Domain Atlas / Lending & credit collections AI
Navy Federal mortgage underwriting & three readings of one gap
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 1 assumed · 8 published baseline.
Navy Federal Credit Union underwrites residential mortgages through what the pleadings and the Fourth Circuit both describe as an 'at-least semi-automated underwriting process' built on a 'proprietary underwriting algorithm' whose variables and weights are, in the complaint's words, 'entirely up to Navy Federal,' which 'maintains secrecy' over both; no third-party underwriting vendor is identified anywhere in the verified record. CNN reported on 14 December 2023 that for 2022 conventional home purchase mortgages the credit union approved 77 per cent of White applicants, 69 per cent of Asian applicants, 56 per cent of Latino applicants and 48 per cent of Black applicants — a spread of nearly 29 percentage points, the widest of the fifty lenders that originated the most mortgages that year — and that holding more than a dozen variables constant Black applicants were more than twice as likely to be denied as White applicants and Latino applicants roughly 85 per cent more likely. Navy Federal disputes that analysis, saying the statistics 'do not appear to have considered several key credit criteria,' and adds that it ranks first among large lenders in the share of mortgages made to Black borrowers, made $3.5 billion in 2022 mortgages to Black borrowers, and counts roughly one in four members as Black. Separately, and unconditionally, the CFPB's own HMDA Data Browser queried directly for this institution (LEI 5493003GQDUH26DNNH17) across all reportable loan types and purposes gives denial shares of originated-plus-denied applications, White against Black, of 23.8 and 45.0 per cent in 2018, 31.0 and 50.8 in 2022, 34.5 and 56.6 in 2023 and 27.4 and 46.8 in 2025; those counts control for nothing whatsoever, and in particular not for credit score, which the public loan-level file excludes by rule. No court and no regulator has ever found that Navy Federal discriminated in mortgage underwriting.[4]
What happened
Navy Federal Credit Union is the largest credit union in the United States: 15.3 million members, $197.1 billion in assets, 25,200 employees and 382 branches as of 31 December 2025, of which 178 sit on or near military installations. Its field of membership is the Department of Defense, all uniformed services, veterans and their families. In 2022 it closed almost 50,000 mortgage loans totalling $16.5 billion against an $84.3 billion mortgage portfolio.
How it underwrites those mortgages is described on the public record in exactly one way and no further. Every applicant completes a Uniform Residential Loan Application collecting name, Social Security number, citizenship, the subject property's address and value, intended occupancy, address history, employment, income, assets, liabilities and expenses, and military service. That data, "along with other inputs," goes through what the complaint and the Fourth Circuit both call an "at-least semi-automated underwriting process" built on a "proprietary underwriting algorithm." What variables that algorithm uses and what weight it gives them "is entirely up to Navy Federal," which "maintains secrecy" over both. No third-party underwriting vendor is named anywhere in the verified record.
On 14 December 2023 CNN published an analysis of Home Mortgage Disclosure Act data. For 2022 conventional home purchase mortgages, it reported, Navy Federal approved 77 per cent of white applicants, 69 per cent of Asian applicants, 56 per cent of Latino applicants and 48 per cent of Black applicants — a spread of nearly 29 percentage points, the widest of the fifty lenders that originated the most mortgages that year. Holding more than a dozen variables constant — applicant income, debt-to-income ratio, loan amount and term, loan-to-value, property value, presence of a co-applicant, applicant and co-applicant sex, credit scoring model, primary applicant age, the tract's minority population percentage and median housing age, and the gap between metro-area and tract median income — it reported Black applicants more than twice as likely to be denied as white applicants and Latino applicants roughly 85 per cent more likely. It published its own limits in the same piece: applicant credit score, available cash deposits and relationship history with the lender are not in the public mortgage data.
That last sentence is the hinge of everything that follows.
Two records come out of one loan file and they are not the same record. The full disclosure submission that goes to the regulator carries the applicant credit score, the debt-to-income ratio and the loan-to-value. The public modified file suppresses the credit score by rule and coarsens other fields. So an analysis built on the public file can be answered, accurately and unfalsifiably, by pointing at what it could not see — and the answer cannot be checked, because the thing it rests on is not published either.
The credit union answered on 18 December 2023. It was "closely examining the allegations"; "the statistics in the article do not appear to have considered several key credit criteria"; it ranked first among large lenders in the share of mortgages made to Black borrowers, with $3.5 billion in 2022 mortgages to Black borrowers and roughly one in four members Black. It had retained Debo P. Adegbile — a former commissioner of the U.S. Commission on Civil Rights — to assess its mortgage lending policies and practices.
Three days earlier, on 17 December, Laquita Oliver and Cherelle Jacob had filed a putative class action in the Eastern District of Virginia. Related suits were consolidated; the operative complaint of 20 February 2024 named nine plaintiffs, eight Black and one Hispanic, pleading the Fair Housing Act, the Equal Credit Opportunity Act, 42 U.S.C. § 1981 and state analogues. The lead plaintiff's account, a pleading allegation, describes the correction channel this deployment offers: after a month-long application that included a hard credit pull which lowered her score and for which she was charged a fee, she was denied; when she complained she was initially ignored and then told she could appeal, but that the appeal would require another hard pull. She abandoned the application.
Civil society named the mechanism early. On 20 December 2023 Consumer Reports asked the Consumer Financial Protection Bureau both to investigate and to give all lenders clearer guidance on algorithmic scoring, arguing that "technology now exists for lenders to identify less discriminatory underwriting models that do not sacrifice accuracy and predictiveness." That test reappeared ten weeks later as question six of a congressional letter to two federal regulators.
Congress ran the loudest channel in this matter and holds no compulsory process over a credit union. On 11 January 2024 ten senators led by Banking Committee Chair Sherrod Brown wrote to HUD and the CFPB asking for a thorough fair-lending review. Congressional Black Caucus members and Sen. Cortez Masto made a parallel call the next day. On 15 February forty CBC members' requested meeting with CEO Mary McDuffie took place; the members reported afterwards that they had asked for "at minimum, aggregate data from Navy Federal regarding credit scores or any other non-public variable that Navy Federal has suggested serves as an explanation," and that "Navy Federal failed to provide this information." Rep. Cleaver said the CEO "could not provide an answer." On 28 February the New Democrat Coalition and the Congressional Hispanic Caucus led twenty-one members in letters to the credit union and to six agencies, asking for underwriting procedures, compliance management systems, fair lending training and monitoring, discrimination-complaint tracking, the stated legitimate non-discriminatory reason for the disparities, and whether any Special Purpose Credit Program exists. On 1 March Reps. Cleaver, Horsford and Kamlager-Dove put ten numbered questions to the NCUA Chairman and the CFPB Director, framing the duties precisely: both agencies "have a legal obligation to supervise Navy Federal for compliance with fair lending laws, to make a referral to the Department of Justice if they have reason to believe the credit union had engaged in a pattern or practice of discrimination, and to engage in any appropriate supervisory or enforcement action." They asked what matters requiring attention had issued in five years, how fair lending findings had affected ratings, whether a Department of Justice referral had been made, and how each agency ensures a lender searches for a less discriminatory alternative. They requested answers by 5 April 2024. No public response has been located.
The same letter is careful about what the journalism does and does not establish: "While CNN's analysis does not prove that Navy Federal has discriminated against current and former servicemembers on the basis of race and national origin, it raises significant questions about unexplained and statistically significant underwriting disparities."
On 21 March 2024 Navy Federal published a statement announcing that Adegbile's review was complete and had found "no race-based decision making." His quoted conclusion: "CNN's analysis was incomplete, and the article's suggestion that Navy Federal discriminates against its members is unsupported. In fact, our review found that when all relevant factors are controlled for, which CNN did not do, the difference in approval rates between Black and White borrowers falls to less than 1%. The remaining difference in approval rates is explained by legitimate, non-race factors like income verification and incomplete credit applications." The report, the model and the underlying data were not released. Adegbile is a partner at WilmerHale, and WilmerHale is Navy Federal's defense counsel in the class action; its lawyers argued the appeal and are counsel of record on the briefs. Plaintiffs' counsel called the arrangement "a classic conflict of interest." The credit union described the work as an "external review," and its statement was later entered into the record of a June 2024 congressional hearing, which is how the operator's account of its own audit became a congressional document. Alongside it the credit union announced a newly created Office of Financial Opportunity under executive vice-president Brittani Ivey. No change to the underwriting process itself was announced and no timeline for one was committed.
A third analysis existed the whole time, run by the supervisor's own economists on data neither disputant had. The NCUA Office of the Chief Economist applied the Federal Deposit Insurance Corporation's Popick logit model to 2020 and 2021 HMDA data — including the credit-score, debt-to-income and loan-to-value fields the public file suppresses — and found credit union Black applicants at roughly 1.60 times the denial odds of white applicants on conventional purchase, 1.54 on rate-and-term refinance and 1.81 on cash-out refinance, with average marginal effects of two to four percentage points and contract-rate premiums of eight to thirteen basis points. The note is industry-wide, names no institution, separately estimates none, and carries the same caveat the operator had invoked against CNN: such results "should not be interpreted as evidence of discrimination ... as such results may reflect unobserved factors."
That is the shape of the evidentiary contest. The party with the public data and no access to the file published a controlled analysis and its own caveats. The party with the file and an interest in the answer published a conclusion without the analysis. The party with the file, the model and the supervisory authority published an industry aggregate that deliberately named no one. Nobody with subpoena power ever put the algorithm on the table.
The courts came closest. On 30 May 2024 Judge Brinkema dismissed the disparate-treatment counts because "the Complaint has failed to allege plausible direct or circumstantial evidence of discriminatory intent," let the disparate-impact theory proceed because "at the motion to dismiss stage ... the statistical disparities reveal a disparate impact among non-white loan applicants and the underwriting algorithm and process is alleged to have caused the disparity," and struck all class allegations. On 9 February 2026 the Fourth Circuit decided the appeal two to one. It is a Rule 23 procedural ruling and nothing more: Rule 23(c)(1)(A), not Rule 12(f) or Rule 23(d)(1)(D), is the source of a district court's authority to decide certification, and before discovery a court may deny certification only if the class allegations fail as a matter of law on their face. It affirmed the denial of a damages class and vacated the denial of an injunctive class, holding the complaint made "a sufficient prima facie showing" of commonality and that the district court "acted prematurely." It reserved the merits in terms: "Of course, discovery in this case might show that the [complaint's] allegations ... are false."
The commonality holding turns on the singular. The panel read the complaint as alleging one form, one algorithm and one at-least semi-automated process applied to every applicant, which would make three questions class-wide: does the credit union use a single algorithm for every applicant regardless of product; if so, does that algorithm rather than some other variable produce the alleged disparate impact; and if so, is its use justified. Judge Richardson, dissenting in part, read the same pleading the other way. It uses the plural elsewhere — "underwriting algorithms or machine learning programs," "internal approval processes," "appraisal policies" — it says nothing about how the algorithm and the loan officers interact, and the disparity "may be caused not by any underwriting algorithm, but by the individual loan officers exercising their discretion in differing ways." He also noted that the approval rates are not uniform across minority groups, the 69 per cent Asian figure sitting twenty-one points above the Black figure, and that a denial decision and a processing-time decision cannot plausibly run on the same inputs — a processing-time decision "may account for little more than the date of the application and the length of Navy Federal's backlog."
While the appeal was pending the supervisory test itself moved. Executive Order 14281 of 23 April 2025 directed federal agencies to eliminate the use of disparate-impact liability. On 4 September 2025 the NCUA issued Letter to Credit Unions 25-CU-04, removing every reference to disparate impact from its Fair Lending Guide and stating that its "examination and supervision processes will no longer include reviews for disparate impact," while continuing HMDA analysis and examinations for disparate treatment. The theory the district court had dismissed is the one the supervisor kept. The theory that survived and was revived on appeal is the one the supervisor stopped examining for.
The consumer-compliance supervisor's posture toward this institution reversed in the same window on an entirely separate matter. On 7 November 2024 the CFPB ordered Navy Federal to pay more than $95 million — $80.6 million in redress and a $15 million penalty — over surprise overdraft fees charged between 2017 and 2022. On 1 July 2025 the Bureau terminated that order and waived any alleged non-compliance. House Financial Services Democrats wrote to CEO Dietrich Kuhlmann in August 2025 asking how consumers would be made whole; the letter records the credit union saying it "firmly believe[s] the CFPB's decision to terminate the order was appropriate." That thread belongs in this file as evidence about a supervisor's posture and capacity, and never as evidence about lending discrimination.
On 25 August 2026 — six months after the mandate issued, and three days before a scheduled status conference — all nine named plaintiffs filed a notice of dismissal with prejudice. Judge Brinkema so-ordered it the same day and the conference was cancelled. The docket text states no reason and discloses no settlement. Discovery into the underwriting algorithm, which the Fourth Circuit had just made available, never produced a public answer.
The public measurement did not move with the controversy. Queried directly from the CFPB's own HMDA Data Browser for this institution across all reportable products, the denial share of originated-plus-denied applications runs, white against Black: 23.8 and 45.0 per cent in 2018; 27.0 and 49.6 in 2019; 27.8 and 50.8 in 2020; 23.7 and 45.5 in 2021; 31.0 and 50.8 in 2022; 34.5 and 56.6 in 2023; 28.0 and 48.3 in 2024; 27.4 and 46.8 in 2025. These are unconditional counts and they control for nothing whatsoever. The absolute gap runs nineteen to twenty-two points across eight filing years and it is still there after the investigation, the letters, the commissioned review, the litigation and the supervisory-policy change. Whether it reflects anything the law forbids is exactly the question none of the mechanisms answered.
The sociotechnical reading
Most cases in this atlas turn on what a system did. This one turns on what could be known about what a system did, and by whom, and it is worth being precise about the difference, because it changes what the file is evidence of.
Nothing here is a finding. No court and no regulator has found that this credit union discriminated in mortgage underwriting. No biased variable, weight or rule has been identified by anybody, and the algorithm was never put on any public record. What exists is a measurement of outcomes, taken from a legally mandated disclosure file, and three incompatible readings of what produced it. That distinction is not a hedge; it is the case's whole content. This is a reporting-based file, not an enforcement-based one, and the atlas needs both kinds.
Start with the structure that makes the dispute unresolvable rather than merely contested. One loan file generates two records. The submission the supervisors receive carries the applicant credit score, the debt-to-income ratio and the loan-to-value. The public modified file, from which every outside analysis in existence has been built, has the credit score removed by rule. Both records are accurate. Neither is contaminated. But they cannot answer the same question, and the field that is missing from the public one is precisely the field the operator says explains the gap. So the outside analysis can be answered by pointing at what it could not see, and the answer cannot be checked, because the object it rests on is not published either. A disclosure regime designed to make lending legible produced, in this instance, a record with a hole in exactly the shape of the argument.
Now watch what that does to each channel of oversight in turn, because the record supplies a complete set.
Journalism had the redacted file. It ran the controls the file supports, published its own caveats about the ones it could not, and got a result. It has no power to compel anything and knows it.
The operator's commissioned review had the complete file. That access is real and this file does not dismiss it: the reviewer could see underwriting variables the public data genuinely lacks. What he published was one sentence. Not the model, not the method, not the data — a number and an attribution of the residual to income verification and incomplete applications. He is a partner at the firm defending the credit union in the same class action, and the credit union nonetheless called the work an external review. When a congressional body asked for even aggregate supporting figures, the letter records that none were provided. A correction channel whose output is a conclusion without a method cannot function as correction however good its inputs were — and this record is the cleanest demonstration of that proposition the atlas contains, because the inputs really were the best available.
The supervisor had the complete file, the strongest model, and the authority. Its economists applied a published federal logit specification and found credit-union denial odds of roughly 1.5 to 1.9 across products, with the caveat that such results should not be read as evidence of discrimination. Then they published it as an industry aggregate naming no institution. That caution is defensible on its own terms and it has a consequence: the only party who could have said something specific about this deployment, using data adequate to say it, chose a form of publication in which nothing specific could be said. Asked in writing which of its supervisory instruments had been used here, it produced no public answer.
Congress had volume and no compulsion. Four letter campaigns, forty members requesting a meeting, ten numbered questions with a deadline. The one request whose compliance is documented went unmet, and the mechanism for that is simply that there is no mechanism. A committee letter is a request.
The courts had the only compulsory instrument. Rule 23 discovery is the sole route by which the variables and the weights could have been put on the record against the operator's wishes. It was struck before it opened in May 2024, reopened by the Fourth Circuit in February 2026 — expressly without deciding anything about the merits — and closed by the plaintiffs' own voluntary dismissal with prejudice in August 2026, three days before a status conference. The docket gives no reason. What is left is that the one channel with teeth was available for six months and produced nothing public.
And in the middle of all this the test itself was withdrawn. Executive Order 14281 directed agencies away from disparate-impact liability, and in September 2025 the NCUA removed the concept from its Fair Lending Guide and from its examinations, while keeping disparate treatment. Read that against the litigation and the symmetry is exact: the district court dismissed the disparate-treatment theory and preserved disparate impact; the supervisor kept disparate treatment and retired disparate impact. The surviving legal theory and the abandoned supervisory test became the same theory. A governance question can be closed by answering it, and it can be closed by retiring the instrument that would have asked it, and only one of those is visible in an outcome.
There is one more thing this deployment gives the atlas that almost nothing else does: a population selected to strip out the usual deflections. The field of membership is the Department of Defense, all uniformed services, veterans and their families. The gap is measured among people who share an employer, whose incomes are documented and often federally guaranteed, and who are applying to the same institution for the same products. Several of the confounders that ordinarily absorb lending gaps are attenuated before the measurement starts, which is why the argument narrowed so quickly to one live question — what does the underwriting stack actually do — and why the failure to answer it is so conspicuous.
Two theories of the answer sit unresolved in the record and this file carries both. The appellate majority read the complaint as alleging one algorithm applied to every applicant. The dissent read the same pleading as silent about how the algorithm and the loan officers interact, and offered dispersed officer discretion as the alternative cause. Nothing settles which. That is why the shape of the deployment matters more here than the size of any number: a semi-automated process where nobody outside can say which part of a decision was the machine's is a process where the two theories are indistinguishable from the outside, permanently, unless somebody opens the file.
Nobody did. The gap is still in the 2025 filing.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.