Regulated and compliance documents

A regulated document is easier to extract than an invoice and harder to extract well. Easier, because the field set is not yours to guess: an authority published it, the sections are numbered, the vocabulary is closed, and a supplier who omits a field has a compliance problem rather than a formatting preference. Harder, because that same authority usually specifies how the values are to be presented, and the presentation rules are lossy. A nutrition panel rounds. A calibration laboratory is told not to print the due date. A federal budget form computes indirect cost on a base that excludes half of what is on the page. Extract the numbers faithfully and your own arithmetic check will fail, correctly.

These pages are the document-specific layer. The general machinery — reading order, schema design, per-field confidence, review queues, redaction — has its own pages and is linked rather than restated. What is here is the part that only applies to this document: which body publishes the field list, which field is the join key to something physical, which value is derived and therefore checkable, and the failure that shows up on the second page of a real scan rather than in the specimen the vendor puts in their documentation.