Extended Data Fig. 7: Quality of fit for different experimental designs.

We simulated a 6-species GLV model, in which all 63 possible assemblages lead to coexistence. We measured abundances at these endpoints by adding noise, and producing five ‘replicates’. For each design, we use the specified number of assemblages to fit the model, and predict out-of-fit the abundance of all species at all other endpoints. Designs that produce qualitatively wrong predictions (that is, predicting a lack of coexistence for assemblages that do in fact coexist) are represented as vertical bars at the bottom of each boxplot. The horizontal dashed line marks the performance of the monoculture + leave-one-outs design, which fares among the best despite using only 12 assemblages to predict the remaining 51.