This hasn't mattered for any cases we've looked at so far because we haven't ever tried to predict attributes that had components in them (e.g., predicting success in RumbleBlocks has no components in it).
In the future when we want to try to predict things with components (e.g., SAI relations in tutors) it might make sense to perform a second structure mapping on the returned concept. This might get closer to the kind of result that structure mapping internally would produce (i.e., mapping depending on concept structure) without the cost of all the internal structure mapping.
This hasn't mattered for any cases we've looked at so far because we haven't ever tried to predict attributes that had components in them (e.g., predicting success in RumbleBlocks has no components in it).
In the future when we want to try to predict things with components (e.g., SAI relations in tutors) it might make sense to perform a second structure mapping on the returned concept. This might get closer to the kind of result that structure mapping internally would produce (i.e., mapping depending on concept structure) without the cost of all the internal structure mapping.