We tested the agents on real reels, then fixed what the judge found.
Every reel went through the Studio's own agents, round after round. A Muse Wonder judge scored every moment, code checked the rest, and each weakness was traced to its cause, changed and measured again.
Every moment, scored. On the strictest judge.
From a report to a daughter talking.
Code counts the habits that make a message sound like a machine: "this reel shows", a self-introduction, a disclaimer, describing the person on screen. The share of moments with at least one, per round:
Every format, on the same reels.
The same reels, the same judge: each moment format before and after the voice changes, and how well each kind of reel comes across, in aggregate.
Checked before anyone sees it. And every connection was tested.
The send gate is code, the verifier is Muse Wonder, and a grandchild's set gets its own child-safety review. Held packs are the checks working: a person reads them first.
Each weakness traced to its cause.
Every change is logged with the weakness behind it and the round it was made in.
The data for a tuned model. And a child model, trained.
The approved moments became a dataset in the family's voice. A small child-safety model was fine-tuned and published.
The whole lab, priced call by call.
Every model call is logged with its price. A round is all usable reels through the agents, with the judge.