Currently, we follow the following process:

1. Randomly sample ~20k revisions
2. Filter our edits by trusted users to get down to ~5k revisions
3. Users label the edits

When doing a second campaign, we should use the model trained on the first campaign's data to filter the edits to an even smaller set.

1. Randomly sample ~20k revisions
2. Filter our edits by trusted users to get down to ~5k revisions
3. Filter edits that are obviously not damaging and saved in goodfaith using the old model (down to ~1 or 2k?)
4. Users label the edits

This will probably belong in editquality autolabel.