Time-to-Detect a New Generator Family

A new open-source face-swap model is released publicly on a Friday. By Monday, your reviewer-confirmed-fake rate has spiked, and your detector's average score on manually-confirmed fakes from this new model is noticeably lower than its average score on fakes from generators already in your zoo.

1. Walk through the full response loop, in order, from "new generator appears" to "detector reliably flags it" — name each stage.
2. Why is this fundamentally a different kind of drift than the slow, seasonal feature drift covered in general ML monitoring material, and what does that difference imply about how you'd staff and tool the response?
3. Propose two concrete leading indicators — signals available before "average score on confirmed fakes drops" — that could have caught this earlier.