Clipping Channel Bench live

Weekly median views per clip across the ten accounts in each arm. The adaptive arm trails the control through week three while it is still learning which moments carry, and gives weeks six and nine back to stretches of catalogue that had nothing in them.

Every run cuts from the same 412 hours of rights-cleared long-form across six source channels, on ten fresh accounts, for ten weeks. The catalogue, the reframe targets, and the posting cadence are identical in every arm, so the only thing that moves between systems is which moments they choose to cut and how they open them.

A clip counts as shipped only once the operator has approved it, and views are read 72 hours after posting and taken as the median across the run — one break-out cut cannot carry a system. Hold at three seconds comes from each platform's own retention curve and is the signal that separates a cut with a reason to keep watching from one that merely started.

The adaptive arm rewrites its own scoring rubric, opening frames, and caption style from what the last fortnight returned, and re-mines catalogue it has already been through under the new rubric. The control arm is the same team with those three frozen at their week-one settings, and it is what the multiple in the headline compares against. Cost per clip is everything billed from ingest to published — model tokens, render, storage, and the posting calls.

We define and run this benchmark ourselves. The other harnesses ran with stock tooling and our prompting, so read the ordering as directional rather than as a settled ranking of the harnesses.