# Final-film QA: what is accepted and what is rejected

The complete 349.458-second export was inspected through one whole-film and five bounded Kie/Gemini passes. These are model outputs, not ground truth. The editable timeline, actual word boundaries, media probes, listening source and decoded pixels remain authoritative.

[Raw whole-film pass](07-adaptation-final-audit/analysis.md) · [opening](08-final-opening/analysis.md) · [reveal](08-final-reveal/analysis.md) · [application](08-final-application/analysis.md) · [payoff](08-final-payoff/analysis.md) · [ending](08-final-ending/analysis.md)

## Accepted with independent corroboration

- The opening mother/photo question, neighbor product reveal, adoption routine, birthday group and final order/care close appear in the actual film. Bounded excerpts place those events correctly; the approved timeline and inspected frames agree.
- The mother is visibly included beside the daughter and father during the birthday payoff. The final separate framed photograph depicts the parents.
- The bounded passes independently transcribe the beginning and ending phrases of their excerpts in Ukrainian without a supplied transcript. They find clear spoken Ukrainian and readable captions. Full normalized word matching and caption coverage are verified separately.
- Small hand, label and background-object approximations are visible in reviewed frames. Their severity varies; the 480p result does not support an assertion of perfect anatomy or label microtext.

## Rejected or not established

- **Whole-film global times:** The model assigns laptop, tea and bathroom events to incorrect times and then attaches defect suggestions to those locations. Those exact defect ranges are not used as editing instructions.
- **Cosmetic improvement:** The broad model calls the generated texture a hydration improvement. Lighting, acting and generated pixels cannot establish a product effect. The script and report do not claim that result.
- **Continuous music:** Four short passes hallucinate soft background music, while the whole-film and reveal passes say none. The actual mix contains exactly one TTS narration asset. All 37 generated clips have no audio streams. The final decoded voice matches that master at the intended gain with −63.53 dB RMS subtraction residual. Direct PCM samples in speech gaps in those four excerpts are recorded in [silence-evidence.json](../../delivery/silence-evidence.json); they contradict continuous audible music. No music was created or added.
- **Extra filler at global ~206 s:** One pass hears “е-е.” No extra lexical word appears in the actual provider alignment. A vowel/hesitation interpretation is not promoted to a confirmed missing caption word.
- **High confidence is not calibrated accuracy:** The short passes correctly locate actions but can still infer generic hand artifacts or music. Their confidence labels are retained as provider output, not adopted as reviewer certainty.

## Decision

Accept the film as the completed fictional narrative adaptation, with the disclosed generative detail and wardrobe limitations. No blocking story, language, timing, caption-layout, decoded-media or delivery failure was found in the checks performed. This is an automated/analyst production review, not a native-human listening panel, every-frame anatomy certification, clinical evaluation or campaign test.