This will fix the vast majority of problems this generated
https://github.com/discourse/discourse/commit/956e3ad2088a9ebe09338b94689f91d5975753e9
It looks for “auto change trust level” something that happens quite a lot being mislabeled as “check email” and for “impersonate” that was mislabeled as “check email”. Once it finds the bounds it takes one off the action number.
It logs what changes it makes in case we have to revert them. We can add a few more conditions but it should vastly improve the data quality.