it doesn't have to be the case, but it often seems to be the case. I think there is a model by which deadliness is usually due to an accident -- note that most

it doesn't have to be the case, but it often seems to be the case. I think there is a model by which deadliness is usually due to an accident -- note that most contemporary deadly pandemic diseases (even flu!) jump from animals to humans, so the deadliness is "accidental" in some sense, it's less deadly in the animal reservoir. But maybe the better way to think of it is entropically/random walk POV. Starting from a base sequence S0, if the virus starts out at a local maximum of deadliness (or at least an increase in deadliness that only has certain mechanistic "channels of escape"), and then optimizes strictly for contagiousness, statistically speaking most mutations in an S1 will make it less deadly (even if marginally so), unless contagiousness and deadliness are mechanistically linked. I think our understanding now is that the first order bottleneck in deadliness is something like "binding to a receptor", so most increases in contagiousness will have a strictly neutral-to-deleterious effect on deadliness.