video-production-harness

Non-negotiables (every state, every host)

1. Copy the reference. Study the whole reference, picture and sound. Keep its people, action, camera, hook, length (±25%) and its shot and setup counts (±30%, from scripts/reference_shots.py); list every change in changes_from_reference, shown first at the script approval.
2. Match its sound and pace. Speech, music and effects follow the reference. Silence only when the reference is silent or the owner asks for it. Never a sine tone, a drone or an empty track. scripts/reference_rate.py measures the reference's words per second (never estimate it); a voice more than 10% slower is regenerated or sped up (at most 1.25×) before lip-sync and picture timing.
3. People and demos stay people and demos. Plan a person, hands or a live demonstration in the reference as one; never plan a text card, dots or a still in its place. One that ends up on screen anyway is an ask finding you recommend fixing.
4. The script comes from write-video-ad-script. Fetch it with catalog_fetch; never write the script inline.
5. Use the real product. Its real photos, its real size and its usage steps. A still packshot gets the real photo through overlay. A pack held or used on screen is made with openai/gpt-image-2/edit, the real product photo as reference image 1, front label facing the camera. A blank, unprinted or label-less brand pack is must_fix. Every use step and product state (dry contents, cooked dish, strip on the nose) comes from a real brand photo sent as a reference.
6. Look real. Phone capture, deep focus, no bokeh, a lived-in room, no retouching or beauty filter, clothes that fit the activity. Every photoreal still and every motion prompt, B-roll included, carries the one light and realism block in libraries, look and rubric: light from a window or lamp, normal exposure, no HDR, matte skin, no over-sharpening, the first frame's light kept.
7. Review the final against the reference, listened or measured, never claimed. scripts/qc_evidence.py must exit 0 on the exact final bytes (it fails only a broken file); then mark the video complete and read the host's final check.
8. Two owner approvals, both in this chat: the script and the ingredients. In review_mode: auto one yes to the full total covers both (see check severity). Agent checks add no other approval step.
9. Quality warns, never blocks. Every finding is must_fix (always redone, never delivered), ask (shown with the price of the fix; in review_mode: auto you decide) or fine (kept). Only money, consent, safety and broken files stop the run. A redo is paid from the approved total when it fits, otherwise after a new yes to its price; never drop a scene. See check severity.

Reading rule. Read one file per call. If the output says it was truncated, re-read that file in parts before acting.