We posit that robots in human-robot interaction, despite differences in embodiment, roles and tasks, require a common set of social reasoning abilities. Hence, we propose a standardized data-driven method for collecting annotations on social errors and competencies across various robotic platforms and interaction scenarios. Therefore, in this paper, we make a call to the community to collect standardized annotations on social errors and competencies. To motivate this need, we bridge the literature between psychology, human-robot-interaction, and general machine learning to explain why the standardized collection of annotations of HRI data across embodiment, roles, and tasks based on social errors and competencies would enable the learning of a common embodiment, task, role agnostic set of social reasoning skills. We conduct an in-depth review surrounding the definitions of social intelligence and the tools developed to measure it, then identify the common central attributes central to social intelligence. Using these attributes, we develop and share a taxonomy of learning signals and an annotation framework that could be used to collect such data. Drawing from the recent technical advances in machine learning, we share ways in which these learning signals could be used in the context of embodied social agents to improve their social behavior.