Explainable artificial intelligence is increasingly needed in high-stakes tabular classification, where predictions should be accurate, auditable, and easy to inspect. We propose GRS-ANFIS, a role-separated neuro-fuzzy model that decomposes inference into a Primary module for main decision formation and a Complementary module for targeted correction. During differentiable training, sigmoid gate values are applied only to consequent coefficients, while the antecedent part receives the original input without soft masking. After each stage, the learned gates are binarized into hard routing masks that define discrete antecedent and consequent subsets for module-specific fine-tuning. The Complementary module is restricted to variables not selected by the Primary module, yielding explicit role separation and disjoint variable usage across modules. To support stable learning in high-dimensional settings, all ANFIS-family models use the same HTSK-style firing computation. Experiments on four tabular benchmarks show that GRS-ANFIS achieves competitive predictive performance while maintaining compact, role-separated rule structures; rule-count compactness is clear, whereas the unified Nauck/HFSi interpretability values are dataset- and variant-dependent. Boundary-focused analysis further shows that the Complementary module mainly improves difficult, low-confidence samples through targeted correction.