I don't know what Poser gives you access to. For instance, as per usual, there are lots of command possibilities for the wavefront object file format. And, just about everyone that implements it has decided to supersede or subsume many of those functions. Probably for the better, to be honest. :) http://paulbourke.net/dataformats/obj/

Poser's Smoothing Tool relies on user-input and placement of selection, first. That random component interacts, as far as I can tell, by doing a simple weighted average of vertice positions weighted for their neighbors and their distance from the object center. ("Object" also being treated as the targeted Group for "object center" purposes.) Sorry, I don't know if there is a specifically named algorithm for what Poser uses, but I really don't know that matters right now. It's probably a variant/simplification of Laplachian smoothing. (No idea, honestly, 'cause it's not easily predictable and can not be user-selected, so I never bothered trying to figure out what it was.) (http://www.faculty.jacobs-university.de/llinsen/teaching/320491/Lecture13.pdf ) Laplachian is also mentioned, below.

I came across this and it may give you some ideas on how to isolate "details" in topologicial features: https://link.springer.com/content/pdf/10.1007/11428831_2.pdf

So, the basic takeaway here is that they have figured out that determining the "angle" being represented is key to preserving detail. And, it makes a heck of a lot of sense for your particular operation. What are topological "details?" We know intuitively what they are, but from the topo point of view, they're larger deviations in relative angle from the surrounding topology. That's a sort of no-brainer. (I think the paper is more about sub-d than "smoothing" and that's why I can's stand the two terms being used interchangeably. :) )

So, forget about pushing verts or, rather, relative delta values, which not only rely completely on the original vert reference, but also the winding order for the object/group. :) Focus on "angles" and stuffs, first, i think. A delta is fairly meaningless without knowing what vert it references. And, to take the entire population of them and just figure out what a "detail" is by examining them individually isn't possible. One has to compare relative verts, from the original reference and the surrounding verts. And, the best way to produce something useful may just be analyzing relative angles, instead, against the surrounding topology. Bigger angles inherently mean larger deviations from the surrounding mesh, which inherently means "surface detail."

<-- Is not a maths dude, sorry. :)

PS - I also don't know what Poser's CR2 format brings to the table in terms of features nor what is a geometrical eccentricity of Poser you could take advantage of. In terms of just pushing-verts-tech, I think the "angle approach" is.. the best angle to approach this? :)