ERA-U: A Convention-Adaptive Evidence Interface for Window-Based Time Series Classification

Window-based time-series classification (TSC) often faces the multi-class window problem: one observation window contains several states but receives one convention-dependent label. We introduce ERA-U, an explicit, no-bypass factorization of pre-aggregation temporal class scores and a low-capacity positional aggregator. ERA-U does not enlarge the hypothesis class of positional weighted pooling followed by linear classification; instead, it makes the intermediate class-score sequence directly available for diagnosis and defines a restricted interface for supervised convention adaptation while keeping the temporal representation fixed. Across five real-world datasets, ERA-U improves mixed-state-window recognition, reveals that many errors select competing states present in the input, and leads conflict-window Macro-F1 in five of six directed convention transfers while updating only nine aggregation parameters or a 525-parameter lightweight interface. These results establish convention shift as a distinct TSC problem and show that an explicit, restricted prediction interface can support accurate diagnosis and low-capacity adaptation.