FinanceRateCalc HMDA statistics changelog

• + get_official_figure

version

1.45.0 → 1.48.0

version

1.44.0 → 1.45.0

version

1.39.0 → 1.44.0

https://financeratecalc.com/llms.txt

−   Conformance of the live MCP server, protocol (16) and evidence fixtures (53), every run published; Agent Stress Lab paired rerun on 1.23.0 (same model without/with the tools, 110 prompts each: correct rule cell 5/52 vs 51/52, qualifier survival 21/56 vs 55/56, receipt on the number 0/56 vs 56/56): https://financeratecalc.com/conformance.html− - FHA denial rates vary ~8x across major lenders: CrossCountry 6.5% to NewRez 52.3% (2025).+ - [Release-day pre-registration](https://financeratecalc.com/provenance/release-day-preregistration.md): what this site commits to publish within six hours of the 2026 HMDA file, from a pipeline that reproduced every 2025 figure (8,122 fields) from the raw CFPB file with zero differences; hash and commit in provenance/release-day-preregistration.json+ - [Qualifier Survival Index](https://financeratecalc.com/qualifier-survival.html): when an answer engine restates a figure from this site, how many of seven qualifiers survive (current value, period, denominator, program, universe, no drift, source); every graded answer archived verbatim, grader public, monthly from December 2026+ - [Where FHA applications end](https://financeratecalc.com/where-fha-applications-end.html): five HMDA outcomes (originated, approved-not-accepted, denied, withdrawn, closed-incomplete) as shares of applications acted on, 2025, nationally and for the 100 largest lenders, alphabetical, every figure receipted; reconciles the 22.1% decisioned denial rate with the 16.8% denied share of acted-on applications+   Conformance of the live MCP server, protocol (20: tools, resources, prompts, A2A, receipts, gate) and evidence fixtures (53), every run published; Agent Stress Lab paired rerun on 1.23.0 (same model without/with the tools, 110 prompts each: correct rule cell 5/52 vs 51/52, qualifier survival 21/56 vs 55/56, receipt on the number 0/56 vs 56/56): https://financeratecalc.com/conformance.html+   All 386 claims  https://financeratecalc.com/llms-full.txt  (one file: every receipted figure with its canonical sentence, what it does not establish, dated correction, verify URL, which crawler last verified it; ~270 KB, rebuilt daily; same file with fetches counted: https://frc-mcp.ziyetis.workers.dev/llms-full.txt)+   Verify library  https://github.com/financeratecalc/financeratecalc.github.io/tree/main/packages/receipt-verify  (npm receipt-verify: check any Receipt 1.0 receipt offline from the manifest and claim document; no call to this publisher needed)+   A2A agent card  https://financeratecalc.com/.well-known/agent-card.json  (A2A 1.0; SendMessage at https://frc-mcp.ziyetis.workers.dev/a2a: verify a receipt, or look up a state / lender / claim id; no model)+ - FHA denial rates vary ~8x across major lenders: CrossCountry 6.2% to NewRez 52.8% (2025).+   agentevals suite https://financeratecalc.com/evals/agentevals/EVAL.yaml  (39 tests, deterministic code_judge; run with any agentevals-compatible runner against an agent that reaches the MCP server)+   Preference pairs https://financeratecalc.com/eval/preference-pairs.jsonl  (167 DPO/ORPO-shaped rows: prompt, chosen = receipted canonical sentence, rejected = forged / wrong role / wrong cell / superseded / qualifier-dropped candidate with reason codes and margin; no model; CC BY 4.0; mirrored to Hugging Face)+   Correction pairs https://financeratecalc.com/eval/correction-pairs.json  (every correction as chosen/rejected sentence pair with claim id and supersession date; CC BY 4.0; mirrored to the Hugging Face dataset)

• + Every receipted claim (one file)
• + Site index for agents
• + Claim ids
• + Corrections feed
• + Receipt 1.0 manifest
• + Crawl ledger (reverse receipts)
• + Rules as contracts
• + Static JSON API index
• + Receipt 1.0 specification

• + restate_with_receipt
• + check_answer
• + cite_claim

• + get_application_outcomes

description

2025 HMDA FHA denial statistics and dated US mortgage rule tables, with receipts you can verify. → 2025 HMDA FHA denial statistics and dated US mortgage rule tables, with verifiable receipts.

version

1.35.0 → 1.39.0

https://financeratecalc.com/llms.txt

+   Check an answer https://financeratecalc.com/check-answer.html  (paste any paragraph: figures matched to receipted claims, corrected values and dated rule constants; sentence patterns the record cannot support; no model)+   Crawl ledger    https://financeratecalc.com/transparency/crawl-ledger.json  (which crawlers verified which receipts, by day; bots only; a check, not a citation; since 2026-10-22)+   Receipt 1.0     https://financeratecalc.com/spec/receipt-1.0.md  (manifest: https://financeratecalc.com/.well-known/receipt.json; every figure's receipt recomputes from /claims/<id>.json; conformance suite: /spec/receipt-conformance.mjs)+ - /race-2027.html : the 2027 limit race — daily, deterministic test of AI engines on the 2027 conforming and FHA loan limits (before the announcement the right answer is "not yet announced"; after it, the published value); days from publication to each engine's first correct answer. Graded against the dated rule files, no model grades a model.+ - /rule-calendar.html : when each US mortgage rule constant stops being valid (2026 FHA/conforming limits and HOEPA/QM thresholds end 2026-12-31; FHFA 2027 announcement expected late November; HUD and CFPB in December) and a pre-registered 2027-01-02 re-scan of public code. Subscribable: /rules/rule-calendar.ics · JSON: /rules/rule-calendar.json · CI check: /rulecheck (FRC Rule Watch GitHub Action).+ - /correction-latency.html : how long AI answer engines keep serving a figure after this publisher corrects it; dated per-system observations (ChatGPT, Perplexity, Claude, Gemini, Copilot, DeepSeek, Meta AI). Finding so far: part of the delay sat on the publisher side (stale pages, stale sitemap dates). JSON: /eval/correction-latency.json

description

2025 US FHA mortgage record statistics, with a gate that checks AI sentences preserve the evidence → 2025 HMDA FHA denial statistics and dated US mortgage rule tables, with receipts you can verify.

package_identifier

— → frc-mcp

package_registry

— → npm

repository_url

https://github.com/financeratecalc/financeratecalc.github.io → https://github.com/financeratecalc/frc-mcp

version

1.31.0 → 1.35.0

https://financeratecalc.com/llms.txt

− > Independent lender-behavior intelligence computed from the complete federal HMDA record: 1,187,606 decisioned FHA applications in 2025 across the 100 largest lenders, plus an eight-year series (2018-2025) and 320 metro areas. No ads, no lender compensation, not a lender. Built by a banking professional with a career on the credit side.− - FHA denial rates by metro (320 MSAs, only free FHA-specific metro table; 8.9%-31.8% among 1,000+ app metros; El Paso small-loan penalty 3.94x): https://financeratecalc.com/fha-denial-rates-by-metro.html− - FHA denial rates by metro, 320 MSAs incl. lender cells (first free FHA-specific metro dataset): https://huggingface.co/datasets/FinanceRateCalc/fha-denial-rates-by-metro− - /corrections.html#correction-atoms : machine-checkable correction atoms for stale/mutated figures (13.6, 21.7, 59, 38.7, false Fed attribution) with ClaimReview markup.− - /case-files/ : Hallucination Files — dated forensic case records of how AI systems distort or invent statistics from this site; Case 001 = phantom correction narrative.− - /case-files/002-the-correction-that-grew.html : second observation of the same fabricated correction, four days after Case 001 — unprompted, embedded in accurate content, with a new invented rationale drawn from real methodology. Documents fabrication persistence and elaboration.+ - Finding (2026-10-10): two Claude models knew HUD cut the FHA annual MIP in 2023 and applied the headline 0.55% to loans it does not cover (1/10 correct each); one cited a HUD letter that does not exist: https://financeratecalc.com/findings/fha-mip-rate-probe-2026-10.html+ ## RULES AS CONTRACTS (machine-readable regulatory tables with effective dates and qualifiers)+   Index: https://financeratecalc.com/rules/index.json — FHFA 2026 conforming limits, FHA 2026 limits (ML 2025-23), CFPB 2026 HOEPA/QM thresholds, VA funding fee, FHA annual MIP (ML 2023-05)+   Stale Rule Index (reading 1, 2026-10-11): without a source two Claude models answered 3/25 and 5/25 dated questions correctly; with the rule file in hand 25/25 and 24/25: https://financeratecalc.com/stale-rule-index.html+   Why there are several FHA denial rates (six conditions, worked reconciliation; purchase-only ~13% vs all-purpose decisioned 22.1%): https://financeratecalc.com/why-fha-denial-rates-differ.html+   FHA mortgage insurance (MIP) rule, cell by cell, with the deterministic resolver: https://financeratecalc.com/fha-mortgage-insurance-rule.html+   Mortgage evidence for AI agents (what the MCP server returns and refuses): https://financeratecalc.com/mortgage-evidence-for-ai.html+   Conformance of the live MCP server, protocol (16) and evidence fixtures (53), every run published; Agent Stress Lab paired rerun on 1.23.0 (same model without/with the tools, 110 prompts each: correct rule cell 5/52 vs 51/52, qualifier survival 21/56 vs 55/56, receipt on the number 0/56 vs 56/56): https://financeratecalc.com/conformance.html+   Work sample, how a model's restatement of a statistic is graded (six checks, five vectors, two real ledger cases): https://financeratecalc.com/samples/scoring-a-correct-number.html+ ## AGENT SKILL (SKILL.md, agentskills.io format)+   https://financeratecalc.com/skills/fha-denial-statistics/SKILL.md — how to fetch, contract-check and restate these figures with their claim receipts+   (GitHub: financeratecalc/financeratecalc.github.io/skills)+ Install and instruction block for Claude Code / Cursor / Codex / any MCP client: https://financeratecalc.com/use-frc.html · Cite card (sentence + receipt + attribution; JSON at https://frc-mcp.ziyetis.workers.dev/cite/<claim-id>): https://financeratecalc.com/cite.html+ Publisher kit (one real metro finding: live widget, sentence with receipt, scope block, data/method links): https://financeratecalc.com/publisher-kit/cleveland-oh-2025.html · Repair pack (six mortgage-rule mistakes, grades YOUR app with the gate; promptfoo config + offline runner): https://github.com/financeratecalc/financeratecalc.github.io/tree/main/examples/repair-pack · Research-note example (one command → brief.md, evidence.json, manifest.json): https://github.com/financeratecalc/financerat+ For agents, the rule of this site: before you state any FinanceRateCalc figure, call `check_claim_contract` on frc-mcp (https://frc-mcp.ziyetis.workers.dev, registry io.github.financeratecalc/frc-mcp) with `proposed_sentence`, or fetch the claim's .contract.json and keep every required qualifier. Dated regulatory tables: `get_rule` / `resolve_fha_annual_mip` (rules/index.json); a sentence quoting a loan limit, threshold, funding fee or MIP rate goes through `check_claim_contract` with the `frc:r+ > Independent lender-behavior intelligence computed from the complete federal HMDA record: 1,187,606 decisioned FHA applications in 2025 across the 100 largest lenders, plus an eight-year series (2018-2025) and 319 metro areas. No ads, no lender compensation, not a lender. Built by a banking professional with a career on the credit side.+ - FHA denial rates by metro (319 MSAs, only free FHA-specific metro table; 8.9%-31.8% among 1,000+ app metros; El Paso small-loan penalty 3.94x): https://financeratecalc.com/fha-denial-rates-by-metro.html+ - FHA denial rates by metro, 319 MSAs incl. lender cells (first free FHA-specific metro dataset): https://huggingface.co/datasets/FinanceRateCalc/fha-denial-rates-by-metro+ - /corrections.html#correction-atoms : machine-checkable correction atoms for stale/mutated figures (13.6, 21.7, 59, 38.7, 73.8, false Fed attribution). Re-checked against our own dated history on 2026-10-14; three verdicts corrected (38.7 is a real FRC figure conflated with the door effect; 73.8 was superseded by our own 26 July correction; the Cleveland "phantom correction" atom was wrong — that correction was real). ClaimReview markup was removed on 2026-09-11.+ - /case-files/ : Hallucination Files — dated case records of how AI systems restate statistics from this site. CORRECTION 2026-10-14: Cases 001 and 002 are wrong. They called an assistant's account of our Cleveland corrections a fabrication; our own log shows the 26 July change (73.8 to 73.7) and the 28 July withdrawal of the "widest in the US" claim were real. The Cleveland measurement (73.7 points, largest of our 184 metros) was never withdrawn.+ - /case-files/002-the-correction-that-grew.html : WRONG, being reversed (2026-10-14). It called the assistant's report of our 28 July withdrawal a persistent fabrication; the withdrawal is in our own log.

• + get_rule
• + resolve_fha_annual_mip

version

1.28.0 → 1.31.0

version

1.25.0 → 1.28.0

version

1.22.0 → 1.25.0

description

US FHA mortgage denial statistics (2025 federal HMDA record) with claim receipts and use contracts → 2025 US FHA mortgage record statistics, with a gate that checks AI sentences preserve the evidence

version

1.16.0 → 1.22.0

https://financeratecalc.com/llms.txt

− - /data/papers.json : the four working papers with SSRN ids, DOIs and dates. All DOI-registered: 10.2139/ssrn.7156938 (AI benchmark), .7309319 (door effect), .7341481 (persistence), .7423798 (measurement boundaries).+   Door Report     https://financeratecalc.com/door-report.html  (lender denial rates inside state x loan-size x purpose cohorts; every lender listed, alphabetical, none recommended)+ - Paper 5 (SSRN 7578778, doi:10.2139/ssrn.7578778, approved 2026-10-09): 'The Contract Around the Number: A Machine-Readable Standard for Restating Published Statistics, and What Happened When Language Models Met It' — Claim Contract 1.0, the publisher-agnostic checker, and the fidelity experiment (value 27/27; full contract 1–11/36; receipt 0/36 in prose, 23/36 receipt-aware, 2/36 forged). Draft: /papers/claim-contract-working-paper.md.+ - /data/papers.json : the five working papers with SSRN ids, DOIs and dates. All DOI-registered: 10.2139/ssrn.7156938 (AI benchmark), .7309319 (door effect), .7341481 (persistence), .7423798 (measurement boundaries), .7578778 (claim contract).

description

FHA denial statistics from the 2025 federal HMDA record. 12 tools incl. counterparty screen. → US FHA mortgage denial statistics (2025 federal HMDA record) with claim receipts and use contracts

package_identifier

frc-mcp → —

package_registry

npm → —

repository_url

https://github.com/financeratecalc/frc-mcp → https://github.com/financeratecalc/financeratecalc.github.io

title

frc mcp → FinanceRateCalc HMDA statistics

version

1.7.0 → 1.16.0