You'd think a platform built around snapshot PDFs would, you know, keep some PDFs. Nope. STG stores none. When you open a record, the viewer assembles the PDF on the spot from its SQL pieces — fields, OCR text, screenshots, attachments. Same components in, identical PDF out, every time.

The canonical PDF can live in our viewer, the customer's ECM, both, or neither — so when a customer exports is their compliance call, not a technical limit. The PDF is a costume, not the body. The longer story STG's canonical store is Azure SQL: it holds the fields, OCR text, screenshot references, preambles, and layout definitions for every record. When a user opens a record, the viewer assembles the PDF on demand from those components — and after Final Capture the SQL for a Transition is locked and immutable, so the same record reassembles identically forever. Because nothing is ever stored as a PDF, the canonical artifact can live wherever the customer wants: in our viewer (assembled on demand), in their ECM (written once at export), in both, or in neither.

AI grounding runs against the SQL text layer, not the PDF. That single architectural choice — separating operational access from custody — is why export timing is a compliance decision the customer makes, not a technical limit we impose.