Auditable Credit Risk Intelligence for the U.S. Financial System: A Scalable Explainable-AI Framework Reconciling Predictive Performance with ECOA, FCRA, and Model Risk Governance
Machine-learning credit underwriting in the United States operates under three families of obligation that are not reducible to one another: a statutory prohibition on discrimination, a statutory duty to disclose the specific principal reasons for an adverse action, and prudential expectations for model risk management. Prevailing practice reconciles these demands by sacrificing model capacity, deploying low-capacity scorecards whose interpretability is structural rather than earned. This study argues that the United States regulatory realignment of 2025 and 2026 makes that settlement less defensible, not more. The Consumer Financial Protection Bureau withdrew the interpretive circulars governing algorithmic adverse-action practice and amended Regulation B to disclaim disparate-impact liability under the Equal Credit Opportunity Act, while leaving the statutory disclosure duty untouched and preserving liability for the intentional use of facially neutral proxies. The federal banking agencies simultaneously replaced the prescriptive 2011 interagency model risk guidance with a principles-based instrument that expressly disclaims enforceable standards. Fewer obligations are now externally specified, and more must be self-specified, self-justified, and self-evidenced. Because effects-based exposure persists under the Fair Housing Act and state analogues, and private litigation is unaffected by federal guidance withdrawal, the value of verifiable self-generated evidence rises as external specification recedes. The paper develops ACRIS, the Auditable Credit Risk Intelligence Stack, a five-layer architecture in which admissibility, explainability, disparity control, and governance evidence are enforced during training and serving rather than audited afterwards. Its layers comprise a provenance-gated feature-admissibility screen bounding residual proxy information through conditional mutual information; a shape-constrained predictive core; a constrained-optimization layer recording an entire searched alternative-model frontier, including rejected candidates and rejection rationales; an explanation engine deriving principal reasons from a counterfactual approval baseline and gating disclosure on a per-decision stability margin; and an append-only, tamper-evident governance ledger. A finite-sample sufficient condition for top-k reason-set preservation under parameter resampling is proved, with its assumptions and failure modes stated explicitly. The framework is not empirically validated. A pre-registered protocol of falsifiable propositions, corpora, temporal validation design, baselines, metrics, statistical plan, and pre-committed disconfirmation criteria is specified so that every central claim can be tested and, if wrong, refuted.