In the DMOJ implementation of elo-MMR, a contest was recently run with a rating ceiling of 1800, but we saw performance values in excess of 3000. There is a belief that these performance values do not make sense (for example, in CF this would not happen, and in AtCoder there is an artificial performance ceiling imposed).

What adjustments should we make, if any, when there are contests which impose a ceiling on the ratings of participants? It seems highly unintuitive that a contest which bounds above the rating of the participant can yield performance values greatly in excess of the bound.