An Agent-Based Model of Comparable Explanator Selection for Machine Learning Models

This paper addresses the problem of selecting explainability methods for machinelearning models in a comparable way, since conclusions may depend on the chosen explainer, its settings, and the evaluation protocol. We treat explainability methods as a candidate space from which a plugin agent selects an explanation for a given model, data type, and resource constraints. To justify the selection rules, we conduct a PRISMAScR review covering 2016–2026, focusing on CAM, SHAP, and LIME, and considering perturbationbased methods, counterfactual explanations, TCAV, and gradientbased attributions. Based on the review findings, we develop a unified, reproducible experimental protocol for images, tabular data, text, and time series, evaluating fidelity, robustness, and computational cost. On this basis, we propose an explainerselection agent that filters feasible method classes and refines the choice via rapid calibration under metric and resource constraints, improving traceability and reducing arbitrary decisions.