The paper discusses the challenges of building robust classification systems in environments where target operating conditions, such as misclassification costs, are uncertain. It introduces the ROC Convex Hull (ROCCH) method, which combines techniques from ROC analysis, decision analysis, and computational geometry to compare and visualize classifier performance in imprecise environments. The ROCCH method decouples classifier performance from specific class and cost distributions, allowing for the identification of potentially optimal classifiers under various conditions. The paper demonstrates that the ROCCH can be used to build hybrid classifiers that perform at least as well as the best available classifier for any target conditions, and in some cases, can even surpass it. The hybrid classifier is efficient to build, store, and update, and it can handle a wide range of comparison frameworks, including optimization of metrics such as accuracy, expected cost, lift, precision, recall, and workforce utilization. The paper also provides empirical evidence that robust hybrid classifiers are necessary for many real-world problems.The paper discusses the challenges of building robust classification systems in environments where target operating conditions, such as misclassification costs, are uncertain. It introduces the ROC Convex Hull (ROCCH) method, which combines techniques from ROC analysis, decision analysis, and computational geometry to compare and visualize classifier performance in imprecise environments. The ROCCH method decouples classifier performance from specific class and cost distributions, allowing for the identification of potentially optimal classifiers under various conditions. The paper demonstrates that the ROCCH can be used to build hybrid classifiers that perform at least as well as the best available classifier for any target conditions, and in some cases, can even surpass it. The hybrid classifier is efficient to build, store, and update, and it can handle a wide range of comparison frameworks, including optimization of metrics such as accuracy, expected cost, lift, precision, recall, and workforce utilization. The paper also provides empirical evidence that robust hybrid classifiers are necessary for many real-world problems.