AUDIT-READY EXPLAINABLE AI FOR FRAUD OPERATIONS: PERSISTED AND REPLAYABLE DECISION ARTEFACTS FOR MODEL GOVERNANCE AND INVESTIGATOR TRUST

Abstract

Fraud detection operations increasingly depend on machine-learning systems to prioritise suspicious events for investigator review. In regulated environments, however, operational defensibility requires more than predictive accuracy and post-hoc explanation. A review, challenge, or audit must be able to reconstruct which model, feature contract, threshold or alert-budget policy, explainer configuration, and workflow actions produced a given alert at decision time. This paper argues that explainability is not the same as audit readiness. It proposes an audit-ready design pattern for fraud operations in which each alert is treated as a governed decision artefact with persisted score and explanation snapshots, version and threshold lineage, investigator disposition records, and monitoring evidence for drift and replayability. The manuscript contributes four core outputs: a six-dimension audit-readiness rubric, a minimum alert-artefact schema, an architecture pattern for persisted and replayable explanations, and an evaluation blueprint that separates predictive quality, explanation quality, workflow utility, and audit readiness. The paper also analyses privacy, security, and retention risks introduced by persisted artefacts and proposes practical controls for role-based disclosure, minimisation, immutable access logging, and evidence-preserving storage. The result is a publication-ready framework for converting explainable fraud models into traceable operational systems.