We contrast predictions from discrete-state models of all-or-none information loss with signal-detection models of graded strength for the identification of briefly flashed English words. Previous assessments have focused on whether ROC curves are straight or not, which is a test of a discrete-state model where detection leads to the highest confidence response with certainty. We along with many others argue this certainty assumption is too constraining, and, consequently, the straight-line ROC test is too stringent. Instead, we assess a core property of discrete-state models, conditional independence, where the pattern of responses depends only on which state is entered. The conditional independence property implies that confidence ratings are a mixture of detect and guess state responses, and that stimulus strength factors, the duration of the flashed word in this report, affect only the probability of entering a state and not responses conditional on a state. To assess this mixture property, 50 participants saw words presented briefly on a computer screen at three variable flash durations followed by either a two-alternative confidence ratings task or a yes-no confidence ratings task. Comparable discrete-state and signal-detection models were fit to the data for each participant and task. The discrete-state models outperformed the signal detection models for 90 % of participants in the two-alternative task and for 68 % of participants in the yes-no task. We conclude discrete-state models are viable for predicting performance across stimulus conditions in a perceptual word identification task.