Scatter in the posterior mean of SBI with respect to the true parameters. The true parameters are fixed to the fiducial parameters π⁰ at approximately the centre of our prior to reduce truncation of the posterior at the prior boundaries which is more significant at lower n_s (see Eq. 7). The bias in the S_diag. estimated covariance is due to ignoring the cross correlations in the datavector - this applies similarly to the neural network (labelled ‘NN’) variances. The similarity of the points in these plots to the plots of the marginal variances of the SBI posteriors against the number of simulations shows that the posterior mean does not scatter significantly more than the MAP.