The initial phase of the inference is monitored by following the systematic drift of the power spectrum and two configurations of the reduced bispectrum towards the fiducial ground truths. After a few thousand samples, we approach the regime of per cent-level agreement with the power spectrum and bispectrum. More samples are needed to enter the typical set from which plausible Gaussian initial conditions can be sampled. Model mis-specification between the ground truth (N-body) and the BORG-EM predicted fields shows up as small discrepancies at the largest scales, as a consequence of putting most weight on fitting the smallest scales where most statistical power resides.