Observed emulator accuracies for ξ₊, averaged between 0.5 and 72 arcmin, with the grid ensemble of trial cosmologies, , shown by the colour maps, when trained on the 26 cosmo-SLICS predictions (upper-left) and 26, 50 and 250 noise-free NICAEA predictions (upper-right, lower-left, lower-right respectively). The training nodes are shown by the magenta circles. The black stars in the upper-left panel show the input parameters of the Takahashi et al. (2012) simulations over our parameters volume (their two highest Ω_m nodes have h and S​₈ values that exceed our boundaries). For each grid in which two cosmological parameters vary, the remaining two are fixed to the corresponding fiducial values from {Ω_m = 0.3251, S​₈ = 0.75245, h = 0.7082, w₀ = −1.254}. The contrast between the upper panels, for which the training cosmologies are the same, indicates the extent to which simulation noise and inaccuracies in both the simulations and theoretical predictions degrade the apparent emulation accuracy.