An O(n log n) language model architecture using wave equation dynamics instead of O(n²) self-attention. Within 5% of standard transformer quality.