The unstructured half of the fabric. Hospital bills, pharmacy invoices, discharge summaries and motor repair estimates arrive as paper and photographs; this turns one into line items, recomputed totals and pricing flags a claims handler can act on.
Paste the document, load one of the sample images, or upload a scan or phone photo of your own. Extraction runs on the multimodal deployment; the totals, composition and reconciliation check are recomputed in code afterwards. The samples are generated documents (tools/make-samples.py) with deliberate issues planted — a mis-footed total, an upgraded room, unbundled items, a part billed as both repaired and replaced.
This is the front door of the claims pipeline — everything downstream depends on getting the document into structure.
Upcoding, unbundling and inflated consumables are the dominant health-fraud patterns, and they are only visible at line-item level — a claim total never shows them. That is why extraction sits ahead of triage rather than beside it.