nasa-gistemp

A small DDPM built from scratch, generating a distribution of next-month temperature anomalies instead of one point estimate. Calibration is reasonable, close to an 80% coverage target. The central prediction overreacts to its input, losing to a linear baseline. A fix attempt made this worse, not better, reported directly.