From Accuracy to Auditability: A Survey of Determinism in Financial AI Systems

In three linesSurvey on reproducibility of AI systems in regulated finance (credit, fraud, AML). Identifies three sources of nondeterminism: post-hoc explanation variance (tabular models), stochastic sampling (graph networks), batch-dependent divergence (LLM agents). Proposes evaluation framework with RBO, D_cos, TDI, PSD metrics for audit readiness.

Summary generated by Claude — human-verified