> **⚠️ SUPERSEDED** — This document is a historical receipt. See `s7-current-understanding.md` for the authoritative current position.

# Totally Spies S7 animation validation — expanded corpus pass

## Purpose

This note validates the earlier conclusion about Season 7's animation type and
style using a much broader official source set than the initial trailer-only
analysis.

Earlier conclusion under test:

- clean vector-like linework
- flat color fill / digital ink-and-paint look
- digital cutout / rigged 2D animation in Toon Boom Harmony
- limited-animation motion discipline with many held or restrained beats
- main AI-video challenge is temporal consistency, not realistic physics

## Corpus used

Official English YouTube downloads now available locally:

- **104 official videos** downloaded
- **85 pure S7** uploads
- **19 mixed `Seasons 1–7` compilations** that still include S7 material

Core receipts:

- inventory:
  - `materials/reference/totally-spies/youtube-official/inventory.csv`
  - `materials/reference/totally-spies/youtube-official/inventory.json`
- all-video midpoint frames:
  - `materials/benchmark/youtube-s7-validation/catalog-frames/`
  - `materials/benchmark/youtube-s7-validation/catalog-sheet-01.png` … `catalog-sheet-07.png`
- continuity split:
  - `materials/benchmark/youtube-s7-validation/continuity-material.json`
  - `materials/benchmark/youtube-s7-validation/continuity-sheet.png`
- benchmark split:
  - `materials/benchmark/youtube-s7-validation/benchmark-material.json`
  - `materials/benchmark/youtube-s7-validation/benchmark-sheet.png`
- motion validation pack:
  - `materials/benchmark/youtube-s7-validation/validation-selection.json`
  - `materials/benchmark/youtube-s7-validation/snippet-manifest.json`
  - `materials/benchmark/youtube-s7-validation/motion-hold-sheet.png`
  - `materials/benchmark/youtube-s7-validation/motion-action-sheet.png`

## What was extracted

### Wide catalog extraction

For **all 104 videos**:

- one midpoint catalog frame per video

### Focused motion extraction

For **44 representative videos** spanning trailers, clips, compilations, full
episodes, bundles, music videos, and character promos:

- one 2-second midpoint motion snippet per video
- four sequential validation frames per snippet

## Expanded model pass

Model used:

- local Ollama `qwen2.5vl:7b`

Saved raw outputs:

- `materials/benchmark/youtube-s7-validation/expanded-validation-qwen2.5vl-7b.json`

Prompts run on:

1. `style-validation-sheet.png`
2. `motion-hold-sheet.png`
3. `motion-action-sheet.png`
4. `continuity-sheet.png` + `benchmark-sheet.png`

## Results

### 1. Style consistency across the broader S7 sheet

On the broad style sheet, Qwen said:

- linework looks **clean and digital**
- coloring looks **vibrant and uniform**
- likely method is **digital cutout / rigged** animation
- however, the still-image-only style sheet **did not support a limited-animation reading** in that one prompt

Interpretation:

- the model still strongly supports the **digital cutout / rigged** reading
- but a still-only contact sheet is a weak tool for judging motion economy
- the same failure already appeared in earlier passes: when shown still frames
  alone, the model sometimes over-reads richness / complexity and under-detects
  limited-animation discipline

### 2. Hold-motion sheet

On the hold-motion contact sheet, Qwen explicitly described the rows as:

- **static backgrounds**
- **repeated character poses**
- only **small adjustments** in expression or posture
- overall **held / restrained motion**

Interpretation:

- this directly supports the earlier claim that many S7 beats rely on held or
  near-held frames rather than dense full redraw
- this is the strongest broader-corpus support for the limited-animation side
  of the earlier conclusion

### 3. Action-motion sheet

On the action-motion contact sheet, Qwen described the rows as:

- still **controlled and limited**
- changes between frames are **sparse**
- motion reads more like **coordinated pose interpolation** than dense,
  continuously redrawn full animation

Interpretation:

- even action-leaning references do **not** flip the show into a high-fluidity
  full-animation regime
- the broader corpus still supports the conclusion that S7 action is stylized,
  controlled, and relatively economical in its frame-to-frame changes

### 4. Continuity vs benchmark consistency

On the continuity-vs-benchmark comparison, Qwen said both sheets appear to
belong to the same production style and support:

- clean vector-like linework
- flat color fills
- digital cutout / rigged animation
- limited-animation discipline

Interpretation:

- the short-form benchmark references are not style outliers
- the long-form episode material and the short-form promo material appear to be
  part of the same production system
- this matters because it means the short clips are valid for benchmarking,
  not just marketing outliers

## Final assessment

### What the expanded corpus clearly reinforces

The broader official S7 corpus strongly reinforces these parts of the earlier
conclusion:

1. **The show is digitally produced, not traditionally hand-drawn full animation.**
   The linework, coloring, and character construction remain consistent with a
   digital cutout / rigged pipeline.

2. **The visual style is highly production-consistent across trailers, clips,
   compilations, promos, and full episodes.**
   The broader corpus does not reveal a second visual mode or an alternate
   production style.

3. **Motion remains restrained even in action-leaning references.**
   The action sheet still reads as controlled, sparse, and key-pose driven
   rather than densely redrawn.

4. **Held / restrained beats are real and recurrent.**
   The hold sheet shows repeated evidence of static backgrounds, small facial /
   posture changes, and limited frame-to-frame variation.

5. **The AI-video problem remains temporal consistency.**
   Nothing in the broader corpus suggests that realistic physics simulation is
   the main challenge. The main challenge is still stable inter-frame behavior
   inside a restrained 2D cutout-style system.

### What remains methodologically weak

One still-sheet prompt did **not** support a limited-animation reading.
That disagreement should remain in the record.

Why it does not overturn the conclusion:

- the same model still called the material **digital cutout / rigged**
- the motion-specific sheets are more relevant than the still-only sheet for
  judging motion discipline
- the motion-specific sheets aligned with the earlier trailer analysis and the
  Gemma 3 second-opinion pass

## Conclusion

The earlier conclusion **still holds** after expanding from a narrow trailer
sample to a much broader official S7 corpus.

Best current reading:

- **animation type**: digital 2D cutout / rigged animation (consistent with
  Toon Boom Harmony workflow)
- **visual style**: clean vector-like linework, flat digital color fill,
  reusable composited background / character layering
- **motion style**: limited / restrained TV-animation discipline, including
  many held or near-held beats and action that is still controlled rather than
  densely redrawn
- **GenAI implication**: the key generation problem is preserving temporal
  consistency and stylistic restraint, not learning complex realistic physics

## Recommended next step

Use the downloaded full-episode continuity set to extract a **dialogue / hold
benchmark pack** and an **action / gadget benchmark pack**. Those will be even
stronger training / validation references than the trailer-only pack because
we now have much broader scene continuity available.