Human in the loop
Also known as: supervised automation, review step
Designing an automated process so a person makes the few decisions that matter, instead of either approving everything or nothing.
Updated
What it means
Human in the loop describes where a person sits in an automated chain. The useful version is not review everything, which defeats the point, and not review nothing, which ships mistakes to viewers.
It is routing the uncertain cases to somebody and letting the confident ones through untouched. The whole design question is where to draw that line, and a Confidence score is what makes drawing it possible.
For editing that usually means the machine assembles and the person adjusts. Choosing which nine seconds of a long clip to keep is mechanical work; deciding that the third attempt was the one that mattered is not.
How it works in practice
The practical lever is a threshold. Anything a model is unsure about is surfaced and the rest is accepted, which keeps the review list short enough that somebody actually works through it rather than skimming.
Review should also be cheap to act on. Being able to nudge a mark a few frames, or drop one shot, is worth more than being able to rebuild the edit from scratch.
Placement matters more than volume. One look at an assembled Rough cut catches the errors that would reach a viewer, while approving individual detections catches errors that mostly cancel out.
What to watch out for
Rubber stamping is the standard failure. A reviewer shown a long list of confident model answers learns to approve without looking, so the step takes time and catches nothing at all.
Putting the person too early wastes them. Approving every Object detection result is tedious and low value, because a wrong box on one frame rarely survives into the finished sequence anyway.
An unclear remit is the third trap. A reviewer who does not know whether they are checking accuracy or taste will do neither consistently, and the same footage gets different treatment each day.
How RawClip handles it
The system assembles a complete highlight and the person is left with the decisions a machine cannot make: whether to keep it, and which moments mattered for reasons that are not visible anywhere in the footage.