Position: Explainability Research Must Prioritize Foundations over Ad-hoc Methods

Collection of recent ML research (July 2026) spanning explainability, vision-language zero-shot methods, geospatial RF planning, domain-contamination control for retrieval, quantum-hybrid wildfire segmentation, sandbox-native RL for language agents, financial sentiment from 10-Ks, and low-latency relay selection for NR-V2X. Key security and safety-relevant themes: (1) Explainability work argues for foundational, human-in-the-loop XAI to ensure explanations drive safe action rather than being ignored; (2) Retrieval/domain-contamination (C3R) and model-exploitation (RENEW/DLHF) papers highlight,