EVIDENCE GATED ROOT CAUSE DIAGNOSIS FOR FINTECH MICROSERVICES UNDER TELEMETRY LOSS

Root cause explanations in financial microservices become operationally useful only when their supporting evidence can be retrieved and checked. This study develops an evidence gated diagnosis framework that separates statistical confidence, evidence availability, and cryptographic integrity. A reproducible computational experiment generates 600 independent synthetic incidents across three abstract failure families and evaluates four nested telemetry missingness conditions, producing 2,400 paired observations. A correlated Gaussian graph model ranks candidate causes; a selective decision rule accepts a diagnosis only when confidence and evidence requirements are satisfied. Across the matched model conditions, unrestricted graph diagnosis achieves 85.6% accuracy, compared with 42.0% for a largest anomaly heuristic. Evidence gating accepts 65.5% of observations without an observed diagnostic error, but rejects 63 correct recommendations beyond confidence filtering alone. Under deliberate model mismatch, accepted accuracy falls to 92.7%, exposing the limits of confidence thresholds. Digital signature tests reject all 1,000 specified integrity violations while accepting deliberately false claims signed by the authorised test key. The contribution is an executable evaluation framework for defensible abstention, rather than evidence of faster human response or regulatory compliance.