ASE's CP2K wrapper does not warn if computations fail to converge. So, we'll get the occasional redox energy that uses an unconverged energy value.
In lieu of fixing ASE (not yet sure how), we should add an outlier screening step to prevent these bad energy values from interfering with learning.