The digital platforms that now mediate a substantial share of social and economic life present institutional theory with a defining paradox. On the one hand, platforms produce behavioral trace data of a granularity and scale unimaginable in the era of classical institutionalism-billions of timestamped reactions to informational signals. On the other hand, the foundational definitions of the discipline, particularly those inherited from Douglass North and W. Richard Scott, fail to travel to this data in any operationally coherent way. Drawing on a systematic review of the A/B testing architectures of Meta and Google, this study demonstrates that the conceptual apparatus of institutions-as-rules is not merely difficult to operationalize but is structurally incompatible with how real behavioral patterns emerge, stabilize, and dissolve on digital platforms. The paper documents three empirical fracture points: the algorithmic disruption of random assignment in platform experiments, the systematic exclusion of demographic groups from informational signals via opaque machine-learning optimization, and the platform-as-institutional-environment problem, wherein the signal generator itself is the primary architect of behavioral spectra. Anchored in Rapoport's subjectivism of system-identification, we propose that institutions must be reconceived as empirically distinguishable, statistically stable clusters of behavioral trajectories-carved from data by the analyst, not discovered as pre-existing entities. This redefinition provides the conceptual foundation for a full three-tier research program encompassing base definitions, a theoretical explanatory scaffold, and a suite of mathematical operationalization tools.