This study aims to critically reinterpret the history of multimodal corpora through the lens of the epistemological tension between two key paradigms for constructing scientific knowledge: ‘control’ for pursuing internal validity and ‘naturalness’ for pursuing external validity. To this end, this paper first systematically outlines the philosophical foundations of science and methodological criteria for each paradigm. Based on this theoretical framework, it analyzes cases of the control paradigm, represented by M2VTS and IEMOCAP, and cases of the naturalness paradigm, such as FOLK, CEJC, and CANDOR. It also examines the characteristics of contemporary corpora like MISP, where the two paradigms intricately intersect. Furthermore, this analysis is extended beyond the audio-vocal channel to the field of sign language corpora, which centers on the visual-motor channel, to analyze the unique methodological dilemmas faced in sign language corpus linguistics. Finally, synthesizing the entire discussion, this study proposes an analytical framework presenting three potential design paths (maximization of internal validity, maximization of external validity, or a compromise) for future corpus researchers to consider according to their research objectives.