Quantifying stability of non-power-seeking in artificial agents