How AI actually reads a G702 / G703

A pay-application package is a hostile document format. The G703 grid varies by GC and by software. Invoices arrive as clean PDFs, phone photos, and fax-quality scans in the same folder. Numbers are printed, handwritten, stamped, and occasionally struck through. Getting reliable structure out of that is where most of the engineering effort goes.

Read it twice, reconcile the difference

The single most effective accuracy technique is embarrassingly simple: read every document twice, independently, and reconcile the two reads field by field. One read comes from a vision model that sees the page layout; a second comes from a different source - the PDF's own text layer when it has one, OCR when it doesn't, or a second model. Where the two reads agree, confidence is high. Where they disagree, the value is flagged for a human instead of being silently trusted. A value that only one source could see never gets treated as certain.

A number that can't be corroborated shouldn't be certified - it should be flagged. Confidence is a first-class output, not an afterthought.

Every value points back to a page

Extraction without provenance is just a faster way to be confidently wrong. Every field a verification engine emits should carry a citation - the exact page, and where possible the cell or line, it came from. That citation is what lets a reviewer click a finding and land on the source in one motion, and it's what makes the output defensible in an audit. If a system can't tell you where a number came from, it can't be trusted to tell you the number.

The 300-page problem

Model context windows and output limits mean you can't just throw a 300-page scanned bundle at a single call. The document has to be planned into its logical constituents - a GC cover sheet, its continuation, each sub's invoices, the waivers - and processed in overlapping windows, with page numbers remapped in code rather than trusted from the model. Do it wrong and a citation opens the wrong page; do it right and the reviewer never notices there were 300 pages at all.