A local-first video production engine. A human authors a
ProductionSpec; local HuggingFace models synthesize the narration, footage, and music inside
it; pure-Python verifiers decide what survives to the final cut.
Where my-AI-story turns a series bible into multi-voice audio, my-AI-scene turns a production guide into a scored, color-graded, assembled video — the engine behind The Journal of Informal Human Protocols. Human-owned constraints → local AI synthesis → deterministic verification at every boundary.
The ProductionSpec — every beat, narration line, footage prompt, and grade — is authored before synthesis. The schema is the contract the whole cut is judged against.
Local HuggingFace models fill it in — Kokoro narration, SDXL/FLUX stills, MusicGen beds — beat by beat. The models are components, not the system.
Whisper transcribes narration and compares it to the script; CLIP scores each still against its prompt; ffprobe verifies the final file. Blocking gates protect the cut; music is non-blocking.
my-AI-scene reliably produces a narrated, illustrated short where every kept frame cleared a gate — the worst case is a slightly simpler video, never a corrupt one.