Abstract page for arXiv paper 2611.06977: BlockLock: Reproducible LLM Inference with Fixed Batch Geometry