The Lesson That Named Its Own Exception
Earlier in this track, the workshop argued that an episode is reproducible from the tool plus the pointer, that a re-run producing a slightly different film is a feature rather than drift, and that everything in between is disposable. The same lesson said exactly when that trade would be wrong: when a render is evidence, when other things depend on it, or when someone has paid for a specific version. On 2026-09-05 the last condition became the normal state of every episode, paid for in approval rather than money, and the workshop reversed its position. What replaced it deserves a close look, because "we need reproducibility now" is the moment most teams reach for the wrong kind.
What the Final Watch Approves
Every finished episode now passes a final watch by the person who publishes it, and that watch does not approve a script. It approves a performance. Consider what an approved film carries that the script does not. In the reference episode, a leftover warm-up word was cut out of the opening by trimming nine-tenths of a second at a measured silence, with no new synthesis at all. In others, subtitle windows were corrected where the alignment ran early, takes were chosen among re-rolls, and images were revised by reference edit. None of those decisions lives in the script, and none survives a fresh run. Speech synthesis gives a different reading every time, and some readings fall into a rushed or slowed mode that the text itself triggers, so a re-roll may not even produce an acceptable take. A rebuild from the tool and the pointer makes a film nobody approved.
So the thing worth preserving changed. The earlier lesson preserved the ability to make the episode. The approval turned one particular result into the product.
Seal, Then Replay
Accepting an episode now seals its production. The seal inventories every file and records its size and hash: the authored scripts, snapshots of the sources, the chosen images with their prompts and provenance, the untouched provider recordings, the processed performances, every declared audio edit, every review — and the rejected iterations too. Beside them it saves the render code, the exact font files, and the tool versions.
Reproduction is then a replay, not a regeneration. It verifies the bundle, builds into a separate empty directory from the archived code and fonts, and runs with synthesis calls disabled. The approved original is never overwritten. And the replay is honest about its limit: a different runtime, encoder, or font version has to be compared before anyone claims the output is equal.
Two supporting mechanisms are what make a seal trustworthy. Audio edits are declared and replayed: a trim is recorded with its measured evidence and applied from a hash-checked baseline, and the untouched provider audio is saved before any trim or breath repair touches it. And every verdict is bound to a file. Validation binds the measured duration, streams, and loudness to the final movie's hash and to its inputs' hashes; the viewing record binds to that same hash; the acceptance records approval of that identity. Landing selects that exact file — never the newest movie in the folder, and never an earlier passing run whose inputs were different.
Where It Lives, and How It Moves
One machine owns the archive. A production made on a travelling laptop completes there, seals, and uploads into an isolated incoming directory on the archive machine, where every sealed file is verified before the production is installed atomically. A different archive already sitting at the destination is a conflict, never an overwrite. If the laptop is offline, it writes a complete pending request — source, destination, procedure, and what counts as done — rather than a pointer to a handoff note, and it does not pretend the central record was updated. Meanwhile the version-controlled repository still receives only what it always did: the tool and a one-line register entry.