Statistical rigor for AI agent evaluation — confidence intervals, significance testing, effect sizes