JAXBench: curated TPU docs beat model scale.

Harvard and Google released the first TPU-native benchmark for AI-generated kernel optimization: 50 JAX workloads, 17 production operators from MaxText architectures (Llama-3.1, DeepSeek-V3, Mixtral, Mamba-2, AlphaFold2) and 33 translated from KernelBench at sizes tuned for high TPU v6e MXU utilization (arXiv 2607.20466). Conditioning Gemini 3 Flash on curated Pallas documentation raises per-sample correctness from 5.8% to 37.3% and solves 48 of 50 benchmarks at 1.28x geomean speedup. Autocomp's beam search reaches 1.36x over XLA. Target-specific context beats a bigger model, which is the whole argument for skills files in one number.