We tested the dual process and unequal variance signal detection models by jointly modeling recognition and source confidence ratings. The 2 approaches make unique predictions for the slope of the recognition memory zROC function for items with correct versus incorrect source decisions. The standard bivariate Gaussian version of the unequal variance model predicts little or no slope difference between the source-correct and source-incorrect functions. We also developed a "bounded" version of this model that did not permit below-chance source discrimination in any region of the evidence space. The bounded version predicts that the source-correct function should have a lower slope than the source-incorrect function. A bivariate version of the dual process signal detection model can predict slope differences in either direction, but it must predict a u-shaped slope for items attributed to the correct source than items attributed to the incorrect source, and source zROC function if the source-correct slope is lower than the source-incorrect slope. Across 4 experiments, results consistently showed that the recognition memory zROC function had a lower the source zROC function for words recognized with high confidence was linear. Only the bounded version of the unequal variance model successfully predicted the full pattern of results.