The notion of computation is ubiquitous in discussions of the aims and advances of cognitive science and neuroscience. This is partly a practical matter, since computational models are used in the analysis of everything from behavioral performance in cognitive tasks to massive brain imaging data sets to fine-scaled neural activities. One plausible alternative understanding of talk of representational states or computational processes in neuroscience is as shorthand for physical states and physical processes that do not otherwise have a convenient taxonomy. Evaluating computationalist hypotheses about neural activity must be a holistic enterprise – requiring simultaneous commitment to a package of hypotheses about content, conventional mappings for computation, and explanatory depth. K. Hardcastle et al. argue that any reliable statistical relationship should count as encoding, and explore a few relatively simple non-linear ones. This allows for much more flexibility in potential codes.