Happy to answer questions. Quick note on methodology: the TFLOPS figure is effective (computed as if doing the full dense multiply) — ROLV doesn't violate hardware peak, it avoids work entirely via structured sparsity. Weights are pulled directly from HuggingFace, output verified with norm hashes and a canonical check. If you want to run a baseline on your own hardware, there's a validation kit at rolv.ai.