Unpacking the Evaluators: How Configuration Shapes the Evaluation of Alignment in Explainable AI

Explainable AI (XAI) evaluation fundamentally requires human-centered evaluation, assessing whether explanations align with human reasoning and expert knowledge. However, such evaluations often require expert-driven manual validation and are difficult to scale, leading to the widespread use of alignment-based metrics, where attribution maps are compared against predefined ground-truth (GT) representations. Despite their popularity, alignment evaluations are implicitly bias towards multiple design choices made during the evaluation process, for example GT representations, metric families, metric formulations (soft vs.\ hard), and thresholding strategies. In this paper, we formalize alignment evaluation as a structured evaluation configuration space through which the influence of these design choices can be systematically analyzed. We further introduce a structured alignment evaluation framework that models interactions between XAI methods, GT representations, metric families, metric formulations, and thresholding strategies across 2D and 3D synthetic and real-world datasets. In addition, we propose a novel taxonomy of XAI output behavior based on local entropy and boundary mass ratio, enabling structured characterization of attribution patterns for alignment evaluation. Our experiments reveal systematic flows(triples of XAI-GT-metric configurations), showing that alignment outcomes provide distinct behavior patterns under certain configurations, such as soft/hard metric implementations. We further observe varying levels of threshold sensitivity across metric families and attribution structures. These findings demonstrate that alignment evaluation cannot be treated as a fixed or invariant measure of explanation quality, and provide practical guidance for designing more reliable and behavior-aware XAI evaluation protocols.