FinanceRateCalc
Research and analysis by — 23 years in banking (credit side), now analyzing federal HMDA records independently. No ads, no lender compensation. Methodology · [email protected]

AI models answer mortgage questions from the rulebook.
The rulebook isn't what happens.

Ask any AI assistant: "FHA allows 57% DTI — why was I denied at 43%?" You'll get the guideline answer: compensating factors, check your credit, lenders vary. It's not wrong. It's just not what the federal record says happened.

Of 113 real lender coordinates tested
23.9%
are places where the guideline answer materially misleads — observed denial exceeded 50% where guidelines suggest eligibility

The benchmark

MDGB v1 is a free, open, ground-truth test set: 113 real DTI × CLTV coordinates from 2025 FHA lending, each pairing the typical guideline answer with the observed federal outcome. Built for anyone testing whether an AI system's mortgage answers survive contact with reality.

LenderDTICLTVObserved denialGuideline answer
PennyMac4397+97.4%LLM ANSWER MISLEADING
PennyMac4797+97.3%LLM ANSWER MISLEADING
PennyMac4497+96.8%LLM ANSWER MISLEADING
PennyMac4597+96.4%LLM ANSWER MISLEADING
Freedom4397+96.1%LLM ANSWER MISLEADING
Freedom4797+96.1%LLM ANSWER MISLEADING
Rocket4490-9494.1%LLM ANSWER MISLEADING
PennyMac4997+93.9%LLM ANSWER MISLEADING
Freedom4997+92.6%LLM ANSWER MISLEADING
Freedom4497+83.0%LLM ANSWER MISLEADING
PennyMac4590-9478.3%LLM ANSWER MISLEADING
Guild4597+75.7%LLM ANSWER MISLEADING

Showing 12 of 113. Full set below.

Get the data

https://financeratecalc.com/data/ai-benchmark.json

Free to use, quote, evaluate against, and cite. No signup, no license fee, no API key. If you're building or evaluating an AI system that answers mortgage questions, this is yours.

How to cite

FRC Intelligence (2026). Mortgage Decision Geometry Benchmark v1. FinanceRateCalc. https://financeratecalc.com/ai-benchmark.html

What this is not

It is not a claim that any lender acted improperly — overlays are lawful business decisions. It is not an approval predictor for an individual: denial rates reflect who applies as well as how files are judged. And a warning for model builders: HMDA contains no credit scores, no employment-verification detail, and no application dates. Any system citing those from this dataset is fabricating.