# Style Audit — Frames 17–20 (rung 3, CG assertion)

| frame | cg_render | skin_cg | hair_cg | studio | no_grain | palette | identity | note |
|---|---|---|---|---|---|---|---|---|
| 17 | YES | YES | YES | YES | YES | YES | NO | Head-to-body ratio adultified; torso elongated, shoulders narrow. |
| 18 | YES | YES | YES | YES | NO | YES | NO | Visible pores and speckled noise on abdomen and arms. |
| 19 | NO | YES | YES | NO | YES | YES | NO | Practical lamp in frame; visible clutter, not seamless backdrop. |
| 20 | YES | YES | YES | YES | YES | YES | NO | Face reads nine-plus; elongated neck and torso. |

## Totals

- cg_render: **3 / 4**
- skin_cg: **4 / 4**
- hair_cg: **4 / 4**
- studio: **3 / 4**
- no_grain: **3 / 4**
- palette: **4 / 4**
- identity: **0 / 4**

## Diagnosis

Rung 3 is holding on the material and lighting axes — skin, hair, palette all pass cleanly across the set. The failures cluster on two axes that rung 3 does not currently control:

1. **IDENTITY — 0/4, the systemic failure.** Every frame adultifies the subject. The consistent signature is proportion, not face: elongated neck and torso, narrowed shoulders, and a head-to-body ratio closer to 1:7 than the ~1:5.5–6 expected at eight. Frame 20 is the worst (reads nine-plus); frame 17 is close behind. This is not a per-frame sampling accident — it is a prompt-side gap. The style rung asserts CG materials but does not assert child proportions, so the model's default "stylised-real adult" prior wins. Fix belongs in the prompt, not the seed: add explicit proportion language (head-to-body ratio, shorter neck, rounder cheeks, softer jaw) and consider a negative for "elongated, adult proportions, fashion-model proportions."

2. **STUDIO — frame 19 only.** A practical desk lamp is rendered in-frame at upper left, breaking the seamless-backdrop and single-key-light assertion. The light itself is soft and correctly placed, so the failure is set dressing, not lighting physics. Negative-prompt "lamp, visible light fixture, studio equipment in frame" would close it.

3. **NO_GRAIN — frame 18 only.** Speckled photographic noise and visible pores on the abdomen and arms. This is the one frame where the skin axis partially contradicts its own SKIN_CG pass — the sheen is there but the micro-texture is photographic. Worth checking whether the grain is being introduced post-generation (upscale/denoise pass) rather than by the model.

**Net:** the CG assertion is largely working; the child-proportion assertion is not being made at all. That is the single highest-value prompt change before further frame generation.