I also find the 3rd criteria counterintuitive. If worlds share features, I would expect these to not be independent.
Agreed, I think this criterion is very strongly violated in practice; a Gaussian process prior with a nontrivial covariance function would be a bit more realistic.
Agreed, I think this criterion is very strongly violated in practice; a Gaussian process prior with a nontrivial covariance function would be a bit more realistic.