As it has already been discussed before, the problem ratings in Chess Mentor don't really reflect the difficulty of the exercises.

While it's no big deal, maybe you could add an option for users to suggest a new (more realistic) rating to a problem they have already tried. Then from time to time, someone in chess.com staff could review the problems which have received enough rating suggestions (say 3 ?) and adjust the ratings if necessary.

It could be the same mechanism as is used for leaving comments on the Tactics Trainer.

I think it makes sense to implement a little system in which each puzzle measures its' worth by what skill level the people getting it wrong are at and what frequency.  It wouldn't have to be perfect to be effective.

Just for a set of data points I have ratings of 500 in  the TT, 800 online, and 1700 on the CM.

I have a large disparity, too.  Online rating is 1650-1670.  My CM rating is 1700, and my TT rating is 1250-1300.