POLICY-CARRYING DECISION MODELS FOR FINTECH A PRACTICAL FRAMEWORK FOR VERIFIABLE, AUDIT-READY AI DECISIONS IN FINANCIAL SERVICES

Abstract

Financial decisions in production systems must satisfy a layered set of obligations: risk tolerance, regulatory compliance, fairness constraints, privacy requirements, and operational service levels.Most machine learning models optimize predictive objectives but treat policy and compliance as external checks.This separation creates avoidable failure modes: decisions that are accurate yet non-compliant, long audit cycles, and limited customer recourse.This paper proposes Policy-Carrying Decision Models (PCDMs): decision systems that emit not only an outcome (approve/decline/route) and calibrated confidence, but also a machine-checkable proof that the decision adhered to an explicit policy expressed in a domain-specific language (FinPol).At inference time, the model (and its surrounding decision logic) produces a decision receipt containing the outcome, explanations scoped to permissible disclosure, and a verifiable policy proof.Optionally, a zero-knowledge variant allows third parties to verify compliance without access to sensitive features or thresholds.