Custom JumpReLU Sparse Autoencoders for mechanistic interpretability. T5Gemma-2 SAEs across all layers. AI safety & interpretability research.