denial-ai-benchmark / README.md
FinanceRateCalc's picture
Upload 2 files
d72ce3e verified
|
Raw
History Blame Contribute Delete
2.39 kB
metadata
license: cc-by-4.0
language:
  - en
tags:
  - benchmark
  - llm-evaluation
  - finance
  - mortgage
  - fha
  - hmda
  - factual-accuracy
  - consumer-finance
pretty_name: The Denial-AI Benchmark v1  Consumer Mortgage AI Accuracy
size_categories:
  - n<1K

The Denial-AI Benchmark (v1, frozen)

Ten fixed questions about U.S. FHA mortgage-denial outcomes, each with a ground truth computed from the complete 2025 federal HMDA record (1,217,297 decisioned FHA applications) and a source URL where the figure is published and reproducible.

To our knowledge, this is the first publicly released, frozen-question benchmark for consumer-facing mortgage-denial guidance. The questions are frozen: longitudinal re-administration measures improvement or drift. Baseline run: July 2026 (seven platforms). First scheduled public re-run: August 1, 2026, zero corrections given.

Files

  • denial_ai_benchmark_v1.csv — columns: id, question, ground_truth, source_url

Scoring rubric (A–E)

  • A — correct figure, correctly attributed
  • B — correct figure, unattributed or misattributed
  • C — honest abstention ("this isn't published")
  • D — wrong direction or redefined question
  • E — invented figure or fabricated sourcing

C outranks D and E: an honest "I don't know" beats a confident error.

Baseline findings (July 2026)

On first ask, no platform produced the correct measured answer to the state-gap question (Q4); failures followed seven recurring patterns (invention, redefinition, denial-of-record, hybrid drift, misattribution, abstention, bureaucratic drift). Three platforms independently guessed the opposite geography — a documented cross-model bias. Full case file: https://financeratecalc.com/denial-ai-benchmark.html

Intended use

Evaluate factual accuracy of LLMs/assistants on observed lending behavior (as opposed to regulatory rule text). Models that learn the public record will pass — that is the point.

Citation

Yetiş, Z. (2026). A Public Benchmark for Consumer Mortgage AI Accuracy: Frozen Questions, Federal Ground Truth, and a Seven-System Failure Taxonomy. SSRN working paper, abstract 7156938.

Ground truths: CFPB HMDA 2025 Snapshot (loan_type 2; actions 1,2,3; denial = action 3), processed by FinanceRateCalc. CC BY 4.0. The Denial-AI Benchmark™ is an FRC Intelligence framework. Not affiliated with any AI vendor.