Low-frequency (LF) words have higher hit rates (HRs) and lower false alarm rates (FARs) than high-frequency (HF) words in recognition memory, a phenomenon termed the mirror effect by Glanzer and Adams (1985). The primary mechanism for producing the mirror effect varies substantially across models of recognition memory, with some models localizing the effects during encoding and others during retrieval. The current experiments contrast two retrieval-stage models, the Source of Activation Confusion (SAC; Reder, Nhouyvanisvong, Schunn, Ayers, Angstadt, & Hiraki, 2000) model and the unequal variance signal detection theory (UVSDT) criterion shift model (e.g., DeCarlo, 2002). The SAC model proposes that two distinct processes underlie the HR and FAR effects, with a familiarity process driving the FAR effect and a recollective process driving the HR effect. The UVSDT criterion shift model assumes that subjects use different criteria when making recognition judgments for HF and LF words, with this single process driving both the HR and FAR effects. Experiment 1 incorporated divided attention and speeded responding manipulations designed to remove the contribution of recollection in the SAC model, thereby eliminating the LF HR advantage. Experiment 2 manipulated the salience of the frequency classes, as the UVSDT criterion shift model requires that subjects are aware of the distinct frequency classes in order to shift their criteria. Across both experiments, model simulations and direct fits of the SAC model demonstrated systematic errors in prediction. While the UVSDT model struggled in fits to Experiment 1 data, the model provided acceptable fits to Experiment 2 data and accurately predicted the general pattern of effects in all cases. Furthermore, state-trace analyses provided compelling evidence in favor of single-process rather than dual-process models of recognition memory, casting serious doubt on the validity of the dual-process SAC model. Finally, the current experiments highlight the importance of obtaining model-based estimates of sensitivity and bias across frequency classes, as the standard practice of conducting direct comparisons of HRs and FARs for HF and LF words confounds bias and sensitivity differences.