“CODA kernels, whether authored by humans or Large Language Models, achieve high performance on representative Transformer workloads.”