FinanceRateCalc · Pilot · human-reviewed

Does your mortgage sentence actually match the source?

Send up to five public claims — from an AI answer, a report, a dashboard or an article — and get back a short Evidence Brief: what the federal record supports, where the sentence goes past it, and a safe rewrite you can publish. Same day, reviewed by a human before it leaves.

What we check, atom by atom

AtomThe audit question
ValueIs the number correct at the source?
UniverseWhich population, program or observation set does it describe?
DenominatorIs the denominator stated, and used correctly?
TimeIs the data year preserved, or has a stale figure survived?
GeographyAre national, state, metro and institution levels being mixed?
QualifierDo the limiting words survive — "associational", "explained variation"?
BoundaryHas an aggregate been carried onto an individual or a lender recommendation?
CitationDo the source, passport and method links travel with the claim?

Every claim comes back with exactly one verdict — PASSNEEDS_QUALIFIEROUT_OF_SCOPE — plus a reason code, a publishable rewrite, and links to the claim passport and contract behind it.

Three worked examples, from our own file

NEEDS_QUALIFIER“Lender identity causes 38% of FHA denials.”
The published figure is 37.97% of the explained variation, and it is associational: HMDA has no credit scores. Safe rewrite: “About 38% of the variation these models can explain attaches to lender identity rather than the applicant file (associational; 859,090 decisions).”
OUT_OF_SCOPE“AmeriSave denies 78.7%, so your application there will probably fail.”
An institution-level aggregate cannot be carried onto an individual application, and a high observed rate is not evidence of misconduct or discrimination. No rewrite makes this claim supportable; it has to be dropped.
NEEDS_QUALIFIER“The FHA denial rate is 22.1%.”
True, but incomplete without its universe and denominator: 2025 forward FHA loans, denied ÷ (originated + approved-not-accepted + denied), HECM excluded, 1,187,606 decisioned applications. Without them the number silently changes meaning.

We describe these as the distance between the evidence and the sentence — not as “the AI lied.” The referee doesn't take sides.

Pilot packages

PackageWhat you getPilot price
Single Claim CheckOne public claim, short audit, safe rewrite$39
Five-Claim AuditUp to five claims, atom-level audit, Evidence Brief$99
Publisher / Team PackTwenty claims, method note, reusable contract checklist for your desk$249

Turnaround: same day for the first two packages when submitted before 18:00 UTC; two business days for the team pack. We invoice after scope is confirmed — no payment before we've told you whether your claims are auditable.

What this service is not. It does not assess personal credit eligibility, predict mortgage approval, select or rank lenders for you, or provide financial advice. We audit public claims for evidence, scope and safe language — nothing else. Do not send borrower files, credit reports, income documents, application data or protected-class information; submissions containing them will be deleted unread. Nothing you send is published without your written permission.

Submit a claim for scoping

Or email [email protected] with “Claim Audit” in the subject.

FinanceRateCalc · Measured, not assumed. · The underlying data is free forever: metro gaps · method · what AI systems get wrong · our own corrections