# Original-only atomic analysis — next-session contract

Status: brief prepared; the new atomic study and its separate report are NOT complete.
This file records the latest owner direction, not newly observed source findings.

## Input and output

Input: the exact original Resilia MP4 identified in `NEXT-SESSION.md`, with both
picture and sound. No generated-film comparisons, adaptation scripts or cast passports
enter the new model context. Existing source notes may be checked AFTER independent
inspection, with disagreements preserved instead of inherited as truth.

Output: `delivery/original-atomic-analysis/index.html`, with original media fragments,
and evidence under `analysis/original-atomic/`. Keep the report separate from production
archives: only source analysis, segmentation/reasons, meta-function and its reasoning.

## Numbered analysis contracts

1. **Identify and watch.** Full original AV → hash/probe, actual video-model request,
   response and whole-film map. Validate source identity, sound access, real response
   coverage through the ending, and disclose truncated/uncertain observations.
2. **Split and examine.** Whole-film map plus source AV → complete scene/shot/event
   ledger. Inspect overlapping AV around cuts; use real frames/PTS to adjudicate model
   timestamps. Distinguish a cut from an actor/prop event inside a continuous shot.
   Validate no unexamined timeline gaps and no duplicate events hidden by overlap.
3. **Explain the construction.** Ledger → scene cards with original before/during/after
   excerpts, observed actions/sound, information and character-state changes, and reasons
   for each boundary or retained shot. Validate that explanations point to actual events.
4. **Derive the meta-function.** Checked cards → callable conceptual contract/pseudocode,
   inputs, state transitions, shot/sound/action decisions, output and earnedness criteria.
   Run it against contrasting real source scenes and retain exceptions. Show why a
   reaction, insert or continuous shot belongs where it does; do not make every scene
   follow one template. This stage derives a model of the original, not a product remake.
5. **Publish and verify.** Analysis and evidence → separate hosted report. Check scene
   and timeline completeness, cited snippets, playback with sound, transition context,
   mobile layout and links. Report missing evidence as missing; a green parser or a
   confident model answer is not an audiovisual-quality verdict.

## Owner feedback to investigate

The owner rejected an edit built from two similar dialogue poses as visibly jerky.
The underlying lesson to test is that a line, a generated clip and an editorial unit
are not necessarily the same thing. Examine how the original builds and connects
meaningful actions, reactions, shot sizes, inserts and voice timing.

Owner-supplied leads in the opening: “Your blood pressure…” / “That high?” /
“And the scan…”; hand contact beginning with “This isn't something we wait on”.
Verify exact wording, speaker, shot/event boundaries and gesture timing in the actual
video. These leads do not establish measured timestamps or require matching patterns
in every later scene. Investigate the whole film for additional or conflicting mechanisms.

Useful fields, where observed: utterance and its picture timing; who hears what;
gaze/hand/body/prop actions; onset, continuation and completion of gestures; reaction
latency and silence; scale/angle/eyeline/staging; cut/insert/continuous action; sound
bridges and ambience/music if present; text already present in the original; visual
and emotional changes; information withheld/revealed; question, consequence and callback.
Absence is valid evidence. Do not invent mechanisms to fill fields.

## Acceptance boundary

Describe the owner's “10/10” reference at this atomic level while keeping “why it
works” as grounded interpretation. Spend and preference do not supply measured
retention, conversion or clinical efficacy. Source claims remain source claims.
No new film, generated pictures, voice, subtitles, product adaptation or provider repair
is part of this next-session objective. Do not start a production stage after the report.