Evidence suggests that readers optimize low-level visual information following the predictive coding principles. Based on a transparent neurocognitive model, we postulated that readers remove redundant visual signals to focus on the informative aspects of the percept (i.e., orthographic prediction error; oPE). Here, we test alternative oPEs by assuming all-or-nothing signaling units based on multiple thresholds (i.e., neuronal output modality). Further, we tested if readers signal predictions from one or multiple units. For model evaluation, we compared model fits of the new oPEs with each other and against the original formulation based on behavioral and electrophysiological data (EEG). We found the highest model fit for the oPE with a 50% threshold integrating multiple prediction units for behavior and the late EEG data (cluster: from 300-800 vs. 360-620 original oPE). The early EEG data was still explained best by the initial hypothesis (cluster: 210-250 ms vs. 150-250 ms original oPE). Also, we trained image recognition models (ResNet18) in a lexical decision task with the new oPE, the original oPE, or letter string images as input. The assumption is that representations should be better suited for learning if they are optimal. Again, the new oPE formulation performs best. Thus, this formulation adequately describes behavioral performance and allows optimal model training. Brain activations, however, showed that the new oPE was only better for describing the late but not the early activations. This pattern of effects might indicate an evolvement of more optimal representations from early to late brain processes.