Working on evolutionary robotics in the 80's and 90's we used to refer to these things as 'scaffolding problems'. For non-trivial systems getting the constrain

Working on evolutionary robotics in the 80's and 90's we used to refer to these things as 'scaffolding problems'. For non-trivial systems getting the constrain specifications right, raising the right scaffolding, can become as hard as handcrafting the solution would have been. It is a pain point in every system that tries to solve a problem through generative meta-heuristic search. Translating intention into a set of explicit constraints that separate desired outcomes from undesired ones that satisfy the letter of the constraints but not the intent gets exponentially harder as the dimensionality of the solution space grows.