MOSAIC outsources 70B-scale transformer inference with masked matmuls and LWE/LPN security, and it still hits BF16-ish accuracy.