In the first part of this talk, I will present some research in progress on time-series-based forecasting.  In particular, I will discuss a method by which the performance of a forecasting model may be improved in a model-agnostic manner, by making use of additional information external to the model: the known past performance of the model in predicting actual observations. The success of the method depends on a balance of information gained from using the time-series of previous forecast performances and the information lost due to fitting noise. In order to gain control over this information flow, I introduce a method for probability measure estimation based on Gaussian Process modeling that allows explicit closed-form computation of Shannon entropies.

In the second part of the talk, I will tell a story of how an ostensible algorithmic failure in Magneto-Hydrodynamic (MHD) code implementations of the Biermann Battery effect led to a re-examination and to a deeper understanding of the physics associated with the effect, as well as to the discovery of a new phenomenon that is potentially observable in laser plasma experiments.