Great breakdown, and "most of the clock time is not creative" is true. One extra filter for the clip scorer: check whether the first sentence of each candidate makes sense with zero context. A lot of the best moments in a talk start with "and that's why" or "so the second thing", which kills a Short even if the middle is great. Either drop those or let the scorer trim to the first self-contained sentence. For talks where the speaker is static, a two-second cold open of B-roll that shows the payoff before the face appears can lift retention a lot. I make those hook shots in ClickPop and drop them in front of the cut; the rest of your pipeline would stay the same.