QCxAI: Parameter-Shift Saliency for Variational Quantum Classifiers

We present QCxAI, a hardware-compatible protocol for saliency in variational quantum classifiers (VQCs) that applies the analytic parameter-shift rule to inputs, requiring only two circuit evaluations per feature. On a $2{\times}2$ benchmark with ground-truth causal pixels, our one-command, seed-controlled pipeline achieves perfect accuracy, 62.5\% perfect saliency matches (25/40), and a ${\sim}232$–$274{\times}$ salient/random confidence-drop ratio; we additionally report clipped-denominator ratios and effect sizes with 95\% CIs across seeds. The study exposes initialization variance and demonstrates that small ensembles stabilize attribution while preserving the two-eval cost. We position QCxAI as a reproducible, systems-oriented baseline for quantum explainability and a practical faithfulness stress test, distinct from performance benchmarking, and designed to translate to near-term hardware.