Document pipeline
Every filing enters as bytes, then earns its structured read.
Ingest, don't summarize
Every document lands in the corpus exactly as received — a degraded scan, an OCR'd PDF, a native filing. Nothing is condensed or interpreted before it's stored, so the original is always there to walk back to.
Read the image, not the OCR
Figures are pulled from the 300 DPI page image itself, never the garbled text layer OCR produces ($109,307.69 arrives from OCR as $108.307.89). The text layer is kept only as a search hint and to route a page to the right format profile.
Log the coordinate
Every extracted figure is logged against the exact page, table, and cell it came from, so a reader can walk back to the source in one click, and an approximate read keeps its ~ marker.
A dollar figure read straight off the auditor's filing is [verified] ; the buildout window inferred from the site plan's phasing language is [inference] .
A figure with no source citation does not enter the record — not even as inference.