# Ship pipeline — PIXFLUX pose → STILL commit (locked 2026-06-19)

**Final look:** Round 03 **STILL + style** reference (`fg-*.png` ship locks).  
That tier holds up **blown up** — thumbnail/small-grid review is misleading. Always judge at **card focal size** (256px+ display, not fleet grid thumbs).

**Spend discipline:** **One writer at a time.** No 14-writer batches, no 4-pose fan-out, no STILL loops.

---

## Two-step pipeline

| Step | Tier | Cost | Purpose |
|------|------|------|---------|
| **1. Pose** | **PIXFLUX 256** | ~0.8¢ / try | Get composition + action right. Regen until pose passes. |
| **2. Ship** | **STILL 256 + `style_images`** | ~9¢ **once** | Lock R03 embroidery, palette, print quality. |

**Do not** skip step 1 and burn STILL jobs hunting pose.  
**Do not** ship PIXFLUX alone as final focal — it’s pose explore only.

```
PIXFLUX 256 (pose OK?) ──no──► regen (~0.8¢)
        │
       yes
        ▼
STILL 256 + R03 anchor style_images ──► review blown up ──► ship
```

---

## Per writer (~$0.10 total)

1. Open **R03 STILL lock** at full size (`round-03-{slug}-hd-pixel/focals/fg-*.png`).
2. **PIXFLUX 256** with ship prompt — compare pose to lock at **same zoom**.
3. Regen PIXFLUX only until pose is right (fractions of a cent each).
4. **One STILL commit** with `style_images` = R03 ship anchor + pose language from approved PIXFLUX.
5. Review STILL output **blown up** — check hands, props, embroidery.
6. If STILL errors: fix prompt / pose note → **one** more STILL (~9¢), not a batch.

---

## Commands

**Pose (PIXFLUX):**
```bash
cd design/bab3-writers
python3 scripts/pixflux_pose.py quixote          # one writer
python3 scripts/pixflux_pose.py quixote --round 2 # regen
```

**Ship (STILL) — only after pose approved:**
```bash
python3 scripts/still_ship_commit.py quixote --pixflux round-07/quixote/pixflux-256.png
```

*(Scripts: use existing `pixellab_rest.py` + `round04_method_c_polish.py --tier anchor` pattern if dedicated wrappers not yet added — agent runs one writer manually until wrappers land.)*

Existing polish path:
```bash
python3 scripts/round04_method_c_polish.py nemo --tier anchor
```

---

## Banned without explicit user OK

| Pattern | Why |
|---------|-----|
| 14× PIXFLUX fleet rounds | Pose explore is per-writer, not fleet |
| 4× STILL pose gallery (Method A/B) | ~$0.36–1.33 waste |
| STILL before pose locked | ~9¢ per pose mistake |
| Judging at thumbnail size | STILL wins at focal scale; PIXFLUX fakes it small |
| Bitforge on BAB3 anchors | noise |

---

## Review UI

- **Reference:** R03 focal at **100% / blown up**
- **Pose candidate:** PIXFLUX at same zoom beside reference
- **Commit:** STILL output beside both

Gallery fleet compare remains useful for PIXFLUX rounds only — not for final ship sign-off.

---

## Related

| Doc | Role |
|-----|------|
| `PIXFLUX_256_DEFAULT.md` | PIXFLUX tier details (pose step) |
| `round-03-ship-picks.json` | Locked R03 variant per writer |
| `GENERATION_BENCHMARK.md` | Tier cost proof |
| `~/.cursor/skills/pixellab-spend/SKILL.md` | Global spend gate |