For a while now I've been building Workshop, and I just put out a paper on the idea underneath it. Here's the short version. Generative video is impressive and, for real work, unusable. It misspells the text on screen. It drifts off brand between runs. Ask for the same video twice and you get two different videos. None of that is acceptable when the video is a product announcement, an earnings summary, or a thousand personalized ads that all have to stay on message.
The usual fix is templates: reliable, repeatable, and rigid. Every video comes out looking like the last one.
I don't think that trade is real. It comes from letting the model hold final authority over the finished artifact. So Workshop splits the job. A language model proposes: mood, narrative, copy, emphasis, written in a small readable format. A compiler authorizes: it decides what is valid, resolves the typography, color, timing, layout, and motion, and emits a plan that renders the same way every time. The model proposes; the compiler authorizes.
The payoff is that the output is deterministic. Same input, same frames. Every decision can be inspected before anything publishes, the text is exactly the text, the brand is exactly the brand, and an agent can produce thousands of variants without any of them going off the rails, because the model never had authority over execution in the first place.
This video summarizes the entire architecture simply
The paper formalizes this as an authority model: which component is allowed to decide that a video is valid. It is not really about video. It is a pattern for anywhere we are starting to let models author things that someone then has to trust. I published the grammar and the intermediate representation openly, because I would rather they become standards than secrets.