As far, as I understand Blum's paper, he doesn't talk about general non-monotone networks, but rather about "standard networks", where all 'not' gates are moved to the front: "The resulting network is a so-called standard network where only input variables are negated."
Could it be, that Tardos' function has exponential monotone network complexity, exponential standard network complexity (with inverters allowed at front), but polygonal complexity in general networks (with not gates/inverters also allowed in the middle)?
Could it be, that Tardos' function has exponential monotone network complexity, exponential standard network complexity (with inverters allowed at front), but polygonal complexity in general networks (with not gates/inverters also allowed in the middle)?