What it does
- Plans lessons with a local language model and turns approved work into queued jobs.
- Runs still-image and video generation in separate GPU stages.
- Records receipts, verifies delivered files, and keeps earlier attempts when a take needs another try.
- Adds graphics and captions, then exports a master with its licence lineage.
- Gives me a browser interface for checking work, selecting takes, and reviewing results.
The local Krea and MiniMax H3 generation lanes are labelled COMMUNITY, with model-specific conditions. Each output retains the most restrictive label in its complete stack.
What was measured
A recorded integration check on 10 October sent the same saved lesson-choice comparison in both candidate orders. Both calls returned and parsed successfully; neither selected my preferred answer. That small check established that the components could exchange the reply. It left the judgment-quality question open.
What I learned
I need to inspect the answer even when every component reports success. Keeping the attempt, delivery receipt, and my own review together makes those disagreements useful.
Status
11 October 2026: actively developed. The integration result above is a dated check; batch completion and output quality require separate validation. The source repository is private.