I've been giving this some thought also. The best idea I came up with was to create an orb-puzzle solver and then use a genetic algorithm to find ridiculously difficult puzzles. But that's probably not feasible for this particular contest (if for no other reason, than you wouldn't have time to run the GA).
Perhaps a more doable option is to choose behavior for most of the orbs, and then choose a potential solution, then set the behavior for the remaining orbs to make that solution work. For example, you decide that the solution should be ABDBCAE. Pick behaviors for A, B, and C. Then choose the behaviors for D and E so that all the doors are open.
Another option (since everything is going to be under tar anyway), is to choose behaviors for all orbs and a solution, then have some doors start open instead of closed to make the solution work.
Of course, neither of those approaches guarantees that there won't be a simpler solution, so you'll probably need an orb solver program to make sure that things are difficult.
____________________________
Progress Quest Progress