Failure-Aware Evaluation for Governed Automation in Convergent Engineering: A Cross-Domain Methods Perspective

Abstract

Convergent engineering brings AI into materials, water, energy, manufacturing, and cyber-physical operations, but common performance averages obscure how failures propagate across those domains. This review develops a cross-domain evaluation method for governed automation. It separates sensing, representation, prediction, decision, actuation, and recovery, then tests each interface for missing data, distribution shift, rare events, adversarial or accidental corruption, and human misunderstanding. Evaluation combines calibration, abstention quality, time-to-detection, decision cost, rollback performance, and the ability to reconstruct a recommendation from source evidence. Policy constraints, model and dataset versioning, access control, independent monitoring, and human appeal are treated as measurable system functions rather than documentation. The synthesis highlights feedback loops in which automated actions reshape later training data and can hide their own errors. No new cross-domain benchmark is reported. Its contribution is a failure-aware AI assurance framework that compares automation by its capacity to remain bounded, explainable, recoverable, and accountable when technical components or institutional assumptions fail.