Marginal eccentricity posteriors from leave-one-instrument-out cross-validation of RV data sets. The legend specifies which data set is ignored. We find that the GP model (left) is more predictive of held out data than the model without a GP (right). Dropping the HARPS (pre) data from the no-GP model also makes the eccentricity nearly consistent with zero ([0.02, 0.12]).