During a pairing session with @BagToad, we caught a few points to improve the evals in our AI spam detection.

There's no explicit E/O or A/C, for this, and I'm going to push a PR for this right after creating the issue.