The shadow mean for a fat-tailed domain is the true population mean that would be observed if you had infinite data; it differs from the sample mean because finite samples cannot capture tail events that have disproportionate impact on the expectation, creating a gap between observed mean and true mean.