Visual confidence refers to our ability to estimate the correctness of our visual perceptual decisions. As compared to other forms of metacognition, meta-perception has attracted a burst of studies recently, no doubt because perception already benefits from strong theoretical frameworks. We have recently refined these existing frameworks by proposing to clearly distinguish sensory evidence from some “confidence evidence” that drives the confidence decision. The problem now is to characterize the properties and consequences of this confidence evidence, and this is the aim of the present proposal. As the number of studies grows, it becomes clear that visual confidence is not simply a noisy estimate of the perceptual decision, but instead depends on a large number of factors. We have identified four axes that we believe will contribute to shape confidence evidence: (1) individual variability, (2) task accessibility, (3) global confidence, and (4) perceptual learning. The purpose of the first axis is to understand which cues are used for confidence, and for this purpose, we will study confidence variability across individuals. Some of the idiosyncratic variability in confidence judgment efficiency might come from a variable temptation to exaggerate the impact of stimulus noise on the estimation of one own performance. In the second axis, we will try to understand what in a task determines the accessibility to visual confidence. In particular, we will test the hypothesis that more high-level tasks, such as face identification, lead to better confidence efficiency that low-level tasks, such as detecting whether two line segments are aligned. The aim of the third axis is to understand how individuals construct a sense of confidence for a task as a whole, not for a single isolated judgment. We will start by carefully studying how confidence builds up within a set of stimuli and compare how such a global confidence compares with a single decision confidence. Finally, in the fourth axis, we will study how perceptual learning benefits from visual confidence. In particular, we will test the extent to which confidence evidence can be seen as an internal error signal that can act as a proxy for an external feedback. We believe that a better understanding of these four fundamental aspects of confidence evidence will help us derive an accurate and useful model of visual confidence, and ultimately of metacognition.