Stage 3 — Form understanding

Template Studio

A form nobody has seen before, delivered as a 300 dpi scan with no embedded fields. Geometry finds the boxes; a page-image model works out what each box is asking; then every question is bound to a canonical ledger slot. Publish once and every future filing of this layout takes the fast path.

Fields detected from vector + glyph geometry

Auto-mapped bound to existing slots

Need a human no confident slot match

New slots proposed extends the ledger schema

Write-back Overlay no field dictionary exists

Detected field graph — page 1

300 dpi · deskewed 0.4°

Boxes are the geometric detector's output, positioned as a percentage of the page so they survive rescaling. Dashed red boxes are detected but unmapped — the system found a question it cannot answer from the ledger, and says so.

Question → canonical slot

Selected field

Why this screen is the differentiator The common approach to a new form is a new prompt: paste the layout, describe the fields, tune until it works. That produces a system whose cost and risk grow linearly with the number of forms, and where nothing learned about one carrier's APS transfers to the next. Here, understanding a new layout produces a reusable artifact — a Form Spec plus a slot mapping. The 36 auto-mapped fields cost nothing because those clinical facts were already extracted for the FMLA certification. Only the 5 genuinely new questions need a human, and answering them once retires the cost permanently.