(a) Schematic showing how we measured dynamical correlations at the strain and species level for a species pair A and B. For any species pair, we defined strain–strain correlations as the highest magnitude of correlation among all four strain pairs from different species (left), never the same species. The species–species correlation for the same pair was simply the correlation between the two species (right). (b) Scatter plot of the dynamical correlation between species in a community and the highest correlation between their corresponding strain pairs. Each point represents one species in one of the 10 communities. The shaded region indicates strain–strain correlations higher than species–species correlations. Inset shows a pie chart of the fraction of points supporting higher strain-level interactions (76%) versus species-level interactions (24%). Triangle indicates a pair of Achromobacter (red) and Delftia (blue) species, shown in (c). Top: relative abundance plots of two uncorrelated species measured over the experiment. Bottom: relative abundances of the minor strains for the same species, which are strongly negatively correlated. (d, e) Schematics of our models showing how species are split into strains, with tunable differences in their consumption rates for each resource, as well as the serial dilution protocol that we simulate, where we slightly change the growth medium from transfer to transfer. (f, g) Scatter plots of the expected dynamical correlations using our models, (f) where strains are ecologically distinct (hypothesis 1) and (g) identical (hypothesis 2), similar to (b). Schematics of the consumption rate matrices for both models (hypotheses) are also shown.