Testing GLOM's ability to infer wholes from ambiguous parts