Right now it re-generates the whole mapping for _every_ armature at each evaluation. While this works well, this is overkill and should be optimized.