Researchers have introduced a method to transform the deterministic Aardvark Weather model into a probabilistic forecasting system capable of disentangling different sources of uncertainty. By attaching specific stochastic mechanisms to each architectural component, the authors separate aleatoric uncertainty, which is captured via learned, input-dependent noise at the observation encoder, from epistemic uncertainty, which is captured through Monte Carlo dropout within the model processor. This approach allows for a formal law-of-total-variance decomposition to attribute forecast spread to either the observing system or the learned dynamics. Experimental results demonstrate that probabilistic finetuning enhances mean forecast accuracy by an average of 4.2% across various variables and lead times. When evaluated against ERA5 data, the resulting ensemble shows a spread-skill ratio of 0.98 over the medium range. Furthermore, while station RMSE remains within 2.4% of the deterministic baseline, the probabilistic model achieves superior Continuous Ranked Probability Score (CRPS) metrics at every lead time. Although the model still trails the operational ECMWF ensemble in performance, this component-attributed uncertainty framework provides a more transparent path toward developing observation-driven digital twins of the atmosphere by distinguishing between errors inherent in observations and errors in the model's predictive logic.