It's intuitively plausible that a Solomonoff type prior would (at least approximately) yield such an assumption.

But even if "intuitively plausible" equates to, say, 0.9999 probability, that's insufficient to disarm Pascal's Mugging. I think there's at least 0.0001 chance that a better approximate prior distribution for "value of an action" is one with a "heavy tail", e.g., one with infinite variance.