Credit Acceptance's net-collections Score, inside CAPS

What this models

This example runs on the Net-collections-class subprime vehicle lending network: 12 components and 22 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: 13 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  The mechanics of this deployment are stated on this board without hedging, and everything else is stated as allegation. That split is not an editorial choice: both parties describe the governed joint the same way. The complaint pleads that the operator scores the proposed transaction and returns an estimate of the percentage of total amounts owed it expects to collect, that the estimate spans post-repossession auction proceeds, post-default recoveries, deficiency judgments and garnishment, that the estimate prices the payment to the dealership, that it does not decide approval, that it does not set the borrower's interest rate, and that the borrower is never shown it. The defendant's own revised dismissal memorandum describes the same system as an internal scoring metric used to calculate the payment to the dealer, and says in terms that it does not affect the terms of a contract. Everything characterising knowledge, intent, incentive or practice is a pleaded allegation, attributed on every surface, and the operator denied all of it and moved to dismiss the complaint in its entirety.

• baseline

  Nothing in this record was adjudicated, and the board is built in that register throughout. The fully briefed motion to dismiss was terminated unruled on 6 February 2026 so the parties could settle; the action was dismissed and discontinued without prejudice on 5 June 2026 on the parties' report that all claims had been settled in principle, with a sixty-day right to reopen; on 23 July 2026 the parties reported agreement on all terms and obtained a forty-five day extension of that deadline. As of the docket's last update no executed settlement, consent judgment or public statement of terms appears. The federal plaintiff was dropped from the case by its own consented motion under Rule 21, granted 29 April 2025, on a supporting memorandum that gives no substantive or merits-based reason and makes no representation about the merits of the withdrawn claims. That is a litigation decision recorded on a docket and is described here as nothing more.

• baseline

  Two forecast figures in this record are easy to conflate and this board keeps them apart. The pleaded portfolio-level average was 64 nationwide and 66 in New York — an expectation of collecting roughly 64 to 66 cents per dollar of TOTAL AMOUNTS OWED, which includes interest. The sharper and load-bearing allegation is different: that on more than 39 per cent of loans nationwide and about 25 per cent of New York loans the projection fell below the AMOUNT FINANCED, the stated principal. The second figure is the one the case turns on, and neither is converted into the other anywhere on this board.

• baseline

  The Score is not an approval gate and no copy on this board says or implies that it screened, declined or denied anyone. Both parties agree it prices the payment to the dealership and decides nothing about the applicant. That is why this deployment has no decision node, no adverse-action pathway and no contest channel: the harm the plaintiffs allege is that an accurate forecast of non-repayment was acted on commercially and was never disclosed, not that a model rejected somebody. It is also why this record carries no protected-class frame — there is no disparate-impact or fair-lending claim in this case, and none is imported here.

• baseline

  Demand 3 and manual capacity 1 come from the operator's own reporting and from an absence the record states rather than a measurement it publishes. Demand: approximately 1.9 million consumers obtained loans through this operator and its affiliated dealers across the pleaded period, the network exceeded 12,000 affiliated dealerships, and the operator reported 11,004 active dealers, 1,456 new enrolments and 84,615 unit assignments in a single quarter of 2026 against an $8.0 billion average portfolio. Capacity: there is no per-application human underwriter, no exception desk and no manual review queue anywhere in this deployment, and the only documented human judgment inside the origination path is a manager's authority to approve a monthly payment above a quarter of gross monthly income, up to 30 per cent — an authority that widens a bound rather than performing the work. No source compares this deployment against its own human counterfactual and nothing here computes one.

• baseline

  The two reconciliations on this board are drawn at intensity zero because the record documents them as absent rather than as narrow, and each carries the operator's contrary position beside the allegation. The printed disclosure set against the deal record: the plaintiffs' whole hidden-finance-charge construction is that the two do not meet, and three state-law counts plead as a failure to disclose that the borrower was not told what the forecast projected; the defendant answered that the cash-price proxy behind that construction was invented for the litigation and that the plaintiffs had abandoned any standalone claim over non-disclosure of the Score, and the court never resolved either point. The recovery apparatus against the ledger: the plaintiffs plead more than 7,000 New York default judgments across 2017 and 2018 in a window when 40 of the thousands of borrowers sued had counsel, and a default judgment is by construction entered on a claim nobody examined. Neither zero asserts that a check is impossible; each records that the cited record documents no such step.

• baseline

  The widths on this board's pathways come from the PAN org's own estimated widths, mapped on one stated rung scale with no exceptions, and every PAN edge for this deployment is marked estimated. Where one pathway here stands for several flows the PAN file draws separately, it keeps the rung of the flow it is named for and its own description carries the others. No source in this record publishes a per-decision defect rate, an override rate, a correction rate or a record-hygiene measure for any store, and none is invented here. The one quantitative series the operator does publish is a calibration record on its own recovery — each assignment year's current forecast collection percentage against the percentage forecast at assignment — and it is never treated as a deployment error rate, because a forecast-versus-outcome gap on collections is not a rate at which anything went wrong for anyone.

• baseline

  Eight pathways here have no PAN counterpart and each is derived from the cited record directly. The disclosure store and its reconciliation against the deal record exist because PAN folds the printed contract and disclosures into the transaction record while the whole governed surface in this case is the difference between the two. The enforcement component and its two pathways exist because the Lab vocabulary has a kind for a downstream action system a record drives and this deployment plainly has one. The read the investigating authorities obtained and the check they run exist because PAN carries that party in its governance block rather than as edges. And three couplings exist because PAN has no edge kind for them at all: a bounded automated parameter screen, one model pricing another model's display, and an operator function tuning the model it owns. Nothing is asserted on the Lab side that the PAN file or the cited record does not already carry.

• baseline

  This board draws the deployment at the coarsest granularity at which every documented mechanism stays distinguishable, so three parties the record names separately are drawn inside a neighbour. The published vehicle book values sit inside the deal record, because nothing documented acts on the valuation tables themselves; the stated price bound written against them stays on the parameter screen, and the plaintiffs' price comparison against them is carried in the deal record's description. The securitisation and funding desk sits inside the pricing function, because both read the operator's own collection performance and the record shows neither acting on it differently for any borrower; its securitisation figures, its servicing fees and the securities-fraud count pleaded over those structures are carried in that function's description. The outside collection attorneys sit inside the servicing arm, because they act for the operator under a referral and work from that arm's own ledger; the default judgments they obtained stay drawn on the reconciliation of what is enforced against the ledger. Where the record describes one flow along two routes, it is drawn once, and the surviving pathway's description carries what the other said.

• baseline

  The bounded parameter screen on this board is the operator's own stated policy and is not a remedy. Nothing in this record orders anything: no consent judgment was entered, no findings were made, no practice change was compelled, and the one settlement that did land against this operator on materially the same theories is a separate Massachusetts action of 1 September 2021 for $27.2 million covering more than 3,000 borrowers, which resolved allegations without any admission and whose required practice changes were not enumerated publicly. A board drawing an ordered gate here would be inventing a control, and this one draws none.

• baseline

  There is no off-network sink on this board and that is a derivation rather than an omission. The obvious candidate is the securitisation leg — roughly $7.34 billion of notes across the shelf deals the plaintiffs listed, and a disclosed August 2026 financing conveying approximately $750.2 million of consumer loans — but a trust holding a receivable is a FUNDING structure, and the Lab's boundary kind and its three egress edges describe client data crossing out of a governed system to an ungoverned consumer of it. Drawing a conveyance of loans as egress would put a data-exfiltration reading on a funding fact. The referrals to outside counsel are the other candidate and are equally not egress: counsel act for the operator under a referral, which is why the PAN carrier draws them as part of the operator network. The consequence is stated rather than hidden: with no external boundary the Privacy gauge does not run on this board, so the privacy posture recorded for the scenario documents the deployment without moving a number, and the exposure it records is carried on the pathways instead.

• baseline

  The people this deployment serves are not in the dynamics and no outcome for any of them is computed from anything drawn here. No credit decision, price, advance, default, repossession, auction result, deficiency or judgment for any person is derived from this network, and no score over any person is authored anywhere in it. The delinquency, repossession, auction, referral and judgment statistics are recorded external observations from a pleading, and the operating and calibration figures are recorded external observations from the operator's own filings. The complaint's exemplar borrower is anonymized in the pleading and stays anonymized here.

• baseline

  One current automation disclosure sits on this board and is fenced off from every allegation in the case. The operator reported that an AI-enabled call agent handled 67 per cent of its inbound customer-service and account-solutions calls in June 2026, up from 27 per cent in March 2026, and that the capability is integrated into core servicing workflows. That is an operator disclosure from 2026 about a channel that postdates the pleaded conduct by years, it is recorded in one node's copy and nowhere else, it carries no pathway of its own, and it is never described as part of the conduct at issue.

What this example does not show

• LITIGATION AND REGULATORY POSTURE, verbatim from the evidence dossier and load-bearing. Resolved by settlement-in-principle without adjudication. The case was DISMISSED and discontinued without prejudice on June 5, 2026 with a right to reopen; the operator's SEC filing of Aug 4, 2026 states the parties have agreed on all terms and are executing documentation, with a potential combined cash payment of $75.5 million covering both this action and a parallel 40-plus-state multistate investigation. No court ever ruled on the motion to dismiss (it was terminated unruled), so there are no findings of fact and no adjudicated liability — every allegation below remains an allegation.
• The $75.5 million figure is the operator's own disclosure of a POTENTIAL payment, it covers BOTH this New York action and a separate forty-plus-state multistate investigation, and as of the docket's last update it is not embodied in any executed agreement, consent judgment or public order. It is never reported here as a New York settlement amount and never as final. No dollar figure in this record is treated as the price of a governance instrument.
• The federal plaintiff's withdrawal is described only in the terms the record supports: on April 24, 2025 the Consumer Financial Protection Bureau filed a consented motion under Rule 21 to be dropped as a party and to withdraw its counsel's appearances, the defendant consented, New York did not object, and the court granted it on April 29, 2025, leaving New York as sole plaintiff. The Bureau's supporting memorandum gives no substantive or merits-based reason and makes no representation about the merits of the withdrawn claims. Contemporaneous trade reporting placed the withdrawal within a broader 2025 pattern of agency enforcement dismissals; that is press attribution rather than an agency statement, and no motive, instruction or internal reasoning beyond the filed record is asserted or implied anywhere on this board.
• The recomputed cost of credit is the plaintiffs' construction, always. Disclosed annual percentage rates averaged about 22 per cent nationwide, and New York contracts disclosed 22.99 or 23.99 per cent against a 25 per cent criminal usury cap. The plaintiffs then treated the difference between the disclosed deal cost excluding interest and the dealership's compensation as a concealed finance charge, and on that recomputation alleged that nearly 90 per cent of New York loans exceeded the 25 per cent cap, about 30 per cent exceeded 40 per cent, and the median recomputed New York rate was about 34 per cent. The defendant argued that the cash-price proxy behind the recomputation was invented for the litigation and is incompatible with the federal disclosure statute's defined terms. The court never resolved it.
• Two forecast figures are easy to conflate and mean different things. The portfolio-level expectation of collecting roughly 64 to 66 cents per dollar of TOTAL AMOUNTS OWED includes interest. The load-bearing allegation is the other one: that on more than 39 per cent of loans nationwide and about 25 per cent of New York loans, the projection fell below the AMOUNT FINANCED — the stated principal. This board keeps them apart and never converts one into the other.
• The Score is not an approval gate and this board never says it screened, declined or denied anyone. Both parties agree it prices the payment to the dealership, decides no application and sets no interest rate. The harm alleged is that an accurate forecast of non-repayment was acted on commercially and never disclosed. Relatedly, the complaint is framed around economic harm to a low-income borrower population and not around a protected class: there is no disparate-impact and no fair-lending claim in this case, and no discrimination frame is imported into it.
• The GPS starter-interrupt allegation is period-bound and is carried with its bound wherever it appears: it is pleaded as a policy of disabling the vehicle within days of a missed payment, and stated to have been used on New York-financed vehicles only until late 2018.
• The operator's AI-enabled call agent — 67 per cent of inbound customer-service and account-solutions calls in June 2026, up from 27 per cent in March — is a current operator disclosure entirely outside the litigation and unconnected to any allegation in it. It appears in one node's copy as what it is, it carries no pathway, it grounds no lever and no stressor, and it is never described as part of the conduct at issue, which predates it by years.
• The 2021 Massachusetts settlement resolved ALLEGATIONS and carried no admission of wrongdoing; the operator said the suit had been vigorously contested and that it looked forward to continuing to serve customers in the Commonwealth. It is cited here as a governed remedy that actually landed and as evidence that these theories are not novel, never as a finding that the conduct occurred. Its required practice changes were not enumerated publicly, which is why they are not priced anywhere on this board.
• The forecast-accuracy series the operator publishes is a calibration record on its own recovery and is never converted into a deployment error rate. Each assignment year's current forecast collection percentage is compared against the percentage forecast at assignment: 2020 running 4.7 points above, 2022 running 8.2 points below, 2025 and 2026 within a tenth of a point. A forecast-versus-outcome gap on collections is not a rate at which anything went wrong for anybody, and no source in this record publishes a per-decision defect rate, an override rate, a correction rate or an independent evaluation of this deployment.
• Borrowers are not modeled. No credit decision, price, advance, default, repossession, auction result, deficiency or judgment for any person is computed from anything on this diagram, and no score over any person is authored anywhere. The delinquency, repossession, auction, referral and judgment statistics are recorded external observations from a pleading, and the operating and calibration figures are recorded external observations from the operator's own filings. The complaint's exemplar borrower is anonymized in the pleading and stays anonymized here; no individual borrower, dealership, employee or judge is named on this board.
• No third-party model vendor appears anywhere in this record and none is invented. The operator builds and runs the system in-house and licenses it to the dealerships it enrols, so there is no model supplier to name, no vendor update to model and no external scoring product 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.

• Credit Acceptance Corporation is a publicly traded indirect subprime vehicle lender: it does not lend across its own counter but buys retail instalment contracts from dealerships enrolled in its program, which pay a monthly fee for access to its Credit Approval Processing System and its servicing. The pleaded scale is that approximately 1.9 million consumers obtained loans through the operator and its affiliated dealers between 2 November 2015 and 30 April 2021, that the network exceeded 12,000 affiliated dealerships, and that consumers obtained more than $4.9 billion in operator-financed loans in 2020 alone, with New York among its top five state markets. The operator's own quarterly report for the period ended 30 June 2026 gives the current shape: 11,004 active dealers in the quarter, a record and up 3.3 per cent year over year, with 1,456 new dealers enrolled; 84,615 consumer-loan unit assignments worth $1.0 billion in the quarter; an $8.0 billion average loan portfolio balance; $90.6 million of dealer holdback paid in the first half of 2026; and a $599 monthly per-dealer program fee. The pleaded borrower population had a median credit-bureau score of 546 and a gross annual income of approximately $35,000.

  empirical
  • Government Complaint, Consumer Financial Protection Bureau and the People of the State of New York v. Credit Acceptance Corporation, No. 1:23-cv-00038 (S.D.N.Y., filed 4 January 2023) https://files.consumerfinance.gov/f/documents/cfpb_credit-acceptance-corporation_complaint_2023-01.pdf
  • Government Credit Acceptance Corporation, Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (filed 4 August 2026), SEC EDGAR CIK 0000885550 https://www.sec.gov/Archives/edgar/data/885550/000088555026000182/cacc-20260630.htm

• The mechanics of the scored artifact are common ground between the parties. When a dealership submits an application through the operator's origination software, the operator scores the proposed transaction — the applicant's credit-bureau attributes, the application data, the deal structure of term, monthly payment, down payment and trade-in, and the vehicle — and returns a Score from 0 to 100 representing its estimate of the percentage of total amounts owed it expects to collect over the life of the contract, inclusive of scheduled payments, late fees, post-repossession auction proceeds, post-default collection recoveries, deficiency judgments, and wage garnishment. The Score prices the payment the operator makes to the dealership; it does not decide whether financing is offered; it does not set the borrower's interest rate; and it is never disclosed to the borrower. The defendant's own revised dismissal memorandum states that the Score 'does not affect the terms of a Contract' and describes it as 'an internal scoring metric that Credit Acceptance uses to calculate the CAC Payment to the Dealer'. The plaintiffs allege, and it is an allegation, that the pleaded portfolio-level forecast averaged 64 nationwide and 66 in New York — an expectation of collecting roughly 64 to 66 cents per dollar of total amounts owed, interest included — and that for more than 39 per cent of loans nationwide and about 25 per cent of New York loans the projection fell below the amount financed, which is the stated principal rather than the total of payments. They further allege that for New York loans originated between 2015 and 2020 whose projected collections were below the amount financed, nearly 70 per cent were sixty or more days past due, repossessed, or sold at auction. The operator denied the complaint in full and moved to dismiss it in its entirety, and no court ruled on any of it.

  empirical
  • Government Complaint, Consumer Financial Protection Bureau and the People of the State of New York v. Credit Acceptance Corporation, No. 1:23-cv-00038 (S.D.N.Y., filed 4 January 2023) https://files.consumerfinance.gov/f/documents/cfpb_credit-acceptance-corporation_complaint_2023-01.pdf
  • Government Memorandum of Law in Support of Credit Acceptance Corporation's Revised Motion to Dismiss, ECF 75 (S.D.N.Y., filed 14 August 2024) https://storage.courtlistener.com/recap/gov.uscourts.nysd.591873/gov.uscourts.nysd.591873.75.0.pdf

• The plaintiffs alleged that the pricing runs to the dealership rather than to the borrower. They pleaded that the operator offered a dealership on average about 72 per cent of the Score-derived projected net collections, that the nationwide average dealer payment was about 22 per cent less than the amount financed, and that recovering roughly 78 per cent of the amount financed therefore sufficed to exceed the cash the operator had actually put at risk — a gap they put at nearly $2,500 per New York loan. They alleged that the dealership's back-end share was rarely paid, with under 12 per cent of new loans nationwide in dealer pools receiving any earnout and total earnout under 2 per cent of the value received by New York dealers. They alleged that because the interest rate does not vary with borrower risk — New York contracts disclosed 22.99 or 23.99 per cent against a 25 per cent state criminal usury cap, and nationwide disclosed rates averaged about 22 per cent — the risk premium was relocated into the amount financed, where it is invisible to comparison shopping and accrues interest. They alleged that the operator's stated origination guardrail was proof of income together with a monthly payment not exceeding 25 per cent of gross monthly income, 30 per cent with manager approval, with no collection of recurring debt obligations, housing cost, food, healthcare, or childcare cost, no debt-to-income ratio, no residual-income calculation, and no adjustment for the number of dependents; that its stated pricing bound permitted a vehicle to be priced at up to 115 per cent of the highest published book value for that make and model without inspecting the vehicle's condition, against a measured median disclosed selling price about 77 per cent over wholesale book value and slightly more than 50 per cent over reported dealer cost; that 90 per cent of loans carried an operator-approved add-on product, at an average vehicle service contract retail cost of $1,545 and an average guaranteed-asset-protection cost of about $782, generating approximately $250 million of add-on revenue nationwide in 2020 and a flat dealership commission of about $385 per vehicle service contract; that the origination software displayed to the dealership, in real time, the projected dealership profit for each vehicle for that applicant and how it changed with each product added, in one pleaded screen moving from $1,605.31 to $4,051.82; and that the operator's internal dealer rating fed back into the size of the dealer advance. The plaintiffs' recomputation of the cost of credit — treating the difference between the disclosed deal cost excluding interest and the dealership's compensation as a concealed finance charge, and on that basis alleging that nearly 90 per cent of New York loans exceeded the 25 per cent cap with a median recomputed rate of about 34 per cent — is their construction: the defendant argued the cash-price proxy behind it was 'invented ... for this litigation' and incompatible with the Truth in Lending Act's defined terms, and the court never resolved it.

  empirical
  • Government Complaint, Consumer Financial Protection Bureau and the People of the State of New York v. Credit Acceptance Corporation, No. 1:23-cv-00038 (S.D.N.Y., filed 4 January 2023) https://files.consumerfinance.gov/f/documents/cfpb_credit-acceptance-corporation_complaint_2023-01.pdf
  • Government Memorandum of Law in Support of Credit Acceptance Corporation's Revised Motion to Dismiss, ECF 75 (S.D.N.Y., filed 14 August 2024) https://storage.courtlistener.com/recap/gov.uscourts.nysd.591873/gov.uscourts.nysd.591873.75.0.pdf
  • Government Office of the New York State Attorney General (2023, January 4). Attorney General James and CFPB Sue Auto Lender for Cheating Thousands of New Yorkers https://ag.ny.gov/press-release/2023/attorney-general-james-and-cfpb-sue-auto-lender-cheating-thousands-new-yorkers

• The information asymmetry is the pleaded governed surface, and it is manufactured inside one system. The plaintiffs alleged that consumers 'do not know about, and certainly do not have access to, the extensive predictive data' the operator was using, and pleaded in three state-law counts, as a failure to disclose, that the operator did not tell the borrower that its algorithms and the Score had forecast projected collections far below the total amounts owed under the loan agreement, and that the projection potentially included amounts collected through late fees, repossession, and wage garnishment. They alleged that the same software instance that shows the dealership a live per-vehicle profit projection then generates the borrower's contract and disclosures. The defendant answered that the plaintiffs had abandoned any standalone claim over non-disclosure of the Score, and that as an indirect lender it has no contact with the consumer until after the dealership's contract is executed and assigned to it. The disclosure regime either side points to — the federal Truth in Lending Act and New York's motor vehicle retail instalment regime — governs the contract form; no disclosure rule in this record requires telling a borrower what a lender's model predicts about them. There is no adverse-action notice, no explanation, no contest path, and no borrower-facing channel of any kind attaching to the forecast anywhere in the record. None of this was adjudicated.

  empirical
  • Government Complaint, Consumer Financial Protection Bureau and the People of the State of New York v. Credit Acceptance Corporation, No. 1:23-cv-00038 (S.D.N.Y., filed 4 January 2023) https://files.consumerfinance.gov/f/documents/cfpb_credit-acceptance-corporation_complaint_2023-01.pdf
  • Government Memorandum of Law in Support of Credit Acceptance Corporation's Revised Motion to Dismiss, ECF 75 (S.D.N.Y., filed 14 August 2024) https://storage.courtlistener.com/recap/gov.uscourts.nysd.591873/gov.uscourts.nysd.591873.75.0.pdf

• The plaintiffs pleaded a quantified recovery apparatus rather than a general complaint about collections. They alleged that within days of a missed payment the operator would, as a matter of policy, disable the vehicle through a GPS starter-interrupt device, a practice they state was used on New York-financed vehicles until late 2018; that contract terms ran 60 to 72 months while repossessed-and-resold vehicles averaged under two years from origination to auction with a quarter auctioned within one year; that a majority of the roughly 1.9 million contracts became delinquent at some point and more than half of borrowers were delinquent within the first year; that the operator repossessed more than a quarter of financed vehicles nationwide and resold about 20 per cent at auction, with approximately 44 per cent repossessed in New York and 21 per cent of repossessed New York vehicles repossessed more than once; that auction proceeds satisfied on average only 29 per cent of remaining amounts owed nationwide and under 28 per cent in New York, leaving an average post-auction deficiency of about $8,500; that the operator made more than 138,000 referrals to debt-collection attorneys nationwide and obtained judgments against more than one in six New York borrowers whose loans reached maturity by May 2021, with New York judgment counts running between roughly 2,500 and 4,200 a year from 2013 to 2019 and more than 7,000 default judgments obtained in 2017 and 2018 in a window when 40 of the thousands of New York borrowers sued had counsel; and that, excluding support personnel, the operator employed nearly twice as many people in servicing and collections as in origination. Separately from every allegation in the case, the operator's own quarterly report for the period ended 30 June 2026 discloses that an AI-enabled call-centre agent handled 67 per cent of inbound customer-service and account-solutions calls in June 2026, up from 27 per cent in March 2026, and that the capability is integrated into core servicing workflows; that is a 2026 disclosure about a channel postdating the pleaded conduct and is not part of it. The pleaded statistics were never adjudicated.

  empirical
  • Government Complaint, Consumer Financial Protection Bureau and the People of the State of New York v. Credit Acceptance Corporation, No. 1:23-cv-00038 (S.D.N.Y., filed 4 January 2023) https://files.consumerfinance.gov/f/documents/cfpb_credit-acceptance-corporation_complaint_2023-01.pdf
  • Government Office of the New York State Attorney General (2023, January 4). Attorney General James and CFPB Sue Auto Lender for Cheating Thousands of New Yorkers https://ag.ny.gov/press-release/2023/attorney-general-james-and-cfpb-sue-auto-lender-cheating-thousands-new-yorkers
  • Government Credit Acceptance Corporation, Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (filed 4 August 2026), SEC EDGAR CIK 0000885550 https://www.sec.gov/Archives/edgar/data/885550/000088555026000182/cacc-20260630.htm

• The operator publishes the forecast's own accuracy, which is unusual and is the most modelling-relevant artifact in this record. Its quarterly report for the period ended 30 June 2026 tabulates, for each assignment year, the current total-loan forecast collection percentage against the percentage forecast when the loans were assigned: 2017 at 64.8 against 64.0 initial; 2018 at 65.6 against 63.6; 2019 at 67.3 against 64.0; 2020 at 68.1 against 63.4; 2021 at 64.1 against 66.3; 2022 at 59.3 against 67.5; 2023 at 62.9 against 67.5; 2024 at 65.1 against 67.2; 2025 at 66.9 against 67.0; and 2026 at 67.1 against 67.2. The company states that it monitors credit quality monthly by comparing current forecast collection rates to initial expectations and periodically adjusts the statistical pricing model for trends identified through that review, and that 'since all known, significant credit quality indicators have already been factored into our forecasts and pricing, we are not able to use any specific credit quality indicators to predict or explain variances in actual performance from our initial expectations.' Accurate forecasting of loan performance is listed first among the company's three published critical success factors. This series is a calibration record of the operator's own recovery against its own expectation. It is not a deployment error rate, no source in this record publishes a per-decision defect rate, an override rate, or a correction rate for this deployment, and no independent evaluation of it exists.

  empirical
  • Government Credit Acceptance Corporation, Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (filed 4 August 2026), SEC EDGAR CIK 0000885550 https://www.sec.gov/Archives/edgar/data/885550/000088555026000182/cacc-20260630.htm

• The matter was resolved by a settlement in principle without any adjudication. On 4 January 2023 the Consumer Financial Protection Bureau and the People of the State of New York jointly filed a 59-page complaint in the Southern District of New York pleading seven causes of action, following state subpoenas running from 2016, a federal civil investigative demand of April 2019, a multistate investigation expanded in August 2020 to 41 further states plus the District of Columbia, notice-of-intent letters in November 2020 and August 2022, and a federal notice-and-opportunity-to-respond letter in December 2021. The case was stayed on 7 August 2023 pending the Supreme Court's decision in Consumer Financial Protection Bureau v. Community Financial Services Association of America and the stay was lifted on 1 July 2024, with a revised motion to dismiss filed 14 August 2024 and fully briefed by 29 October 2024. On 24 April 2025 the Bureau filed a consented motion under Federal Rule of Civil Procedure 21 to be dropped as a plaintiff and to withdraw its counsel's appearances; the defendant consented, New York did not object, and the court granted it on 29 April 2025, leaving New York as sole plaintiff. The Bureau's supporting memorandum offers no substantive or merits-based reason and makes no representation about the merits of the withdrawn claims; it is a litigation decision recorded on a docket and not an adjudication or a determination that the claims lacked merit. Contemporaneous trade reporting placed the withdrawal within a broader 2025 pattern of agency enforcement dismissals and carried the operator's statement, through its Chief Legal Officer, that the case 'never should have been brought'; that context is press attribution rather than an agency statement. The case was reassigned to Judge Jesse M. Furman on 28-29 January 2026, and on 6 February 2026 the court deferred ruling on the fully briefed motion to dismiss and terminated it to facilitate settlement, stating it would restore the motion if settlement failed. On 5 June 2026 the court, advised that all claims had been settled in principle, ordered the action dismissed and discontinued without costs and without prejudice to a right to reopen within sixty days, and noted that it would not retain jurisdiction to enforce a settlement agreement unless the agreement were made part of the public record. On 23 July 2026 the New York Attorney General's office reported that the parties had agreed on all settlement terms with documentation to be executed and obtained a forty-five day extension of the reopen deadline. The operator's quarterly report of 4 August 2026 states that it offered $45.0 million in September 2025 to settle this action jointly with the parallel multistate investigation, that in January 2026 it reached preliminary alignment on material terms including a potential cash payment of $75.5 million covering both matters, and that until the matter is fully and finally resolved it intends to defend itself. As of the docket's last update no executed settlement, consent judgment, or public statement of terms appears. There are no findings of fact and no adjudicated liability.

  empirical
  • Government Docket, People of The State of New York v. Credit Acceptance Corporation, No. 1:23-cv-00038 (S.D.N.Y.), CourtListener/RECAP https://www.courtlistener.com/docket/66698748/consumer-financial-protection-bureau-v-credit-acceptance-corporation/
  • Government Memorandum in Support of Consent Motion by Plaintiff Consumer Financial Protection Bureau to Withdraw as a Plaintiff, ECF 86-1 (S.D.N.Y., filed 24 April 2025) https://storage.courtlistener.com/recap/gov.uscourts.nysd.591873/gov.uscourts.nysd.591873.86.1.pdf
  • Government Order dismissing and discontinuing the action, People of the State of New York v. Credit Acceptance Corporation, ECF 94 (S.D.N.Y., 5 June 2026) (Furman, J.) https://storage.courtlistener.com/recap/gov.uscourts.nysd.591873/gov.uscourts.nysd.591873.94.0.pdf
  • Government Credit Acceptance Corporation, Quarterly Report on Form 10-Q for the quarter ended 30 June 2026 (filed 4 August 2026), SEC EDGAR CIK 0000885550 https://www.sec.gov/Archives/edgar/data/885550/000088555026000182/cacc-20260630.htm
  • Reference Civil Rights Litigation Clearinghouse, University of Michigan Law School. Consumer Financial Protection Bureau (CFPB) v. Credit Acceptance Corporation, 1:23-cv-00038 (S.D.N.Y.) case page https://clearinghouse.net/case/48237/
  • Trade press American Banker (2025, April 25). CFPB drops lawsuit against auto lender Credit Acceptance https://www.americanbanker.com/news/cfpb-drops-lawsuit-against-auto-lender-credit-acceptance

• One version of these theories against this operator did produce a governed remedy. On 1 September 2021 the Massachusetts Attorney General announced a $27.2 million settlement in Suffolk Superior Court, described by that office as the largest of its kind, resolving allegations that the operator made high-interest subprime loans 'it knew or should have known many borrowers would be unable to repay', imposed hidden finance charges violating the state's 21 per cent usury cap, engaged in unlawful collection practices, and failed to tell investors that higher-risk loans were placed into securitisation pools. Relief covered more than 3,000 Massachusetts borrowers and comprised cash, debt forgiveness, credit-bureau deletion of related negative marks, and required changes to loan-handling practices; the required practice changes were not enumerated publicly. The operator did not admit wrongdoing, stated that the suit had been 'vigorously contested', and said it looked forward to continuing to serve customers in the Commonwealth through its financing programs. The settlement resolved allegations and is not a finding that the conduct occurred; it is recorded here as evidence that these theories are not novel and that one jurisdiction's version of them landed a remedy.

  empirical
  • Government Office of the Massachusetts Attorney General (2021, September 1). In Largest Settlement of Its Kind, AG Healey Secures $27 Million for Thousands of Massachusetts Consumers from Subprime Auto Lender https://www.mass.gov/news/in-largest-settlement-of-its-kind-ag-healey-secures-27-million-for-thousands-of-massachusetts-consumers-from-subprime-auto-lender
  • Trade press American Banker (2021). Subprime auto lender reaches $27M settlement with Massachusetts AG https://www.americanbanker.com/news/subprime-auto-lender-reaches-27m-settlement-with-massachusetts-ag

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

• Mark AI-written records — Provenance labeling
• Upgrade model — Improve the model
• Review on schedule — Oversight cadence & retrospectives
• Understand the system — Understand the system
• Gate vendor updates — Vendor quality gate
• Gate record entries — Human-in-the-loop write gating
• Review the riskiest first — Risk-tiered oversight
• Check copied records — Reconcile copied records
• Pause AI on alarms — Deployment circuit-breaker

Documented case histories

• Credit Acceptance Corporation's net-collections score
• 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
• Wells Fargo refinance underwriting & the bridge nobody could build
• 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