MS-CPAT-LOB v4 combines multi-scale CNNs, state-space models, and efficient transformers to forecast limit order book dynamics with regime awareness. Raw LOB tensors flow through feature engineering, multi-scale convolutions (1×1, 3×3, 5×5), SSM selective scan compression, and sparse attention fusion, with a regime classifier routing through mixture-of-experts layers. This hybrid architecture reduces computational complexity to O(n log n) while capturing both local patterns and long-range temporal dependencies in high-frequency market data. Fork and customize this diagram on Diagrams.so to adapt the fusion strategy, attention mechanism, or regime routing for your own financial ML pipeline.