Almost Nothing Survives a Run
A finished episode leaves behind a working directory full of substantial artifacts: the fetched material, the brief, the script, dozens of image plates, dozens of audio takes, video clips, the final render, and a long log of every judgment made along the way. None of it is committed. The repository gains, at most, one line in an index file.
The reasoning is a single sentence: the episode is reproducible from the tool plus the pointer. The material already lives properly in the system that owns it. The tool is versioned. So everything in between is derived, and derived things do not need to be stored — they need to be rebuildable.
The Uncomfortable Part
Here is where this design says something most engineers would flinch at. Re-running the pipeline on the same pointer does not produce a byte-identical video. Speech synthesis is not deterministic; generated art is not deterministic. The second render would be a slightly different film.
And the doctrine's response is that this is a feature, not drift — the variation is treated as acceptable rather than as a defect to engineer away. That is a genuinely unusual position and it is worth understanding rather than copying blindly. It works because of what reproducibility is for here. Nobody needs to prove this exact file was produced by that exact code; there is no audit requirement, no regulatory replay, no bug that will be diagnosed by diffing two renders. What people actually need is the ability to make the episode again if the file is lost, and a slightly different take satisfies that completely.
Change the purpose and the answer changes with it. If a render were evidence, or a build artifact other things depended on, or something a customer had paid for a specific version of, then non-determinism would be unacceptable and every generated asset would need to be pinned. The lesson is to ask what your reproducibility is actually protecting against before paying for the strong version of it.