# Fresh source study: adjudication and limits

The pinned original hash matches. Local probes establish source duration 325.681678 s
and previous Omni master duration 65.625 s. Original source remains read-only. Complete
AV proxies preserve both streams; original excerpts cover 0–325.681678 with no gaps.
Contact sheets contain 109 source frames/33 previous-film frames. These are sampled
direct visual checks, not continuous frame-by-frame inspection.

Gemini's full-source narrative/style passes and the previous-film pass received actual
complete picture and sound. Nine original excerpt requests supplied actual AV with
known source offsets. All returned evidence remains raw and immutable. The 240–280 v1
response stopped after its second second; it is rejected as incomplete. Its v2 was
reviewed through the ending. No provider completion alone establishes source coverage.

Corrections to raw model output:

- The source is spoken **English**. A draft extraction-review prompt mistakenly named
  Vietnamese; it was stopped before any excerpt request was submitted and corrected.
  Whole-video prompts did not presume the language. All retained excerpt input receipts
  use language identification from actual sound.
- “Dad is lying” is an interpretation of reassurance versus visible worry, not directly
  observed knowledge of the father's intention. “Bypasses skepticism” is a hypothesis,
  not a measured viewer response.
- The previous-film reviewer called the mother “Larisa.” The production identity is
  Oksana; Larysa is the neighbor. It also asserted the final spoken line “Don't crop
  anything,” whereas the visible caption around 56 s reads “Сьогодні ти поруч.” Do not
  adopt that source-free rewrite as a transcript.
- Whole-model exact times, strict 180-degree adherence, “high-speed shutter,” precise
  production technology and a sharp reality shift at 4:59 are unverified. Sampled
  frames support broad staging/light differences, not those technical conclusions.
- The previous reviewer called its timecodes measured. Only ffprobe runtime and local
  extraction offsets are measured here. Emotional scores and timing are model estimates.
- “Suspiciously identical lighting” in previous proof frames is not a continuity failure:
  matched light/camera were intentional comparison controls. Actual smoothing/age drift
  needs pixel review; source-matching cinematic lighting is a separate criterion.
- Some excerpt models speculate that the mentor is a family member. Doorway recognition,
  the child's friend-at-school lead and the later home visit support the neighbor role.
- The 240–280 v2 calls a package ingredient “Lecithin”; visible source packaging elsewhere
  reads CoQ10 ubiquinol. This ingredient detail is not adopted. Exact medical numbers
  are unnecessary for our cosmetic adaptation and remain source ad assertions.
- The source contains incidental music according to excerpt listening. Our owner's
  narrative production still explicitly excludes a music stage; do not call this a
  literal reproduction of the entire original soundtrack.

Independent source discoveries carried into the skill: different audiences justify
repeated explanation; hospitality changes the scene before teaching; a physical search
has a relationship consequence; accepting a trial leaves residual doubt; gratitude
returns from mentor to seeker; proof returns to the opening witness and hand motif.
No clear multiple-failed-products or envy sequence was found. The original contains
substantial overt sales exposition and unverified medical/competitor claims.

The new owner rule requires actual matching originals at every generation. The current
text-only speech endpoint cannot supply that modality; voice generation has not started.
The native Omni endpoint supports an original video excerpt but changes duration
behavior when video_list is used. Its source-fidelity behavior is not yet tested here.