Output token
Also known as: completion token, response unit
The unit a model measures its own answer in, which is why a pipeline asks for rankings rather than essays about the footage.
Updated
What it means
An output token is a piece of the answer rather than a piece of the question. Where Input token counts what was sent to the model, this counts what came back.
On video work the two are wildly unbalanced. Frames make the input enormous while a useful answer is often a short list, so output is usually the smaller half by a long way.
That asymmetry is worth designing around. Asking for a verdict is cheap to receive, while asking for an explanation of every clip multiplies the answer without improving the Shot scoring that depends on it.
How it works in practice
Structured answers keep the count down and the usefulness up. A ranked list of stretches with a one line reason each can be parsed directly, which is what Quality scoring consumes.
Capping the answer length is the blunt control. A limit forces the model to lead with its conclusion rather than reasoning its way there, which is normally what a pipeline wanted anyway.
Streaming matters only for interactive use. A queued job reads the whole answer at the end, so nothing is gained by receiving it piece by piece, and a Batch processing design never needs it at all.
What to watch out for
Cutting the limit too hard truncates mid answer, and a half written list is worse than a short one because the parser cannot tell the difference between a cut off and a genuinely short result.
Long answers also invite invention. The more a model is asked to say about a clip, the more room it has for a Hallucination nobody checks, because each extra sentence is another claim.
Verbose reasoning is easy to mistake for rigour. A paragraph justifying a choice reads convincingly whether or not the stated reason drove the choice at all.
How RawClip handles it
The analysis asks for rankings and marks rather than prose, so answers stay short and machine readable. A model that writes less about the footage leaves less room for anything it made up to reach the finished cut, and it keeps the whole pass fast enough to run over an upload of several hundred clips.