[mjburgess]

Yes, this is essentially what i mean by "distributions are sampled from distributions", ie., there's a subjective uncertainty in the choice of model but also an uncertainty in the very determination of that uncertainty.

You can model a plausible "final stable distribution", after all these recursions, with a power law.

This makes intuitive sense, if you consider how science works: all the confirmatory evidence in the world doesn't help, all the information lies in the single refutative point. This is how power laws work.

So basically we're always operating under a heavily under-determined region with high uncertainty, and we can only improve that by disconfirmatory apparent outliers.