Context-dependent signals shape the representation of sensory stimuli in the brain and determine which stimuli are amplified to guide behaviour. Top-down attentional signals were proposed to modulate the apical gain of pyramidal neurons in the sensory cortex, but the exact nature of this interaction remains poorly understood. Long-range orbitofrontal cortex (OFC) projections were found to respond to changes in task contingency, targeting somatostatin-expressing (SST+) neurons in the sensory cortex. We suggest that the apical gain modulation reflects the stimulus value, i.e. the future reward estimation, which is learned via reward prediction error (PE, Error 1). While improving on the task, the variability in the PE decreases, and the model confidence increases. The variability in the PE is itself postulated to be learned by a higher cortical area. When a change in the task contingency occurs, the PE variability suddenly increases, causing a task-prediction error (2nd-order PE, Error 2). The occurrence of this task-prediction error results in a model doubt that we postulate to be represented in the OFC. Upon task reversal, OFC reverses the top-down attention by apical inhibition via local SST+ interneurons. The computational model is supported by Ca2+-imaging data from both the lateral OFC and primary somatosensory cortex (S1) acquired in mice during reversal learning of a texture discrimination task. The theory shows how a hierarchy of PEs forms the attention-modulated sensory representation in S1 (Error 1) and how it reshapes this representation through model doubt formed in OFC (Error 2) with rule reversal.