ROTOR: Rotational Orientation Token Organization & Recall — An Attention-Free Multi-Head Non-Abelian Spinor Architecture
ROTOR (Rotational Orientation Token Organization & Recall) is an attention-free, non-abelian geometric sequence architecture designed and invented by Prannessh K.V.A. ROTOR eliminates the quadratic O(N × M) compute and memory bottleneck of Transformer Cross-Attention by accumulating incoming context streams into independent 3D coordinate frames on the 3-sphere S³ ≅ SU(2) using 4D pure-real unit quaternions. Query streams interrogate accumulated frames via a non-abelian conjugation sandwich with strictly sign-invariant readouts, compressing sequences into a constant 4.5 Kilobyte memory footprint. ROTOR natively supports both ROTOR-15M (Dense Edge) and ROTOR-50M-MoE (Sparse Top-2 Mixture-of-Experts) configurations. Official Hugging Face Repositories: https://huggingface.co/Prannesshkva/ROTOR-15M https://huggingface.co/Prannesshkva/ROTOR-50M-MoE