Review for NeurIPS paper: Certifying Strategyproof Auction Networks

Additional Feedback: 49, 59, 74 – The discussion here talks a bit loosely about certifying the strategyproofness of auctions. However, this is not quite what gets certified, and this is an important caveat that should be made more explicit. In particular, what gets certified is the amount by which agents can manipulate on particular type profiles. This makes no guarantees about the incentives on any other type profile, so it is a bit misleading to describe the approach as certifying the extent of strategyproofness full stop. The abstract (11) is more careful and explicit about this. Based on the description of the 3 trained networks (225-230), none of them seem to use it.