Reading documents
What the page reads, the loan file read by code, how pages become documents, and how every value is anchored to the words on the page.
Updated 2026-09-26
Reading documents
Drop files in any combination: the loan file, the borrower's documents and, once the AUS has run, the DU findings or the LPA feedback certificate. A drop can be one document in forty images or several documents in one file. PDF, JPEG, PNG, HEIC, TIFF, XML and ZIP are taken.
The loan file
The MISMO 3.4 export from Arive or EASE is read by code, along the paths of the DU Specification: the loan terms, the product and purpose, the AUS, the application and closing dates, the employers with their start and end dates, the liabilities and the property. The loan's payment, ratios and cash to close are worked out from it. What it states is shown once to confirm and never asked again.
What it reads
| Kind | Documents |
|---|---|
| Wages | Paystubs, W-2s, verification of employment (Form 1005) |
| Returns and business | Form 1040 with Schedules 1, B, C and E; Forms 1065, 1120 and 1120-S with Schedule K-1, Form 8825 and Form 1125-E; Form 1099-DIV; profit and loss statements, balance sheets, business account statements, partnership agreements, business verification |
| Benefits | Social Security award letters, tax status verification |
| Assets | Bank, investment and retirement account statements; gift letters and gift transfers; retirement access and liquidation; settlement statements |
| Liabilities and credit | The credit report, creditor statements, liability analysis |
| The loan | DU findings, the LPA feedback certificate, the application supplement, the loan estimate, lender product quotes, the broker compensation agreement, Form 70 with its rent schedule, title verification |
Forty-eight document types, each with its own field list and its own checks. A DU findings report or LPA feedback certificate in its published layout is read by code in a fraction of a second; a scan or another layout goes to the reader like any other page.
Files it turns away
Every file is scanned for malware first, then typed by its bytes and checked for size. A password-locked file, one carrying scripts or embedded files, an unsafe zip or a duplicate is refused and named with the reason. A file that still cannot be opened is set apart, and the rest of the loan reads on.
Pages into documents
A drop is split into documents by its page metadata, its printed page numbers and the same reader that reads the fields. When the reader is unsure which pages belong together, it asks: "Are these 3 pages one paystub?" Yes, one paystub keeps them. No, fix the pages opens the group: retype the document, move a page, remove one, add one.
Documents are read one at a time, pages in order, each page with the previous one as context. On the installed profile: about 15 seconds for a one-page document, about nine minutes for a packet of 15 documents and 25 pages.
Anchored to the page
Every value the page keeps points at the words it was read from: the page, the words and the quote. A number's page opens in a drawer, the whole page with those words boxed. A value the reader was unsure of is asked: the crop, what was read, Confirm or Correct.
Checks between documents
Documents are compared with one another, field for field and period for period: year-to-date earnings on a paystub against the employer's own statement for the same period; W-2 wages against the year's wage basis, with printed pretax deductions accounted for; tax schedules through their own lines; employer names, pay frequencies and employment dates across every document and the loan file.
A disagreement is a contradiction. Both sources stay with their pages, and the page asks which is right: each reading as a card with its crop, Use this on the right one, or Neither is right to type the value. See Questions.