The explanation is quite simpel, the model will correct the Ra/Dec coordinates, so only in the case there is absolute no backlash, orthogonal, polar errors etc ... the model will NOT work, that's why I recommended "no model".
As soon as these errors exists the model will try to correct for those and they will skew the slew, it sounds counter intuitive and it even makes you wonder why a model works .... wel the issue is that if you gather to much sync points in a very close manner, the equation will have unforeseen convergence errors ... a model works best with lots of points all over the sky but not to close to each other ... EQMOD mention this, and I guess all other softwares that build models, never really tested with model points a few pixels away from each other ... hence their formula goes crazy ...
/Yves