Three openings.
One demo.

Artificial Analysis’s checked image-to-video-with-audio table lists Dreamina Seedance 2.0 720p at 1174 ±7 Elo and H3 Max post-trained by fal at 1195 ±10. These are different named configurations and blind viewer preferences, not your probability of an acceptable take. Neither number justifies an invented conversion rate. [5]

OpenRouter documents that frame_images takes precedence when supplied with input_references. Validate the actual route instead of silently dropping part of your direction. No live model comparison was run for this page. [6]

# One product demo; three openings

You are my production-engineering agent. Start with a film, not a studio UI.
This article is context, not authorisation to spend or publish.

## Inputs to resolve
- Actual product recording and a claim it demonstrably supports.
- Original/authorised character, scene and voice references.
- Audience, aspect ratio, and an explicitly authorised cash and human-time budget.
- Provider route/capabilities; capture a current quote for the exact configuration.

If inputs are missing, run the labelled local fixture and return a plan. Do not invent a feature, approval, viewer outcome, timestamp or price.

## Film
30 seconds: 6s hook + 14s real demonstration + 6s compatible reaction + 4s end card.
Propose three distinct hooks around the same verified problem. First create a timed script and still-based rough cut. Obtain approval before generating motion.
Keep only the tail that actually remains compatible. A changed line, pose or eye line may invalidate shared footage.

## Implement the smallest missing boundary
Keep script, approved references, current story state, voice, source takes, timeline, captions, billing records and approval separate.
Store immutable input/output hashes and dependencies. A text-only caption edit rebuilds captions, render and approval, not the source performance.
Validate exact provider capabilities; do not silently send conflicting frame/reference modes.
Reserve the next upper-bound charge. Reconcile unknown jobs; do not interpret a timeout as a free failure. Cap attempts; permit HOLD, simplified scope or rejection of every take.
Treat model observations as untrusted evidence. Missing review remains unknown. Compare delivered media, not only script descriptions.
Never publish without a separate explicit instruction and valid final approval.

## Required outputs
brief.json; shots.json; state.json; references/; audio/; takes/; jobs.json; timeline.json; captions.srt; drafts/; review.json; costs.csv.
Attach rejected takes, actual provider spend, active human repair minutes and missing-evidence notes.

## Run these cases before connecting providers
1. Group five synthetic labels into two categories; label it as a fixture.
2. Fix 'too' to 'two'; verify original source bytes remain unchanged.
3. Change a spoken line; invalidate voice, performance, dependent edit and approval.
4. Remove part of review evidence; HOLD.
5. Add 'saves 90%' without evidence; HOLD.
6. Exhaust the budget; HOLD before dispatch.
7. Receive an ambiguous timeout; retain original job identity/reservation; HOLD for reconciliation.
8. Compute 3 openings × 3 max takes: repeated generation 108 seconds; reusable reaction 72 seconds. These are maximum planned generated seconds, not quality claims.
9. Price the Fast route using measured acceptance against the same quality floor, not cost/second alone. Include discarded draft cost in any two-stage strategy.

## Validation
First execute the offline fixtures. Then propose a bounded, separately authorised matched-footage test. Freeze the brief and quality floor. Report accepted-film cost, human repair time and viewer comprehension by episode. Do not turn repeated checks on the same footage into independent evidence.
The inherited Python kit tests only local policies and a subtitle-file fixture. Passing it is not validation of model-generated films or production security.