What makes a clip work on its own

The first version of moment detection scored segments on how much people talked. It found the loudest ninety seconds in every recording, which turned out to be almost never the most interesting ninety seconds — and the clips it produced started mid-sentence and ended mid-thought.

The problem was the question. “What is the most engaging part of this recording” is unanswerable without knowing who is watching. “Which parts of this recording still make sense with the other fifty-eight minutes removed” is a much narrower question, and it has an answer.

Four properties

A clip that works alone tends to have four things, and a clip missing any one of them tends to fail in a way you can name.

• An opening: the first sentence has to be a reason to keep watching, not a run-up.
• Standalone: nothing in it depends on something said earlier that is not in the clip.
• A payoff: the idea lands inside the clip rather than in the minute after it.
• Pace: enough happens per second that the middle does not sag.

These are what the clip score is made of, and each one carries a sentence about this particular clip rather than a bar on a chart. That is deliberate. A number you cannot argue with is a number you either believe completely or ignore completely, and neither is useful when you are deciding what to publish.

A score you can disagree with is more useful than a score you cannot check.

What it still gets wrong

Running jokes. A callback to something from forty minutes earlier reads as a standalone failure, because it is one — the clip genuinely does not make sense alone, even though it is the best thing in the episode for the people who heard the whole thing.

We have not solved that, and we would rather score it honestly and be overruled than quietly promote it. Overruling is one click; a scoring model that flatters itself is a model nobody checks.