Generally, the past “tuning” is done to low absolute values. Instrumental SST and what the sky “sees” are not always the same. So if you are basing what the model limits are when the initial values or “tuning” criterion are you can come up with all sorts of reasons why “models” won’t work.
If the models are properly initialized, that would be the information need to evaluate how well they may be able to perform. Once that GFLD example gets to 300K, it starts leveling off as it should.
Personally, I wouldn’t write them off completely because I am a cheap bastard, they can produce something useful, but I think cleaning house in a few “institutions” might stimulate a little more attention to detail.