The point about deciding what "accepted" means before collecting data also works for scoping a paid offer. If the buyer's acceptance test is written into the scope up front, such as the inputs you need, the format of the deliverable, and how many revision rounds are included, the human cleanup time stops being an open-ended cost. It also lets you see which task types lose money after review time is counted, so you can narrow or drop those before repeating them. A blended average across all work would hide exactly that.