AI models are accurate numerically but not contextually. To solve this I have developed a verification system. All numbers used have been checked and provided with a receipt. All data comes from the entire 2025 US mortgage record.

About Me

Branch Manager for one of the largest banks in my country. 23 years within the Banking industry and has worked in nearly all roles. Including a Data Analyst.

Over my 2 decades I have noticed lending decisions aren’t solely based on credit history. They also depend on the internal policies of the bank. To fully understand this I looked at the data within the US HMDA dataset which is arguably the most transparent credit dataset.

Data Insights

There are large discrepancies in the rejection rates of similar FHA applications between lenders. ("Door Effect")

However there is no reason why. I don’t try to speculate. This entire website was made so no assumptions could be made from this.

What I Have Done So Far

Processed 1,187,606 FHA credit decisions from the 2025 HMDA dataset. (Excluding Reverse Mortgages)

Claimed 386 times. With receipts.

Provided one of only two sources for 2025 FHA Rejection Rates. (The other is paywalled). It is the only source to show the denominator.

MCP Server. (16 Read Only Tools) Being reviewed to be added to the Anthropic connector list.

Hugging Face Datasets. Zenodo DOI. Transparent Error Log showing 18 types of errors.

Test Suites Scoring 20/20. And 12/12.

Developed a small drift classifier model recently. Scoring 0.996 exact matches.

Key Insight

AI models can be numerically accurate, however inaccurate contextually. Missing information such as the population, time period, loan type and what exactly the number proves.

All numbers on my site have receipts to show where they came from. How many and when. I don’t claim there to be one true rate. I simply provide you with a number, denominator and what it can’t prove.

Funding

$400 ($600: Second-provider grader runs and repeat fidelity tests.$) Covers the cost of processing and publishing the 2026 HMDA dataset on the first day (March 2027).

$6K ($500: GPU training for a small open-source model using the receipt reward.The rest is my evenings. I run this whole project alone after my regular bank shift. The basic goal ($) Covers the cost of processing data and some of my time. If the full funding amount ($18K) is met I will be able to publish the 2026 dataset, create another test suite and develop version 2 of the model.

Why I Am Not Selling

I have a strict policy of not taking any money from the lenders we are measuring.

In this market almost all revenue streams would come from the lenders. By having a grant I can ensure the data is completely unbiased and independent.

Within 3 Months

Publish results from second grade evaluations. Along with raw files.

Drift Classifier will be live on Hugging Face. Along with evaluation metrics.

Connector will be published on the Anthropic directory. Along with usage metrics.

Have at least one independent paper citing the dataset. (StatRanker.org cited in Oct 2026).

I feel this data project is important. Data is how I breath.