During its training, the model uses a curriculum that linearly reduces the chance of sampling simulated data from 70 % to zero, gradually shifting emphasis from simulated trajectories to real-world observations. Across 30 disease‑location forecasting tasks, Mantis 2 achieved the best aggregate probabilistic accuracy over short, medium, and long horizons, and missed roughly 30 % fewer outbreak peaks than Chronos‑2, the next‑best model. The peaks it correctly anticipated were identified more than six weeks in advance on average. However, the evaluation is retrospective; improved forecast scores do not automatically guarantee better real‑time decisions or reliability in operational settings.