There's a gap between researcher-crafted evaluation frameworks and how end users actually use AI products, and this gap is exploited in ways that render traditional reward models useless.