Box models are the standard approach to modeling and predicting the decay of soil carbon into CO2. In this paradigm, models are made more expressive by adding boxes (e.g., another type of soil carbon) or catalysts (e.g. microbes). Yet such model changes always add fitting parameters and sometimes increase the intrinsic complexity of the model by making it nonlinear. This is a particular challenge in the context of soil carbon, where we have quite limited data on carbon stocks and stock changes, meaning that even relatively simple box models are challenging to fit responsibly. We borrow an alternative statistical approach from studies of organic carbon in ocean sediments. Instead of a finite number of boxes with discrete kinetics, this 'disordered kinetics' approach describes a continuum of decay kinetics described by a statistical distribution (e.g. a log-normal or a power law). Focusing on data reporting on the 10-100 year dynamics of soil carbon, we show that models of this type perform better than standard box models with many fewer parameters.