Wells Fargo refinance underwriting (CORE/ECS)

What this models

This example runs on the Bank-class rules-and-scorecards refinance underwriting network: 13 components and 29 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 1 assumed · 10 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  D48-derived org (Phase 6, lending-credit-collections), re-derived at the coarsest faithful granularity in September 2026. REGISTER FIRST, because it governs every value on this diagram. No court and no regulator has ever found that this bank discriminated in refinance underwriting. Class certification was denied on 5 August 2025 for failure of Rule 23(a)(2) commonality; the pending summary-judgment motion was never decided; both petitions for permission to appeal were denied; and seven of the eight named plaintiffs settled confidentially and were dismissed with prejudice in 2026. What is agreed is a MEASUREMENT: the operator did not dispute the arithmetic of the published approval counts, and both sides' statisticians agreed a statistical disparity along racial lines exists. What is unresolved is CAUSATION, and it is unresolved because no adjudicator ever chose between the two expert accounts. No rule, attribute, weight or scorecard column has been identified as defective by anyone. The digital-redlining theory, the surname-and-geocode race imputation, the uncorrected appraisals, the credit-score overlays, the workload and disincentive figures and the lost-paperwork account are all ALLEGATIONS from the operative complaint, carried as such wherever they appear.

• baseline

  THE SECOND REGISTER RULE, WHICH IS THIS DEPLOYMENT'S OWN: this is not an artificial-intelligence underwriting model and nothing here draws it as one. On the undisputed facts recited in the public class-certification order from the operator's own sworn declarations, the front-end workflow tool guides the flow of the loan-origination process and retrieves, stores and displays application data, engine outputs and final lending decisions, and it is not an automated underwriting system and does not make lending decisions or calculate or assign risk classes. A separate business-rule engine assigns risk classes using business rules, business services and data attributes, executing tens of thousands of separately identifiable rules across sixteen business services, its largest holding about fourteen thousand rules over more than thirteen hundred data attributes. And the scorecards are two simple scorecards whose logic, per the same declaration, can be fully stated on two or three sheets of paper, neither utilising artificial intelligence, each simple enough that an applicant's score could be calculated by hand. The operator's position, accepted as undisputed for certification purposes, is that these are distinct applications. The governance object here is a very large deterministic rule base plus human discretion, and that is what is drawn.

• baseline

  TOPOLOGY. Thirteen nodes and twenty-nine pathways, drawn at the coarsest granularity at which every documented mechanism of this deployment is still a separate thing on the board. THREE model nodes, because the order recites three distinct decision channels and the operator's own accepted position is that they ARE distinct: a rule base, two scorecards, and the two enterprise systems run alongside and controlled by nobody here. Drawing them as one would adopt the pleading's framing over the sworn record. ONE input source, drawn exactly as the PAN file draws it, carrying the credit-bureau attributes and the appraisal together with the split between them stated in the copy. THREE record stores because one decision writes three records with three different readerships: the complete workflow record the operator holds, the published extract with the credit score excluded by rule, and the copy that reached a court file under a protective order. THREE operator classes, because the sources document three groups with different authority over the error: the desk that works the file and decides it, the internal model-risk and fair-lending function, and the rule owners who author the cut-offs. THREE reviewers because the record documents three separately constituted examining channels with three different data endowments: an expert examination that read everything under seal, a forum that compelled the production and decided on proof rather than on the system, and a correspondence channel that asked for the algorithms and holds no power to compel them.

• baseline

  WHAT THE COARSER DRAWING CARRIES IN WORDS RATHER THAN AS SEPARATE ELEMENTS, and none of it is dropped. The PAN user block lists four operator classes; loan processors are drawn here inside one desk with the underwriters, because everything the record says about processors is a pleaded allegation and nothing sworn or adjudicated shows them acting on the error differently from the desk they feed. The difference between the two is real and is stated rather than drawn: processors handle the file and underwriters hold the decision, and the PAN file weights the underwriters' correction well above the processors'. The engine's class and the enterprise answer reach the desk through the workflow record, because the order says that tool displays the engine's outputs; the scorecards' insight message stays a direct pathway because the order recites it to the underwriter by name. Several relations the PAN file lists are narrated on the element they belong to rather than drawn: the rule owners reading outcomes, the underwriter's departure from the class stopping at the file, the processor ordering and reading the inputs, the tool supplying application data to the engine, and the published record reaching the internal function in no way that matters. Four others are drawn as pathways of their own: the published record into the engine, at zero; the rule owners' hold on the scorecards' policy surface; policy reaching the desk; and fair-lending training reaching the desk.

• baseline

  ABSENCES ARE DERIVED TOO, and five of them are load-bearing. There is NO model-to-model self-loop and its absence is the most deliberate choice on this diagram: a monoculture loop asserts that one system scores every case so that a single flaw repeats as correlated error, and this deployment is the opposite of that on its own sworn record — three separate channels, one of them run by parties outside this operator entirely. The one model-to-model relation drawn is a CHECK, and it runs at zero. There is NO enforcement node: a denial is a decision, and no downstream action system driven automatically by the record is documented anywhere here. There is NO worklist: the pleaded workload figures describe volume against staffing, never a queue object, and a node drawn from them would be a decoration. There is NO guardrail and NO retriever: no automated output screen and no retrieval component appears anywhere in this record. There is NO external boundary and no egress pathway: the published record is a legally mandated publication and the sealed production is a court-supervised transfer under an order, and both are drawn as the record stores they are rather than as data crossing a boundary without a guardrail.

• baseline

  WHERE THE LAB SHAPE DIVERGES FROM THE PAN SHAPE, and nothing is asserted here that the PAN file does not already record. Three divergences. First, PAN has no edge kind for a check at all, so two of its peer edges are REDRAWN here as checks — a monitoring finding reaching the rule owners, and fair-lending training reaching the deciding desk — because a channel that corrects a decision is inhibiting in the Lab's vocabulary and reinforcing in PAN's; their widths still come from PAN on the stated mapping. Two further checks have no PAN counterpart and are derived from the cited record: the expert findings reaching the forum, and the enterprise answer set against the operator's own stack. Second, PAN carries the court, the legislature and the regulators in its governance block as actors, two of them with narrow dynamical authority and one, the independent auditor, listed with none so that the gap is visible rather than silently filled. The Lab draws the examining channels as reviewers with their own reads and their own outbound pathways; the forum's pathway into the engine runs at zero, and the correspondence channel's only outbound pathway reaches the institution rather than the engine, which is the same statement in the Lab's vocabulary. Third, PAN carries the protective order in its narrative and its docket source rather than as a store; the Lab draws the sealed production as a record store, because a complete copy that exists and is readable only under an order is a record with an access property, and that access property is this deployment's subject.

• baseline

  BASELINES, and exactly how far the PAN org carries them. The PAN entry for this deployment holds thirty-three edges, one of the larger edge lists in this fleet. Every pathway drawn here that has a counterpart among them mirrors that edge's rate on a single rung mapping (0.50 and above to 3, 0.30 to 0.49 to 2, 0.06 to 0.29 to 1, below 0.06 or documented absent to 0) with no exceptions, including the two peer edges redrawn here as checks; the pathways without a counterpart are derived from the cited record directly and each says so on its own line, and no width moved when the drawing was coarsened. Four contrasts are load-bearing. The pathway from the published record into the decision runs at zero, so the only account of these decisions the outside world can read never touches the thing it is trying to explain. The widest read into an examining channel that has no power to compel runs into that same published record, while the read that carries everything runs into a store under a protective order. The pathway into the engine from the one examining channel that held the power to compel runs at zero, and the channel that held only the power to ask and to publish is drawn with no pathway into the engine at all: its demand, and the silence that followed it, ride on its one drawn route, into the institution. And the reconciliation of the sealed copy against the published counts runs at zero, which is not a repair that was skipped but a comparison no rule and no forum ever permitted.

• baseline

  DEMAND 3 / CAPACITY 2, and the pair is a reading rather than a default. Demand 3 rests on the public record's own scale statements: the court described the case as an attempt to sue about hundreds of thousands of home loan decisions at once, numerosity was conceded on the plaintiffs' estimate of at least 119,100 members for a class already narrowed to minority applicants approved by an enterprise system or by the operator's own scorecard and then denied, and the disputed year is a refinance-boom year whose underlying published data covers roughly eight million refinance applications across all lenders. Capacity 2 is deliberately held off both ends. It is not 3, because the operative complaint alleges processors carrying two to three times their historical monthly volume, staff terminated or departed and not replaced, and a systematic disincentive to check the work. It is not 1, because every word of that is a pleaded allegation in a case that ended without a merits ruling, and because the operator's sworn account, which the plaintiffs did not dispute, is that underwriting is done by human underwriters and the engine never supersedes their judgment, which is a real human comparator. No headcount, backlog, override rate or staffing figure for any class here appears in any adjudicated record.

• baseline

  EVIDENCE STATUS, labelled where it is used, because this record mixes tiers harder than most and the tiers are the case. PRIMARY COURT RECORD, and it is unusually load-bearing: a sixteen-page order reciting the system's architecture from the operator's own sworn declarations on facts the plaintiffs did not dispute, together with the operative pleading, the settlement notice, the dismissal order, the stipulation and the docket. PRIMARY GOVERNMENT: the consumer bureau's own report establishing that credit score and the fields naming the scoring model are excluded from the published loan-level data, and a Federal Reserve study using the confidential file that does contain it. CONGRESSIONAL PRIMARY: two March 2022 documents, one demanding the data and algorithms with a deadline and one asking two agencies to review. OPERATOR TIER: the annual report, the credit-programme announcement and its dollar figures, the response to the measurement, and the model-risk group's published paper. JOURNALISM, ALWAYS ATTRIBUTED: the 2020 refinance approval rates, which are carried only because two independent congressional documents restate them, and about whose method this board makes no claim at all because the article itself was not retrievable. DEFENCE-SIDE CONSULTING, LABELLED AS SUCH: the measured explanatory ceiling for models built on the published record, whose factual points are corroborated by the two government sources and whose framing is not neutral. PLEADING, ALLEGATIONS ONLY: everything in the operative complaint. No parameter on this diagram is set by the journalist's figure or by the operator's figure; both are recorded description.

• baseline

  THE STRUCTURAL FACT THIS BOARD IS DRAWN AROUND, stated as a property of the record rather than as a measurement. The evidence that would settle causation sits on both sides of a wall, and there are two walls rather than one. The published record is missing the applicant credit score by rule, so no analysis built on it can control for the variable the operator says explains the gap, and its measured ceiling is real and independent of this dispute: models built on that record explained at most about fifty-six per cent of ten large lenders' decisions, and even the confidential version that carries the score leaves most of the variation unexplained. The complete record and the rule base exist and are complete, and they reached a court file only under a stipulated protective order, in a proceeding that stopped before the merits with its briefing extensively sealed. Add the allocation of burden and the shape closes: the party seeking the explanation must show robust causality, so the operator's own inability to identify what drove the disparity counted against the claimants rather than against the operator, and the system's own articulation became the defence when the court held that thirteen hundred attributes under fourteen thousand rules showed commonality is not at all obvious. Nothing on this diagram computes a harm to any applicant, and nothing could.

• assumed

  Applicants are not in the dynamics. No approval, denial, price, rate or household outcome for any person is computed from anything drawn here, and no score over any person is authored anywhere in this network. The 2020 refinance approval rates, the class-size estimate, the rule and attribute counts, the workload figures and the remediation dollar figures are recorded external observations and they set no parameter. Two conflation guards ride with them. The seventy-two per cent is THIS operator's own white-applicant approval rate for 2020 refinance applications and is not the same quantity as the roughly seventy to seventy-one per cent Black-applicant approval rate reported across all other lenders in the same coverage; the two appear together in the sources and are frequently confused. And the litigation drawn here is the 2022 to 2026 underwriting-approvals matter: a differently captioned 2011 case and a 2012 federal settlement at the same operator concerned discretionary pricing and subprime steering, and the 2011 case appears in the 2025 order as precedent against certification rather than as a related claim.

What this example does not show

• Litigation posture, carried exactly as the dossier states it. In re Wells Fargo Mortgage Discrimination Litigation, No. 3:22-cv-00990-JD (N.D. Cal., Judge James Donato), consolidating six private actions on 18 January 2023 with a nationwide putative class pleaded under the Equal Credit Opportunity Act, the Fair Housing Act, 42 U.S.C. section 1981 and California statutes. There was NEVER a motion to dismiss: Wells Fargo Bank, N.A. answered on 17 May 2023 and the case went straight to discovery and Rule 23 briefing. On 5 August 2025 class certification was DENIED for failure of Rule 23(a)(2) commonality; both petitions for permission to appeal were denied by the Ninth Circuit on 8 January 2026; seven of the eight named plaintiffs settled confidentially and were dismissed WITH PREJUDICE between 7 May and 1 June 2026; interim lead counsel was relieved; and the last remaining individual claim was stayed on 28 July 2026 through 25 September 2026, with the court noting that further requests to continue the stay are not likely to be granted. The pending summary-judgment motion was never decided. The correct current posture is: effectively resolved without any adjudication of the merits, no class ever certified, and no finding of discrimination by any court or regulator. Do not imply that the settlements admit or establish anything; the notice of settlement extinguishes claims without any concession, the terms are confidential and no amount is public.
• NO FINDING OF DISCRIMINATION EXISTS, and nothing on this board should be read as one. No court and no regulator has found that this bank discriminated in refinance underwriting. The digital-redlining theory, the allegation that race was imputed from surname and geocoded location, the uncorrected-appraisal allegation, the credit-score-overlay allegation, the workflow decision states, the intake platform, the processor workload and disincentive figures, the monthly internal lending-breakdown report and the nine-months-then-approved-the-next-day account are ALL allegations from the operative complaint. None was adjudicated and none was conceded.
• The system is not what its popular description says it is, and the difference is load-bearing. On the undisputed facts recited in the class-certification order from the bank's own sworn declarations, the front-end workflow tool 'is not an automated underwriting system and it does not make lending decisions or calculate or assign Credit Risk Classes'; a separate business-rule engine assigns those classes; and the two scorecard models are, per the declaration, each 'a simple scorecard whose logic can be fully stated on 2-3 sheets of paper', neither of which 'utilizes artificial intelligence', each 'simple enough that an applicant's score could be calculated by hand'. The bank's position, which the plaintiffs did not dispute, is that these are distinct applications and that the combined artefact named in the pleading does not exist. Do not describe this deployment as an artificial-intelligence or machine-learning underwriting model.
• The 2020 approval rates are JOURNALISM, always attributed. Wells Fargo approved 47 per cent of Black homeowners' completed 2020 refinance applications, 53 per cent of Hispanic or Latino homeowners' and 72 per cent of white homeowners', the largest racial gap among major lenders and the only major lender to reject more Black refinance applicants than it approved. Those figures come from a news analysis of federal mortgage disclosure data and are carried here only because two independent congressional documents restate them; the article itself was not retrievable for verification and its methodology note could not be read, so this board makes NO claim about what the analysis controlled for. The bank did not dispute the arithmetic of the counts and attributed the gap to 'additional, legitimate, credit-related factors' including credit scores, home appraisals and broader economic inequity, with a spokesman saying the analysis was 'designed to present a skewed picture of our lending efforts'. That attribution has never been tested by any adjudicator.
• CONFLATION GUARD on the headline numbers. The 72 per cent is THIS bank's own white-applicant approval rate for 2020 refinance applications. It is NOT the same quantity as the roughly 70 to 71 per cent Black-applicant approval rate reported across all OTHER lenders in the same coverage, which also appears in the sources and differs slightly between the two congressional restatements. The two are frequently confused. The cross-lender comparison figures describe the market rather than this deployment and are deliberately not carried as this board's numbers.
• The expert exchange is the centre of this file and it is SEALED. The plaintiffs' expert ran a regression controlling for key underwriting factors and found statistically significant approval-rate disparities favouring white applicants that 'cannot be explained by legitimate underwriting factors'; at a concurrent expert hearing the bank's expert 'did not substantially disagree with the logistic analysis ... or with her findings of a statistical disparity along racial lines', objecting that without individual file review one cannot tell whether a denial was wrongful and that 'you will almost always find an underwriting disparity, on average, for minorities'. Neither report, neither specification and none of the underlying data is public. No point estimate from either side is carried on this board, because a number from a sealed report is a number nobody can check.
• The published-data ceiling is measured and is independent of this dispute. Credit score and the free-form fields naming the scoring model are excluded by rule from the public loan-level disclosure data, so no public-data analysis of any lender can control for the variable this bank says explains the gap. On the confidential regulator-held file that DOES contain credit score, a nine-million-application study reports explanatory power of about 39.8 per cent. A 2025 defence-side analysis of ten large lenders' 2024 filings found public-file models never exceeding about 56 per cent. That defence-side work is consulting output by a firm serving defendants in fair-lending litigation; its factual points are corroborated by the consumer bureau's own report and by the Federal Reserve study, and its framing is not neutral.
• The independent literature does not resolve this in either direction and is carried because omitting it would imply that it does. A Federal Reserve study using the confidential file finds that minority applicants have significantly lower credit scores and higher leverage, are less likely to receive approval from race-blind government automated underwriting systems, that observable applicant-risk factors explain most racial disparities in lender denials, and that differential treatment has played a limited role in recent denial disparities. Other peer-reviewed work finds algorithmic lenders discriminating less than face-to-face lenders in pricing while Latinx and Black borrowers still pay measurably more. Neither speaks to any single lender's internal system, and the plaintiffs themselves cited the first of them in the operative complaint.
• VERIFIED ABSENCE of any regulator action, stated as an absence. No publicly announced investigation, enforcement action or finding by the consumer bureau, the housing department, the justice department or the prudential supervisor concerning this bank's refinance underwriting was located, and the bank's most recent annual report discloses this private litigation and no government proceeding about mortgage-lending discrimination while disclosing several unrelated government matters. Two near-misses must not be conflated with it: the large December 2022 consumer-bureau consent order at the same bank, terminated in early 2025, covered auto lending, deposit accounts and mortgage rate-lock extension fees and NOT refinance underwriting; and a December 2023 press account of a supervisory matter requiring attention concerns discretionary mortgage PRICING exceptions rather than underwriting approval, and could not be verified at source.
• Do not confuse the two mortgage matters at this operator. This board draws the 2022 to 2026 underwriting-approvals litigation. A differently captioned 2011 case and a 2012 federal settlement at the same bank concerned discretionary PRICING and subprime steering; the 2011 case appears in the 2025 order as precedent against certification, not as a related claim.
• The bank's remediation is carried and is not erased: a $150 million special purpose credit programme announced on 13 April 2022 to lower rates and refinancing costs for Black homeowners it already serviced, plus $60 million in grants aimed at roughly 40,000 homeowners of colour in eight markets through 2025. Those are operator-reported and unaudited figures for a programme announced one month after publication, conceding no disparity in underwriting, and no change to any underwriting rule was announced alongside them. They are described here and they set no parameter on this board.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• The public class-certification order of 5 August 2025 in In re Wells Fargo Mortgage Discrimination Litigation recites, on facts the plaintiffs did not dispute, a three-part home-lending stack that is not an artificial-intelligence system. CORE is a front-end workflow tool that 'guides the flow of the loan origination process' and 'retrieves, stores, and displays' application data, Risk Engine outputs, and final lending decisions, and it 'is not an automated underwriting system and it does not make lending decisions or calculate or assign Credit Risk Classes.' A separate application, the Risk Engine, assigns Credit Risk Classes using business rules, business services, and data attributes; it 'currently executes tens of thousands of separately identifiable business rules' grouped into 16 business services, the largest of which, Get Risk Decision, 'contains 14,000 rules' and '1,300+ data attributes.' ECS comprises exactly two scorecard models, 11419 for government loans and 11960 for conventional loans, each of which per the bank's declaration 'is a simple scorecard whose logic can be fully stated on 2-3 sheets of paper,' 'neither utilizes artificial intelligence,' and each is 'simple enough that an applicant's score could be calculated by hand'; they map credit-bureau attributes to a score and thence to a Credit Risk Class and also generate risk insight messages displayed to underwriters. Wells Fargo's position, accepted as undisputed for certification purposes, is that these are distinct applications and that 'there is no such thing as CORE/ECS.' Underwriters 'are instructed that the Risk Engine never supersedes the judgment of the underwriter' and '[u]nderwriting is done by human underwriters.' Separately, Bloomberg News reported on 11 March 2022, from an analysis of Home Mortgage Disclosure Act data covering roughly 8 million 2020 refinance applications, that Wells Fargo approved 47 per cent of Black homeowners' completed 2020 refinance applications, 53 per cent of Hispanic or Latino homeowners', and 72 per cent of white homeowners' — the largest racial gap among major lenders and the only major lender to reject more Black refinance applicants than it approved. Those figures are carried here because a Senate letter of 16 March 2022 and a Senate Banking Committee release of 17 March 2022 independently restate them; the article itself was not retrievable for verification and no claim is made about what its analysis controlled for. Wells Fargo did not dispute the arithmetic of the counts, attributed the gap to 'additional, legitimate, credit-related factors' including credit scores, home appraisals, and broader economic inequity, and a spokesman said the analysis was 'designed to present a skewed picture of our lending efforts.' That attribution has never been tested by any adjudicator, and no court and no regulator has ever found that Wells Fargo discriminated in refinance underwriting.

  empirical
  • Government United States District Court for the Northern District of California (2025, August 5). Order Re Class Certification, In re Wells Fargo Mortgage Discrimination Litigation, No. 22-cv-00990-JD (Dkt. 328), via the RECAP Archive https://storage.courtlistener.com/recap/gov.uscourts.cand.391955/gov.uscourts.cand.391955.328.0.pdf
  • Investigative Donnan, S., Choi, A., Levitt, H., & Cannon, C. (2022, March 11). Wells Fargo Rejected Half Its Black Applicants in Mortgage Refinancing Boom. Bloomberg News (article not directly readable from this environment; figures corroborated through two congressional documents rather than read at source) https://www.bloomberg.com/graphics/2022-wells-fargo-black-home-loan-refinancing/
  • Government Warren, E., & Wyden, R. (2022, March 16). Letter from Senators Elizabeth Warren and Ron Wyden to Wells Fargo CEO Charles Scharf regarding mortgage-refinancing discrimination https://www.warren.senate.gov/imo/media/doc/2022.3.16%20Letter%20to%20Wells%20Fargo%20on%20Refinancing%20Discrimination.pdf
  • Government U.S. Senate Committee on Banking, Housing, and Urban Affairs (2022, March 17). Brown, Colleagues Call for Review of Wells Fargo Refinancing Process (with the letter to the Secretary of Housing and Urban Development and the Director of the Consumer Financial Protection Bureau) https://www.banking.senate.gov/newsroom/majority/brown-colleagues-review-wells-fargo-refinancing-process
  • Trade press Fortune (2022, March 16). Wells Fargo rejected half of all Black homeowners' refinancing applications https://fortune.com/2022/03/16/wells-fargo-rejected-black-homeowners-refinancing-applications

• The evidence that could settle causation in this matter sits on both sides of a wall, and there are two walls. On the public side, the Consumer Financial Protection Bureau's March 2023 report on the Home Mortgage Disclosure Act rule states that 'Credit score and free form text fields used to report the name and version of credit scoring models are excluded from the public loan-level data,' so no public-data analysis of any lender can control for the one variable this bank says explains the gap. The explanatory ceiling on that public record is measured rather than argued: a 2025 defence-side analysis of ten large lenders' 2024 filings found public-data models never exceeding about 56 per cent explanatory power and argued that sampled manual loan-file review remains necessary to confirm any statistical indication, matching the interagency fair-lending examination procedure in which regression is the scoping step and file review is the confirmation step; and on the confidential regulator-held version that DOES contain credit score, a Federal Reserve study of roughly nine million applications reports explanatory power of about 39.8 per cent, leaving most of the variation in loan outcomes unexplained. That defence-side analysis is consulting work published by a firm that serves defendants in fair-lending litigation; its factual points are corroborated by the two government sources and its framing is not neutral. On the private side, the internal rule base and the two scorecards were produced only in discovery under a stipulated protective order, and the class-certification and summary-judgment records are extensively sealed, in a proceeding that stopped before any merits ruling. The consequence is structural rather than rhetorical: an analysis built on the published record can be answered, accurately and unfalsifiably, by pointing at the variable it could not observe, and that answer cannot itself be checked, because the object it rests on is under seal.

  empirical
  • Government Consumer Financial Protection Bureau (2023, March). Report on the Home Mortgage Disclosure Act Rule Voluntary Review https://files.consumerfinance.gov/f/documents/cfpb_hmda-voluntary-review_2023-03.pdf
  • Government Bhutta, N., Hizmo, A., & Ringo, D. (2022). How Much Does Racial Bias Affect Mortgage Lending? Evidence from Human and Algorithmic Credit Decisions (Finance and Economics Discussion Series 2022-067, Board of Governors of the Federal Reserve System) https://www.federalreserve.gov/econres/feds/how-much-does-racial-bias-affect-mortgage-lending.htm
  • Trade press Chernin, A., Oka, S., & Oswald, K., Cornerstone Research (2025, November 19). When Mortgage Data Can't Prove Discriminatory Lending (defence-side expert analysis; framing not neutral) https://www.cornerstone.com/wp-content/uploads/2025/11/When-Mortgage-Data-Cant-Prove-Discriminatory-Lending.pdf
  • Government In re Wells Fargo Mortgage Discrimination Litigation, No. 3:22-cv-00990-JD (N.D. Cal.), docket entries 1 to 348 through July 28, 2026, via CourtListener and the RECAP Archive (docket 63052766) https://www.courtlistener.com/docket/63052766/in-re-wells-fargo-mortgage-discrimination-litigation/

• The two testifying statisticians converged on the measurement and diverged on causation, and the case turned on that gap rather than on either analysis. As recited in the class-certification order, the plaintiffs' expert Dr. Amanda Kurzendoerfer ran a regression controlling for key underwriting factors and found statistically significant approval-rate disparities favouring white applicants that 'cannot be explained by legitimate underwriting factors.' At a concurrent expert hearing held on 12 February 2025, Wells Fargo's expert Dr. Marsha J. Courchane 'did not substantially disagree with the logistic analysis ... or with her findings of a statistical disparity along racial lines'; her objection was causal, that without individual file review one cannot tell whether a denial was wrongful, and that 'you will almost always find an underwriting disparity, on average, for minorities.' Neither expert report, neither specification, and none of the underlying data is public, because the expert record is sealed. The burden allocation is the other half of the shape. The order records the plaintiffs' characterisation of Wells Fargo's own documents as showing that it 'could not itself identify what was driving the disparity' and that it 'found that there were three potential proxies for race in its ECS Model — major derogatories, average months in file, and recent inquiries — which could be causing the disparity'; the court held that this cut AGAINST the plaintiffs, because the burden of showing class-wide causation was theirs. In May 2021, six members of Wells Fargo's own Corporate Model Risk group had published a paper warning that historical data skew and automated feature engineering can 'miss the potential for correlated surrogate variables causing proxy discrimination,' that black-box algorithms have 'potential for serious harm' in consumer lending, and that models 'must be continually monitored for disparate impact testing'; the plaintiffs quoted that paper back at the bank in the operative complaint. Every allegation in that complaint — the 'digital redlining' theory, race imputation by surname and geocoded location, uncorrected appraisal values, credit-score overlays above agency minimums, the A1/A2/C1/C2 decision states, the point-of-sale intake platform, the processor workloads rising from about 30 applications a month to more than 50 and sometimes nearly 100, the systematic disincentive to check the work, the monthly internal report circulating the racial breakdown of lending, and the account of nine months of lost paperwork followed by approval the day after a federal housing agency was notified — is an allegation that was never adjudicated and never conceded.

  empirical
  • Government United States District Court for the Northern District of California (2025, August 5). Order Re Class Certification, In re Wells Fargo Mortgage Discrimination Litigation, No. 22-cv-00990-JD (Dkt. 328), via the RECAP Archive https://storage.courtlistener.com/recap/gov.uscourts.cand.391955/gov.uscourts.cand.391955.328.0.pdf
  • Vendor Zhou, N., Zhang, Z., Nair, V. N., Singhal, H., Chen, J., & Sudjianto, A., Corporate Model Risk, Wells Fargo (2021, May). Bias, Fairness, and Accountability with AI and ML Algorithms (operator-authored technical paper) https://arxiv.org/abs/2105.06558
  • Government In re Wells Fargo Mortgage Discrimination Litigation, No. 3:22-cv-00990-JD (N.D. Cal., filed March 24, 2023). Amended and Consolidated Class Action Complaint (Dkt. 114) (pleading; allegations only), via the RECAP Archive https://storage.courtlistener.com/recap/gov.uscourts.cand.391955/gov.uscourts.cand.391955.114.0.pdf

• Six private actions filed in early 2022 were consolidated on 18 January 2023 before Judge James Donato as In re Wells Fargo Mortgage Discrimination Litigation, No. 3:22-cv-00990-JD (N.D. Cal.), pleading a nationwide putative class under the Equal Credit Opportunity Act, the Fair Housing Act, 42 U.S.C. section 1981, California's Unruh Civil Rights Act, and the California Unfair Competition Law. There was never a motion to dismiss in the consolidated case: Wells Fargo Bank, N.A. answered the Amended and Consolidated Class Action Complaint on 17 May 2023 and the case went straight to discovery, expert work, and Rule 23 briefing. On 5 August 2025 class certification was DENIED for failure of Rule 23(a)(2) commonality. Numerosity had been conceded on the plaintiffs' estimate of 'at least 119,100' members for classes narrowed to minority applicants approved by an external automated underwriting system or by Wells Fargo's own ECS and ultimately denied. The court wrote that the plaintiffs 'did not present any classwide evidence whatsoever of robust causality' and had 'focused like a laser on the statistical disparity in application denial rates, without anything in the way of explanatory factors,' and that 'the fact that Wells Fargo potentially analyzed 1,300+ data attributes pursuant to 14,000 rules indicates that commonality with respect to denial of a mortgage is not at all obvious here'; it relied on Wal-Mart Stores v. Dukes and treated the undisputed presence of human underwriter discretion as bringing the case within the line of authority refusing certification of discretionary-policy lending classes. It reached no other Rule 23 element and made no merits finding. Both petitions for permission to appeal, Ninth Circuit Nos. 25-5262 and 25-5267, were denied on 8 January 2026. The case then resolved without any ruling: a Notice of Settlement was filed on 7 May 2026 covering four named plaintiffs and any parties represented by that counsel who were potential class members, the court granted dismissal with prejudice and relieved interim lead counsel on 14 May 2026, three further plaintiffs were dismissed with prejudice between 8 May and 1 June 2026, and the last remaining individual claim was stayed on 28 July 2026 through 25 September 2026 with the court noting that further requests to continue the stay are not likely to be granted. Settlement terms are confidential and no amount is public; the notice extinguishes claims without any concession. The pending summary-judgment motion was never decided. No publicly announced investigation, enforcement action, or finding by any federal regulator concerning Wells Fargo refinance underwriting discrimination has been located, and the bank's Form 10-K for fiscal 2025 records that 'In August 2025, the district court denied class certification and plaintiffs' interlocutory appeal of the decision was denied in January 2026' while disclosing no government proceeding or investigation concerning mortgage-lending discrimination and disclosing several unrelated government matters. One month after publication, on 13 April 2022, Wells Fargo announced a $150 million Special Purpose Credit Program to lower rates and refinancing costs for Black homeowners it already serviced plus $60 million in WORTH grants aimed at roughly 40,000 homeowners of colour in eight markets through 2025; those are operator-reported, unaudited figures for a programme that concedes no disparity in underwriting.

  empirical
  • Government United States District Court for the Northern District of California (2025, August 5). Order Re Class Certification, In re Wells Fargo Mortgage Discrimination Litigation, No. 22-cv-00990-JD (Dkt. 328), via the RECAP Archive https://storage.courtlistener.com/recap/gov.uscourts.cand.391955/gov.uscourts.cand.391955.328.0.pdf
  • Government In re Wells Fargo Mortgage Discrimination Litigation, No. 3:22-cv-00990-JD (N.D. Cal.), docket entries 1 to 348 through July 28, 2026, via CourtListener and the RECAP Archive (docket 63052766) https://www.courtlistener.com/docket/63052766/in-re-wells-fargo-mortgage-discrimination-litigation/
  • Government In re Wells Fargo Mortgage Discrimination Litigation (N.D. Cal., May 7, 2026). Notice of Settlement; Request for Dismissal Pursuant to Federal Rule of Civil Procedure 41; and Request to be Relieved as Interim Lead Counsel (Dkt. 337), via the RECAP Archive https://storage.courtlistener.com/recap/gov.uscourts.cand.391955/gov.uscourts.cand.391955.337.0.pdf
  • Government United States District Court for the Northern District of California (2026, May 14). Order Granting Request for Dismissal Pursuant to Federal Rule of Civil Procedure 41; and Request to be Relieved as Interim Lead Counsel (Dkt. 342), via the RECAP Archive https://storage.courtlistener.com/recap/gov.uscourts.cand.391955/gov.uscourts.cand.391955.342.0.pdf
  • Government In re Wells Fargo Mortgage Discrimination Litigation (N.D. Cal., May 7, 2026). Stipulation of Dismissal, With Prejudice, as to the Claims of Plaintiffs Aaron Braxton, Paul Martin, Gia Gray and Bryan Brown (Dkt. 338), via the RECAP Archive https://storage.courtlistener.com/recap/gov.uscourts.cand.391955/gov.uscourts.cand.391955.338.0.pdf
  • Vendor Wells Fargo & Company (2026). Annual Report on Form 10-K for the fiscal year ended December 31, 2025, Note 12 (Legal Actions), Home Mortgage Lending Discrimination Litigation https://www.sec.gov/Archives/edgar/data/72971/000007297126000133/wfc-20251231.htm
  • Vendor Wells Fargo & Company (2022, April 13). Wells Fargo Expands Efforts to Advance Racial Equity in Homeownership (operator announcement; unaudited operator figures) https://newsroom.wf.com/news-releases/news-details/2022/Wells-Fargo-Expands-Efforts-to-Advance-Racial-Equity-in-Homeownership/default.aspx
  • Trade press ABA Banking Journal (2025, September). Class certification denied in mortgage discrimination lawsuit against Wells Fargo https://bankingjournal.aba.com/2025/09/class-certification-denied-in-mortgage-discrimination-lawsuit-against-wells-fargo/
  • Trade press Banking Dive (2025, August). Wells Fargo won't face class-action mortgage discrimination lawsuit https://www.bankingdive.com/news/wells-fargo-wont-face-class-action-mortgage-discrimination-lawsuit-digital-redlining/757091/

Where this connects

Institutional pressures in this domain

• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.

All of them in context on the Lending & credit collections AI domain page.

Levers available here and the patterns behind them

• Upgrade model — Improve the model
• Check with a second model — Cross-model verification
• Check copied records — Reconcile copied records
• Understand the system — Understand the system
• Review the riskiest first — Risk-tiered oversight
• Review on schedule — Oversight cadence & retrospectives
• Mark AI-written records — Provenance labeling
• Peer sharing rules — Peer-edge governance

Documented case histories

• Wells Fargo refinance underwriting & the bridge nobody could build
• Automated underwriting with its fair-lending testing on the record
• Cleared on the numbers but faulted on the explanation
• The governance an enforcement action had to write
• M-Shwari & Kenya's Digital Credit Market
• Citi Retail Services Judgmental Review & the Armenian surname screen
• Santander Consumer USA subprime vehicle loan scoring
• Credit Acceptance Corporation's net-collections score
• Navy Federal mortgage underwriting & three readings of one gap
• Enova International servicing defects & the debits nobody authorised
• Equifax Online Model Server coding error (2022)
• TransUnion's OFAC Name Screen & the people who could not sue
• Dave ExtraCash: an advertised ceiling, an automated amount, and a case that never asks how the amount is set
• Hello Digit's automated-savings algorithm
• Oportun's legal-collections filing pipeline