The goal is not to model a single skill - i.e. a single deliverable, but to model a skill in an evolutionary sequence of skills. What we mean is that we can show a next-step function (as in an inductive sequence), where the function shows how the next skill in the sequence has higher fitness and is derived from the previous step by viable mutations.
We therefore have to model the likely mechanisms of the next step function.
A model of a skill has to satisfy output constraints. The model has to deliver information to control skeletal action through muscle-based control of joint rotations.
The model has to fit in with reasonable information input from perceptual mechanisms. We start with gravity and vision. For reasons of manageability, we cannot expect to model the perceptual process right from the physical input device - but the input data assumptions have to be fairly reasonable (even if simplified).
There has to be a timing component for the model that is reasonable.
There has to be an abstract storage model for the skill as information specification that is reasonable for the transformation of the genetic encoding of the skill to brain representation to action.
The abstract model for the genetic encoding of the skill has to be compatible with meiosis - the recombination of skills from both parents.
The abstract model of the skill has to allow for reasonable internal processing that is part of the skill, such as action that is conditional on perceptual input, or action that is dependent on prediction and planning.