In this paper we follow up on the results from our previous publication (Lobsien et al 2018 Nucl. Fusion58 106013), where it was found that stellarator coil design optimization can be substantially improved by using a stochastic optimisation approach. In that paper performance was quantified by lower (better) and more narrow (more robust) distributions of the penalty functions at the end of the optimisations. Here, we evaluate and compare the various coil sets of the previous paper but seek a verification and deeper understanding of the physics performance by replacing the relatively simple penalty function estimate with more accurate ones from state-of-the-art calculations of MHD stability, neoclassical transport in the -regime, fast particle confinement, and gyrokinetic behavior. The investigation shows that stochastic stellarator coil optimization generally outperforms the earlier non-stochastic stellarator optimization, also when using these more accurate metrics, generally confirming and quantifying the better performance. We do find some discrepancies, indicating that the penalty function does not represent a physics performance optimum perfectly. For example, as pointed out by others before us, the depth of the magnetic well is not a sufficiently good proxy for MHD stability, and the neoclassical transport can be significantly reduced in configurations that have relatively high field errors and therefore high penalty values. Thus, our work points to areas where better physics model inside the optimisation loop are needed, than what is currently represented by our penalty function.