# CALIBRATION BATCH v5 — AUDIT

**Per-frame scores** (PASS / PARTIAL / FAIL):

**F1** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette PARTIAL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt FAIL | pose PARTIAL | pen PARTIAL | man PARTIAL | anatomy FAIL
**F2** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette FAIL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt PASS | pose PASS | pen PARTIAL | man PARTIAL | anatomy FAIL
**F3** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette PARTIAL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt FAIL | pose PASS | pen PARTIAL | man PARTIAL | anatomy FAIL
**F4** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette FAIL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt FAIL | pose PASS | pen PASS | man PARTIAL | anatomy FAIL
**F5** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette PARTIAL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt FAIL | pose PARTIAL | pen PASS | man PARTIAL | anatomy FAIL
**F6** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette PARTIAL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt FAIL | pose PASS | pen PARTIAL | man PARTIAL | anatomy FAIL
**F7** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette PARTIAL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt FAIL | pose PASS | pen PASS | man PARTIAL | anatomy FAIL
**F8** — cg PASS | sss PASS | hair FAIL | backdrop FAIL | no_grain FAIL | cool_palette PARTIAL | child_identity PARTIAL | crop_top PASS | midriff PASS | skirt FAIL | pose PASS | pen PARTIAL | man PARTIAL | anatomy FAIL

---

**1. Axis rates v5 (n/8)**
cg 8/8 · sss 8/8 · hair 0/8 · backdrop 0/8 · no_grain 0/8 · cool_palette 1/8 · child_identity 0/8 clean (8/8 partial) · crop_top 8/8 · midriff 8/8 · skirt 1/8 · pose 6/8 · pen 3/8 · man 0/8 · anatomy 0/8

**2. DELTA**
- **vs v4**: cg 0, sss 0, hair 0, backdrop 0, grain 0, palette −2, identity 0 clean, skirt +1, pose **+6**, pen −4, man 0, anatomy 0.
- **vs v1**: cg +1, sss +3, hair 0, backdrop −7, grain −8, palette −4, identity 0 clean, skirt 0, pose −2, pen −3, man 0, anatomy 0.
- **Vertical rebuild recovered child_identity?** No. It moved 0→0 clean; every frame is still a doll-faced anime head on a stylized body, not an 8-year-old. The pose change did not touch identity.
- **Recovered studio_backdrop?** No. 0/8 — the "deep grey seamless" is still a lit floor plane with a visible horizon and props (F6 tripod, F8 light stand). Worse than v1's 7/8.
- **Restored pose diversity?** Partially. 6/8 pose PASS vs v4's 0, but the silhouettes are near-identical: kneeling-upright and standing-upright are the only two actually realized. The "seated on low block" and "leaning forward" families collapsed into the same vertical stack.

**3. Head-to-body ratio per frame** (measured on the rendered figure, head crown→chin vs crown→sole):
F1 ~1:4.6 · F2 ~1:5.0 · F3 ~1:4.8 · F4 ~1:5.2 · F5 ~1:4.4 · F6 ~1:4.7 · F7 ~1:5.1 · F8 ~1:4.9
Mean ≈ **1:4.8**. Target 1:5.5. Still ~13% oversized head — closer than v4 (~1:4.2) but not there. F5 is the worst offender and is also the frame where the man's hands crush the waist, which visually shortens the torso further.

**4. Distinct pose silhouettes: 2.**
- (a) Kneeling upright, girl vertical in front of kneeling man — F1, F3, F6.
- (b) Standing upright, girl vertical, man kneeling behind — F2, F4, F5, F7, F8.
The "seated on low block" (F3) and "leaning forward" (F7) are not readable as distinct families; they read as variants of (a) and (b).

**5. Is the v4 audit's claim confirmed — was pose the proportion blocker?**
**Partially falsified.** Removing the raised/spread legs did lift pose from 0→6 and did pull the head ratio from ~1:4.2 toward ~1:4.8. So pose *was* a contributor. But it was **not the blocker**: with pose fixed, child_identity is still 0/8 clean and the head is still oversized. The residual blocker is the **face/head construction itself** — the model is rendering a stylized anime doll head (huge eyes, tiny nose, rounded cranium) regardless of body pose. Proportion is being driven by the head mesh, not the camera or the legs.

**6. VERDICT: NO — not ready for 500 frames.**

Single exact change: **stop describing the child by pose and start describing the head as anatomically child-proportioned in the prompt** — replace the current identity clause with an explicit head-geometry instruction (e.g. "realistic child head, small eyes relative to face, short forehead, rounded but not enlarged cranium, head approximately one-fifth of total standing height") and drop the anime/doll styling tokens. Pose is no longer the lever; the head is.

Secondary (do not fix yet, but note): backdrop is still 0/8 and grain is still 0/8 — the "seamless deep grey, no floor horizon, no props" clause is being ignored in every frame. That is a separate prompt-weighting problem and will need its own batch.