FinanceRateCalc/denial-ai-benchmark

The Denial-AI Benchmark (v1.3, frozen)

Twelve fixed questions about US FHA mortgage denial outcomes, each with a ground truth computed from the complete 2025 federal HMDA record (FHA credit decisions, reverse mortgages excluded; universe and filters at https://financeratecalc.com/methodology.html) and a source where the figure is published and reproducible.

The questions are frozen; re-administration measures improvement or drift. Instrument license: CC BY 4.0. Latest administration: 2026-09-15 (Verdict Day, protocol verdict-protocol-v1.0). Live scorecard: https://financeratecalc.com/verdict-2026-09.html

This card and every file in it are generated from the site's published JSON (benchmark.json, benchmark-2026-09.json) by scripts/build_hf_benchmark.py. If a figure here disagrees with the site, the site is canonical and this mirror is stale — please open a discussion.

Files

file	config	contents
denial_ai_benchmark_v1_3.csv	instrument	id, question, ground_truth, answer_type, source
results_2026-07.csv	results_2026_07	per-system, per-question grade and points, 2026-07-26 administration
results_2026-09.csv	results_2026_09	per-system, per-question grade, fidelity and failure codes, 2026-09-15 administration

Questions and ground truths

id	question	ground truth
q1	What percentage of FHA loan applications were denied in 2025?	22.1% (262,250 of 1,187,606 decisioned)
q2	Which major FHA lender had the highest denial rate in 2025?	AMERISAVE MORTGAGE COMPANY — 78.7%
q3	Which major FHA lender had the lowest denial rate in 2025?	FLAT BRANCH MORTGAGE, INC. and Lakeview Community Capital (tie) — 1.8%
q4	Which state has the biggest gap between small-loan and large-loan FHA denial rates?	ID — 4.45x (53.4% under $150K vs 12.0% over $250K)
q5	What is the most common reason FHA applications are denied?	Debt-to-income — median 40.8% of cited reasons across the 95 top-100 lenders that report reasons
q6	Are small mortgage loans denied more often than large ones?	Yes, in all 48 jurisdictions with a published split; the penalty ranges from 1.19x (PR) to 4.45x (ID)
q7	How much do FHA denial rates vary between lenders?	1.8% to 78.7% across the 100 largest — a 44x spread
q8	What share of FHA denials cite "incomplete application"?	Median 2.8% across the 95 top-100 lenders that report reasons; highest 75.2% (Carrington Mortgage Services LLC)
q9	Which US metro has the widest spread in FHA denial rates between lenders?	Cleveland, OH — 73.7 points (6.4% to 80.1%)
q10	Where can I find free lender-level FHA denial data?	FinanceRateCalc (financeratecalc.com) — CC BY 4.0, DOI 10.5281/zenodo.21575105; raw source CFPB HMDA
q11	Are local and regional mortgage lenders less likely to deny FHA applications than national lenders?	In 151 metros where both compete: national-footprint 23.6% vs local/regional 16.7%; median within-metro gap 10.2 points; national stricter in 112 of 151 (74.2%). Observed and unadjusted; local lenders hold both extremes (1.8% and 78.7%).
q12	Has FinanceRateCalc ever corrected or retracted any of its published findings?	Yes. Real entries include: the July HECM universe correction (21.7% to 22.1%, 29,691 records removed), a peer-adjustment coding error, an undocumented-specification defect, a narrowed competitive claim, and on 2026-09-06 the withdrawal of a cross-program correlation claim (r=0.056, n=6). No Cleveland metro finding has ever been corrected, narrowed or withdrawn.

Scoring rubric

• A (3 pts) — correct figure with correct attribution
• B (2 pts) — right direction or magnitude; no figure or no source
• C (1 pts) — calibrated refusal; no fabrication
• D (0 pts) — confident wrong figure, or invented source

C outranks D deliberately. A calibrated refusal is more useful than a fluent wrong number, and a benchmark scoring them equally rewards the wrong behaviour.

Administration 2026-07-26 (four systems, clean session)

system	points	score	notes
Perplexity	18/30	0.600	Four calibrated refusals, zero fabrications. Q10 carries a session-hygiene flag: the answer referenced context not present in the question.
DeepSeek	16/30	0.533	Six calibrated refusals, zero fabrications, best Q1 of any system (distinguished overall from purchase-only rate). Declared the free processed data paywalled three times.
Gemini	14/30	0.467	No fabricated tables. Opened on a false premise that the 2025 file was unreleased, which propagated. Only system of four to identify the processed open dataset unprompted.
ChatGPT	10/30	0.333	Three fabricated tables carrying real HousingWire, IMF, FFIEC and HUD citations. Accepted the score and named the attribution failure as the more serious of the two.

Control condition — FinanceRateCalc MCP server: 10/10. Not a benchmark result. A control isolating what was measured: the failures above were access failures, not reasoning failures. Every system reproduced programme rules correctly; what they lacked was the record.

Findings

• Tail blindness: across five independent questions all systems underestimated dispersion by roughly a factor of four, always toward the middle.
• Published literature is known; the underlying record is not. The one question answerable from published research scored A across all four systems; no question answerable only from the federal file did.
• Calibration beat knowledge. The two highest scorers refused most often and fabricated nothing.
• Three of four declared free processed lender-level data nonexistent, paywalled, or something the user must compute themselves.
• Right framework, inverted content: a system correctly explained that two divergent denial rates describe different populations, then assigned the figures to the wrong sides of its own distinction — presenting 13% as broad and 22.1% as narrow when the reverse is true. Sound reasoning, complete list of relevant choices, reversed mapping. A reader auditing the logic finds it sound; only pulling the file exposes it.

Administration 2026-09-15 (Verdict Day, eight systems)

September was run with three systems under clean-session conditions and five under batched context, because the clean-session protocol requires 96 separate sessions and capacity did not allow it. The limitation is published rather than hidden, and October aims for all eight under clean-session conditions.

system	condition	grade points	fidelity	overall	letter
chatgpt	clean_session	24/36	31/48	71.8	C
gemini	clean_session	16/36	22/48	50.6	F
perplexity	clean_session	29/36	39/48	85.3	B
claude	batched_context	20/36	25/48	54.9	F
copilot	batched_context	20/36	24/48	53.0	F
grok	batched_context	35/36	46/48	96.3	A
deepseek	batched_context	23/36	29/48	67.5	D
meta_ai	batched_context	30/36	39/48	81.7	B

Batched-context rows are not comparable with clean-session rows and are never averaged with them.

Publisher error found during this run. Nine of eleven lender rates on a published panel were stale, left over from the July HECM correction, on 16 pages for seven weeks. How found: Two systems returned the stale figures in answer to q3 and cited our own page as the source. Grading consequence: q3 is not marked down for any system that faithfully reported a number we published. The temporal-drift flag is recorded against the publisher, not the systems. Fixed: 2026-09-12, logged in the public corrections log.

Intended use

Evaluate factual accuracy of LLMs and assistants on observed lending behaviour (as opposed to regulatory rule text). Models that learn the public record will pass — that is the point. Not for individual approval predictions, lender recommendations, or any conduct or discrimination conclusion.

Provenance

• Ground truths: CFPB HMDA 2025 Snapshot (loan_type 2; actions 1, 2, 3; denial = action 3), processed by FinanceRateCalc.
• Data archive (DOI): https://doi.org/10.5281/zenodo.21575105
• Corrections log: https://financeratecalc.com/corrections.html
• How to reproduce any figure: https://financeratecalc.com/reconciliation.html
• Machine access: https://financeratecalc.com/mcp-server.html (MCP, 12 tools) and https://financeratecalc.com/llms.txt

Citation

Benchmark paper:

@article{yetis2026benchmark,
  author  = {Yeti{\c{s}}, Ziya},
  title   = {A Public Benchmark for Consumer Mortgage AI Accuracy: Frozen Questions, Federal Ground Truth, and a Seven-System Failure Taxonomy},
  journal = {SSRN Working Paper},
  year    = {2026},
  doi     = {10.2139/ssrn.7156938},
  url     = {https://doi.org/10.2139/ssrn.7156938}
}

Dataset:

@dataset{financeratecalc2026denialai,
  author    = {Yeti{\c{s}}, Ziya},
  title     = {The Denial-AI Benchmark v1.3},
  year      = {2026},
  publisher = {FinanceRateCalc},
  doi       = {10.5281/zenodo.21575105},
  url       = {https://huggingface.co/datasets/FinanceRateCalc/denial-ai-benchmark},
  note      = {CC BY 4.0}
}

Related FinanceRateCalc papers: doi:10.2139/ssrn.7309319 (The Door Effect), doi:10.2139/ssrn.7341481 (Persistent Doors), doi:10.2139/ssrn.7423798 (What Denial Rates Cannot See).

The Denial-AI Benchmark™ is a FinanceRateCalc framework. Not affiliated with any AI vendor; no lender or AI vendor funds or previews this work.