A long-standing debate in the recognition-memory literature concerns which model provides the best account. Prominent candidates in this debate are the unequal-variance signal detection model (UVSD), the dual-process model (DPSD), and two versions of the mixture model (MSD). The present work evaluates a recently proposed ROC-based method for comparing these models (Dede, Squire, & Wixted, Neuropsychologia, 54, 51-56, 2014). This method consists of evaluating the pattern of residuals produced by each model's best fits to ROC data. Previous results showed that the DPSD produced systematic residuals while the UVSD did not, a difference that was interpreted as evidence for the superiority of the latter model. Using a linear mixed model (LMM), we evaluated each model's residuals for 883 individual ROCs. LMM results revealed the presence of systematic residuals in all candidate models, indicating a general failure of these models to capture some of the regularities found in the data. We discuss different ways that current signal detection models can be modified or extended in order to meet the challenge that these systematic residuals represent.