That's not correct - we simply have no benchmark data to prove whether
it is good or bad. It has only been thoroughly tested on single chains.
It *appears* to function rather well on multichain models, however.
Then, how is
the best way to evaluate them? Each chain should be evaluated
independently? Should I take only the best molpdf score?
You certainly should not use the molpdf score - the Modeller scoring
function does a great job for optimization, but it is not designed to
discriminate good models from bad models at assessment time.