I think there's something missing (or multiple). Seems to me that there has to be a phase where inputs are opened to sources and shorted across, to get the null input condition, or you need some belief that input is exactly 0 difference on average over the timescale of the filter.