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 **⚠️ SUPERSEDED** — This document is a historical receipt.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

# Gemma 4 26B pack confirmation

## Purpose

Confirm or adjust the current S7 benchmark-pack conclusions using `gemma4:26b`.

Raw output:

- `materials/benchmark/youtube-s7-validation/packs-eval-gemma4-26b-chat.json`

## Result

Gemma 4 26B **confirms the current results**.

It does not change the current priority order:

1. character / costume consistency
2. dialogue / hold temporal consistency
3. action / gadget control

## What Gemma 4 reinforces

### Dialogue / hold

Gemma 4 emphasized:

- identity stability
- facial micro-expression fidelity
- background / gadget stability
- costume consistency during restrained movement
- texture boiling / shimmering
- background "breathing"

This strongly supports the current conclusion that held and low-motion scenes
are one of the highest-risk failure zones for a generic model.

### Action / gadget

Gemma 4 emphasized:

- gadget geometry preservation
- edge preservation during high-energy motion
- maintaining the cutout / layered look under pressure
- avoiding gadget melting, color bleeding, and motion smearing

This confirms that the action problem is still not "make it more realistic";
it is "keep it structurally stable and readable while moving."

### Character / costume consistency

Gemma 4 emphasized:

- identity drift
- costume morphing
- silhouette degradation
- texture bleeding from the environment
- the need for character anchors to stay fixed across contexts

This confirms the current memo's view that character and wardrobe locking are
still the first gate.

## What Gemma 4 adds

Compared with the earlier Qwen and Gemma 3 runs, Gemma 4 sharpens the failure
language in a useful way:

- **texture boiling / shimmering** on static lines
- **background breathing** in low-motion scenes
- **gadget melting** into hands or nearby shapes
- **color bleeding** across costume boundaries
- **silhouette degradation** under motion or lighting change

These do not overturn the conclusions, but they are good additions for future
scoring rubrics.

## Final assessment

Gemma 4 26B confirms the current operational reading:

- the project should optimize for **identity stability**,
  **held-frame temporal consistency**, and **controlled gadget/action
  readability**
- the key target remains **restrained, structurally stable cutout-style
  animation**, not generic AI fluid realism