How we knew COVID was over (and what our models had to unlearn)

This post refines the ML model lifecycle management framework by detailing three distinct actions: refitting, respecifying, and holding a model. It elaborates on the failure modes associated with each action, such as 'chasing noise' with off-cycle refits, 'carrying ghosts' by retaining stale assumptions, and 'respec-as-panic' through overcorrection. The post introduces a decision-making process based on whether the underlying data generation process has changed, providing a flowchart to guide these choices and emphasizing that updating models is not free.