The rise of large-scale machine learning applications consuming unprecedented amounts of data necessitates throughput-oriented machines. Dataflow computing has thus regained traction lately due to its ability to capitalize on fine-grained parallelism in both the vector and pipeline dimensions. Furthermore, modern reconfigurable dataflow accelerators (RDAs) provide the ability to express flexible memory orderings, enabling algorithm developers to explore detailed synchronization schemes tailored to each application. However, there lacks a generalized framework for reasoning about event orderings that suits the unique constraints of RDAs like strict pipeline timing. Such an abstract model can also provide stronger correctness guarantees by making the programming model more amenable to formal verification. In this talk, we describe the shortcomings of existing frameworks and propose a new approach called Tags based on a generalized notion of the token-credit system. We also outline potential future works in integrating Tags into new languages and compilers, as well as exploring relationships between Tags and session types.