testnet where one exists.

The inspection bench

Watch one invoice go through every check.

Step 0 of 6
    Nothing here is made up for the page

    Where every value comes from.

    The invoices are fictional (Qingmar Steel Co. does not exist) and the money is test USDC. The outputs are not: each one was produced by the code in the repository and saved, and the links let you check it. To run it yourself, use the commands in the README.

    • The scanThe reader's own result on the PDFs in reader/corpus/: fingerprint, hidden text dropped with its reason, the verdict before any AI call. Re-run it with tools/read_invoice.py --dry-run.
    • The AI's answerThe live reading of the honest invoice on NEAR AI Cloud, saved whole in docs/proof/reading-INV-2208.json: request, response, signature and attestation report. It re-checks offline.
    • The contract's checksThe reason codes are the contract's own (src/inspection.rs). All four contract verdicts are live receipts on guard.cleartopay.testnet: the honest invoice was CLEARED and paid 7.80 USDC on 2 Nov 2026, minutes after the supplier's 72-hour wait ended.
    • What happens nextThe follow-up text is what tools/next_step.py printed for the live receipts: facts from the signed records, a draft for a person to read, or "no draft" for a seizure.
    The AI reads. Code decides.

    Every receipt on the ledger · Source on GitHub · @cleartopay