
Chapter 2 of 7
The only magician who reveals every trick.
Every value comes with its receipt: the page it was found on, the exact quote, and who voted for it.
The receipt Example
An illustrative receipt for a small invoice. Every value shows its page, its quote and its status.
| Field | Value | Where it came from | Status |
|---|---|---|---|
| Seller VAT ID | 11 characters | p. 1 · check digit ✓ | confirmed |
| Gross amount | 54.88 EUR | p. 1 · “Total due: 54.88 EUR” | confirmed |
| Shipping | 13.49 EUR | p. 1 · “Courier shipping 13.49” | confirmed |
| Check | net + VAT = gross ✓ | passed | |
| Sale date | — | readings differ: 14 Mar vs 4 Mar | disputed |
For every field
- the status: confirmed, disputed, unreadable or not in the document
- the value, the exact quote and the page number
- what every model read, in each of its two passes
- which models contradicted themselves or disagreed
- whether the arbiter was called in
- the result of check digits and cross-checks
For every document
- the readability of each page
- the document type and the template version used
- which models read it and how many tokens each one used
- the time taken and any errors
The content of your documents and the models' answers never go into system logs. They live only in the receipt inside your own set.
Written down vs. worked out
Some values are read: they sit on the page in plain sight, like a number, a date, an amount or a name. Others are inferred from the content, for example "is the person registered at this address a child?". Both go through the same agreement rule, but two models can agree on the same wrong conclusion, and no check digit can catch that. So the receipt counts inferred fields separately, and you decide how far to trust them.