GRAIL: Gini/Graph-Routed Adaptive, Interpretable Learning for Accurate and Efficient Tabular & Mixed-Modality Inference

I present , a neural architecture that couples an oblique decision router regularized by Gini impurity, a graph-aware feature router, and calibrated early exits for efficient inference. That yields rule-like decision paths and cluster-level rationales while remaining competitive on accuracy with millisecond-level CPU latency. We provide a practical training recipe---stability-safe Gini annealing, route-consistency regularization, knowledge distillation from strong tabular teachers, and validation-calibrated early exits---that delivers strong accuracy without sacrificing transparency. On standard tabular benchmarks we observe competitive accuracy with sub-millisecond p50 latency, near-one-hot routing purity, and human-readable explanations.