I'd say at least one of the assumptions in using this particular modeling strategy is that all the neccessary and sufficient parts of the eventual model are observable (without additional instrumentation) in the behavior (verbal and non-verbal) of the exemplar (e.g. Carl) in the relevant context(s).  The question then becomes, "How do you know if what you're modeling lends itself to this strategy?". I'd say at least one way is to just go for it. Pick one context, know what evidence you need to know that you've been successful, and pretend to be Carl.  If you don't get the evidence, it could mean any number of things, including: 1. You've missed some difference(s) [that you're neurologically capable of detecting] between Carl's physiology (I'm guessing we're talking micro-muscule movements here - including those of the vocal tract) and your own in that context. 2. There's some quality about Carl (e.g. bone structure, eye color, hairstyle, size of the resonating cavitys in his vocal tract, muscle strength, etc.) that is necessary to the task at hand, and yet is potentially non-trivial for you to adapt to. 3. There's some factor outside of what you've defined as the context (pre-ordered deck of cards, other people who happen to be present, family ties, time of day, etc., etc.) 4. There's some difference you can replicate physiologically, but can't necessarily detect without additional instrumentation due to the limitations on what we're calling first attention. It could be any number of things.  I would suggest that the domain of NLP[modeling] is precisely those skills in which the problem(s) in reproducing excellence are mostly due to reason 1 (above), and possibly 3 or 4 or parts of 2 depending on how you think about it.  Keep in mind that these categories of problems I'm proposing are very informal. The point however, is that assuming what you're modeling has mostly to do with 1, you start out by not caring how it works, just if it works when you pretend well.  Assuming that it's reasonable to believe the problem is with 1, and you're not getting results, you find out how to pretend better. The point in witholding judgement as to how it works at first is a big piece in making sure you take in as much information as possible when pretending.  The very act of forming theories and/or naming things limits people's attention. So, in the beginning you mimick everything (including, in milton's case, bringing a pillow along with you, and pretending your arm is limp) until you can get results.  Then you know you're onto something and you can begin to weed out what seems to be unnecessary, thereby arriving at a (hopefully compressed) description of "how". I hope that helps... -Emery