Statistical Framework for Lifelong Expert Model Pool Management

This method optimizes streaming systems by deciding whether to reuse, spawn, or defer expert models based on conditional discrepancy. The framework uses a restarted e-detector to ensure finite-time anytime validity and conservative risk management in expert selection.