Provenance, Governance, and Human Review for Multimodal Zero-Shot Anomaly Detection: For Reproducible Human-Ai Collaboration in Operational Assurance

Abstract

This critical review examines multimodal zero-shot anomaly detection and governed industrial diagnosis. It asks how provenance and review authority should be attached to model-assisted decisions. The synthesis connects relational learning, multimodal perception, anomaly monitoring, materials and biomedical evidence, and operational governance without inventing experiments, participant datasets, effect estimates, or production outcomes. Ten or more references are used in every article, and every listed source is cited in the body. Particular attention is given to visually plausible explanations masking localization or calibration failure. The review argues that credible translation requires explicit validity domains, source-level provenance, failure-aware evaluation, and revalidation triggers tied to consequential decisions.