From Model Proliferation to Deployment Intelligence: A Risk-Traceable System-Aware Model Selection Framework

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The rapid growth of machine-learning architectures has shifted an important engineering challenge from developing additional models to selecting the most suitable trained candidate for a specific operational environment. Benchmark metrics provide essential evidence, but they do not determine how competing model behaviours should be interpreted when deployment risks and priorities differ. This paper proposes the System-Aware Model Selection (SMS) Framework, implemented through the Deployment-Aware Model Selection Algorithm (DAMSA), as a three-layer post-training decision architecture. The pre-ranking layer translates operational hazards into measurable criteria and distinguishes core, composite, descriptive, and mandatory roles. The candidate-control and reference-ranking layer applies feasibility and Pareto screening with a transparent SMS Index, while the post-ranking layer evaluates F1-score influence, preference robustness, and bounded metric perturbations. Across five scenarios, YOLOv7 was selected for traffic monitoring and pharmaceutical inspection, with first-rank acceptabilities of 92.8% and 91.2%, respectively. In defence surveillance, YOLOv8s was preferred to the highest-mAP YOLOv11s because its Precision was 18.9 percentage points higher. YOLOv8s was preference-stable in microplastics detection. In CCPP monitoring, Random Forest became preferable when the Generalization Stability priority exceeded 12.99%. Removing F1-score did not change any vision recommendation.