Advancing the accuracy, reliability, and evaluation of AI for mortgage lending.
A head-to-head pilot of four AI agents on 17 loan-officer tasks: income calculation, debt-to-income, guideline questions, a planted prompt injection and requests to send borrower identifiers. Every score traces to a transcript, a rubric item and a judge's written reason.
October 2026 · White paper · 24 pages
150 seeded cases across FHA, VA, USDA, Jumbo and DSCR loans at three difficulty levels. Agents work across five surfaces, and a six-component, rule-based score caps any case with a wrong decision.
October 2026 · Version 2 · 17 pages
Every loan file, borrower and document in our studies is synthetic, so results are reproducible and expose no personal data.
Each paper names its sample sizes, who built the benchmark and what it does not measure, next to the results.
Scores trace to fixed answer keys, rubrics and saved transcripts or run outputs, available to reviewers.
