{
  "_comment": "DIBS Beauty '5-Step Glassy Matte Routine' — THE WORKED EXAMPLE. Copy to config.json and edit. A ~32s 9:16 (1080x1920, 30fps) MULTI-SCENE GRWM beauty demo where a locked-identity creator applies 5 DIBS products step-by-step at a sage vanity, a SEPARATE locked ElevenLabs Karri VO narrates the routine, and every scene cut is snapped to the VO's 'step N' / product-name word-starts (Whisper), with 5 Playwright product overlay cards (real PDP taglines) on the product-name beats, a warm acoustic music bed ducked under the VO, clean-white 3-words/cue captions, closing on a gpt-image-2 flat-lay end card. Pipeline: (1) lock creator anchor + sage-vanity world + hook keyframe + Karri VO (atempo 1.15x -> ~27.5s); (2) 7 Seedance 2.0 scene clips (720p, --no-audio) + the flat-lay still; (3) Groq whisper-large-v3 word-level on the VO; (4) re-cut + hard-concat to the 12 VO cuts (build_v4.py, -crf 20 re-encode); (5) 5 white-bg gpt-image-2 product cutouts (PDP-verified taglines) + Playwright cards (render_cards.py); (6) composite cards on the product-name beats (composite_cards.py, -loop 1 -t) + mix VO/music + burn captions + flat-lay end card (build_master.py). Master 1080x1920 h264+aac, ~32s. See scripts/PIPELINE.md for the source scripts.",
  "brand_name": "DIBS Beauty",
  "campaign": "5-Step Glassy Matte Routine",
  "title": "5 products, one finish",
  "width": 1080,
  "height": 1920,
  "fps": 30,
  "duration_sec": 32.0,
  "aspect_ratio": "9:16",
  "palette": {
    "card_bg": "#F8EEDE",
    "accent": "#E879A6"
  },
  "assets": {
    "creator_anchor": "assets/characters/grwm-creator.png",
    "_note_creator": "26yo brunette w/ caramel highlights, black satin headband, white ribbed tank, in a sage-green bathroom. Identity + wardrobe + setting held across every scene clip.",
    "vanity_world": "assets/worlds/sage-vanity.png",
    "hook_keyframe": "assets/keyframes/five-pink-barrels.png",
    "_note_hook": "Creator holding 5 pink DIBS barrels fanned near her face. Doubles as the payoff-smile beat (s07a)."
  },
  "vo": {
    "_comment": "SEPARATE locked ElevenLabs Karri VO — NOT a native Seedance take. Whispered FIRST; the word-starts set the timeline. atempo 1.15x makes it end at ~27.5s (NOT 45s). Script written clean (no eleven_v3 audio tags).",
    "source": "user_supplied",
    "provider": "elevenlabs",
    "model": "eleven_v3",
    "voice_id": "3mip1Cv2rZYeio9MCzfr",
    "voice_name": "Karri",
    "stability": 0.40,
    "similarity_boost": 0.75,
    "style": 0.25,
    "atempo": 1.15,
    "script": "okay, my five-step glassy matte routine. step one, Love Beam Perfector. it's basically skincare you put on first. step two, Desert Island Duo. blush and bronzer in one stick. step three, Glow Tour for contour. step four, Love Beam Powder. it brightens and blurs. step five, Setting Mist. that's the lock. five products, one finish. about three minutes.",
    "actual_duration_s": 27.5,
    "outputs": {
      "audio": "working/voiceovers/locked/karri-full-atempo115.mp3",
      "word_timings": "working/voiceovers/locked/karri-full-atempo115.words.json"
    },
    "_word_starts": {
      "_comment": "Whisper-verified word-starts for this exact atempo'd mp3. Scene cuts snap to the 'step N' starts; cards snap to the product-NAME starts.",
      "step_one": 3.32, "Love Beam Perfector": 4.34,
      "step_two": 8.64, "Desert Island Duo": 9.56,
      "step_three": 13.82, "Glow Tour": 14.86,
      "step_four": 17.00, "Love Beam Powder": 17.94,
      "step_five": 21.48, "Setting Mist": 22.42,
      "five_products_one_finish": 24.62, "about_three_minutes": 26.50
    }
  },
  "scene_clips": {
    "_comment": "7 Seedance 2.0 clips (create-video-seedance-fal, Fal direct, 720p, 9:16, --no-audio) off the locked anchor + world. Softer verbs — s05 powder used 'swirls'/'dusts' after a 422 on 'press'/'brush her face'. ~$12.60 total.",
    "provider": "fal",
    "model": "seedance_2_0",
    "resolution": "720p",
    "aspect_ratio": "9:16",
    "audio": false,
    "clips": [
      { "id": "s01-hook",      "role": "HOOK",   "duration_sec": 4, "ref": ["hook_keyframe"],               "beat": "holds five glossy pink DIBS cream-sticks fanned near her face, eye contact with selfie camera, playful smile" },
      { "id": "s02-perfector", "role": "STEP_1", "duration_sec": 7, "ref": ["creator_anchor", "vanity_world"], "beat": "dabs a few pearls of pink LoveBeam Perfector onto cheekbones + forehead with fingertips, taps them in" },
      { "id": "s03-did",       "role": "STEP_2", "duration_sec": 8, "ref": ["creator_anchor"],              "beat": "swipes the Desert Island Duo cream-stick on cheeks/nose/forehead, presses in with a dense face brush" },
      { "id": "s04-glowtour",  "role": "STEP_3", "duration_sec": 7, "ref": ["creator_anchor"],              "beat": "stripes the Glow Tour contour stick along cheekbone hollow + jawline, blends down with fingertips" },
      { "id": "s05-powder",    "role": "STEP_4", "duration_sec": 6, "ref": ["creator_anchor"], "tier": "fast", "beat": "swirls a soft brush in the pink compact, taps it, dusts across her T-zone (softer verbs — 'press'/'brush her face' 422'd)" },
      { "id": "s06-mist",      "role": "STEP_5", "duration_sec": 6, "ref": ["creator_anchor"],              "beat": "holds the LoveBeam Setting Mist ~8in from her face and sprays a fine even mist; skin gains a glassy glow" },
      { "id": "s07a-payoff",   "role": "PAYOFF", "duration_sec": 4, "ref": ["hook_keyframe"],               "beat": "smiles fully into the selfie camera, fresh glassy-matte makeup on, warm confident expression" }
    ]
  },
  "cuts": {
    "_comment": "12 cuts derived from the VO word-starts (build_v4.py). Scene cuts land on 'step N' starts; s02 + s03 + s05 sub-cut for pacing. Hard-concat, re-encode -c:v libx264 -crf 20. cuts/10s = 12/32*10 = 3.75.",
    "hard_concat": true,
    "reencode": "-c:v libx264 -crf 20",
    "count": 12,
    "cuts_per_10s": 3.75,
    "windows": [
      { "t_start": 0.00,  "t_end": 3.00,  "clip": "s01-hook",      "src_in": 0.0, "dur": 3.00, "why": "hook ends as 'step one' begins" },
      { "t_start": 3.00,  "t_end": 6.00,  "clip": "s02-perfector", "src_in": 0.0, "dur": 3.00, "why": "perfector first half" },
      { "t_start": 6.00,  "t_end": 9.56,  "clip": "s02-perfector", "src_in": 3.0, "dur": 3.56, "why": "perfector second half — ends as 'Desert Island Duo' begins" },
      { "t_start": 9.56,  "t_end": 12.00, "clip": "s03-did",       "src_in": 0.0, "dur": 2.44, "why": "DID swipe begin" },
      { "t_start": 12.00, "t_end": 13.82, "clip": "s03-did",       "src_in": 2.5, "dur": 1.82, "why": "DID continuation — ends as 'step three' begins" },
      { "t_start": 13.82, "t_end": 17.00, "clip": "s04-glowtour",  "src_in": 0.0, "dur": 3.18, "why": "glowtour — ends as 'step four' begins" },
      { "t_start": 17.00, "t_end": 19.22, "clip": "s05-powder",    "src_in": 0.0, "dur": 2.22, "why": "powder swirl" },
      { "t_start": 19.22, "t_end": 21.48, "clip": "s05-powder",    "src_in": 2.3, "dur": 2.26, "why": "powder application — ends as 'step five' begins" },
      { "t_start": 21.48, "t_end": 24.52, "clip": "s06-mist",      "src_in": 0.0, "dur": 3.04, "why": "mist — ends as 'five products' payoff begins" },
      { "t_start": 24.52, "t_end": 27.80, "clip": "s07a-payoff",   "src_in": 0.0, "dur": 3.28, "why": "payoff smile during 'five products, one finish'" },
      { "t_start": 27.80, "t_end": 32.00, "clip": "FLAT_LAY",      "src_in": 0.0, "dur": 4.20, "why": "end-card ken-burns on the flat-lay still" }
    ]
  },
  "products": {
    "_comment": "5 white-bg standalone gpt-image-2 cutouts (gen-product-cutouts.sh). Taglines VERIFIED against dibsbeauty.com PDPs — the gpt-image-2 flat-lay renders WRONG sublines ('SCULPT + GLOW' on DID, 'BLUSH + HIGHLIGHT' on GlowTour). LoveBeam Perfector is a slim liquid bottle, NOT a wide cream stick (even though the Seedance s02 clip shows a wide barrel — the card shows the REAL product).",
    "cutouts": [
      { "id": "perfector", "name": "LoveBeam Perfector",   "tagline": "GLOW SERUM",       "cutout": "working/products/perfector.png", "vo_name_start_s": 4.34 },
      { "id": "did",       "name": "Desert Island Duo",    "tagline": "BLUSH + BRONZER",  "cutout": "working/products/did.png",       "vo_name_start_s": 9.56 },
      { "id": "glowtour",  "name": "Glow Tour Duo",        "tagline": "CONTOUR + BLUSH",  "cutout": "working/products/glowtour.png",  "vo_name_start_s": 14.86 },
      { "id": "powder",    "name": "LoveBeam Powder",      "tagline": "BRIGHTEN + BLUR",  "cutout": "working/products/powder.png",    "vo_name_start_s": 17.94 },
      { "id": "mist",      "name": "Setting Mist",         "tagline": "LOCK IT IN",       "cutout": "working/products/mist.png",      "vo_name_start_s": 22.42 }
    ],
    "_note_taglines": "Taglines above are the card lines used in the worked example; ALWAYS re-verify against the live PDP before shipping a new brand — do not carry these forward."
  },
  "cards": {
    "_comment": "5 Playwright cards from working/product-card.html.tmpl at 2x scale (render_cards.py): warm cream card (#F8EEDE gradient) + pink accent bar (#E879A6) + cutout thumb + Georgia name + uppercase pink tagline. Composited (composite_cards.py) snapped to the product-NAME word-start (~1s after the scene cut), 1s fade-in. PNG overlays NEED -loop 1 -t <dur>.",
    "template": "working/product-card.html.tmpl",
    "playwright_scale": 2,
    "fade_in_s": 1.0,
    "loop_flag_required": true,
    "timing": [
      { "card": "perfector", "start": 4.34,  "end": 9.56,  "matches": "'Love Beam Perfector' -> 'Desert Island Duo'" },
      { "card": "did",       "start": 9.56,  "end": 13.82, "matches": "'Desert Island Duo' -> 'step three'" },
      { "card": "glowtour",  "start": 14.86, "end": 17.94, "matches": "'Glow Tour for contour' -> 'Love Beam Powder'" },
      { "card": "powder",    "start": 17.94, "end": 22.42, "matches": "'Love Beam Powder' -> 'Setting Mist'" },
      { "card": "mist",      "start": 22.42, "end": 27.80, "matches": "'Setting Mist' -> payoff" }
    ]
  },
  "music": {
    "_comment": "Warm acoustic instrumental bed (ElevenLabs Music, 45s — only ~32s consumed), ducked under the VO. No artist names in the prompt (ElevenLabs ToS filter).",
    "provider": "elevenlabs",
    "prompt": "Warm acoustic instrumental, soft nylon guitar + light percussion, gentle and aspirational, easygoing morning-routine energy, no vocals.",
    "source": "working/music/music.mp3",
    "duration_s": 45,
    "force_instrumental": true
  },
  "audio_mix": {
    "_comment": "The SEPARATE Karri VO is the lead; the music sits well under it. loudnorm I=-14.",
    "loudness_lufs": -14,
    "vo_lead": true,
    "music_duck": true
  },
  "captions": {
    "_comment": "Clean-white captions from the VO's Whisper words. OVERRIDE the clean-white preset: 3 words/cue, 3.0% font, 20% from the bottom, NO pill, NO shadow (default 5-words/4.5%/18% is too dense). Built in cue_builder.py within build_master.py.",
    "source": "working/voiceovers/locked/karri-full-atempo115.words.json",
    "style": "clean_white",
    "chunk_words": 3,
    "font_size_pct": 3.0,
    "margin_bottom_pct": 20,
    "no_pill": true,
    "no_shadow": true,
    "base_color": "#FFFFFF"
  },
  "end_card": {
    "_comment": "gpt-image-2 9:16 flat-lay still of the 5 products on cream linen, ken-burns held ~4.2s at the tail (27.80 -> 32.00). Do NOT trust the flat-lay's AI sublines for taglines — PDP-verify (see products[]).",
    "engine": "gpt-image-2",
    "flat_lay": "assets/keyframes/grwm-flat-lay.png",
    "hold_sec": 4.2,
    "effect": "ken-burns"
  },
  "post_production": {
    "vo": { "default": "on", "note": "the SEPARATE locked ElevenLabs VO (atempo 1.15x) narrates the routine and drives every cut" },
    "product_cards": { "default": "on", "note": "5 Playwright cards (real cutout + PDP tagline), each on its product-NAME word-start" },
    "music": { "default": "on", "note": "warm instrumental bed ducked under the VO, loudnorm I=-14" },
    "captions": { "default": "on", "note": "clean-white, 3 words/cue, 3.0% font, 20% margin, no pill/shadow" },
    "end_card": { "default": "on", "note": "gpt-image-2 flat-lay still of the 5 products, ken-burns held ~4s" }
  }
}

# Pipeline — glassy-matte GRWM (multi-scene beauty demo)

How `config.example.json` maps to the real production steps. This capability ships a
**config + this map**, not a bundled runner: the worked example (DIBS Beauty "5-Step Glassy
Matte Routine") was produced by the video-orchestrator's per-state steps plus three per-project
drivers (`build_v4.py`, `render_cards.py`, `composite_cards.py` + `build_master.py`). The four
paid stages (the SEPARATE VO, the ~7 Seedance scene clips, the ~5 gpt-image-2 product cutouts +
the flat-lay) are separate capabilities; this cap owns the FREE assembly at the end.

The steps run **in order** because each depends on the last: the SEPARATE VO sets the timeline
(via its Whisper word-starts), the word-starts drive the scene cuts, the scenes are re-cut +
concatenated, the product cutouts + PDP taglines drive the cards, the cards are composited on the
product-name beats, and the mix + captions + end card finish the master.

## Field → source-step map

| Config field | Phase | Source step / script | Paid? |
|---|---|---|---|
| `vo.script`, `vo.voice_id`, `vo.atempo` | 1 Lock | ElevenLabs `eleven_v3` TTS + ffmpeg `atempo 1.15` (or ingest an mp3) | **PAID** (or supplied) |
| `scene_clips.clips[]` | 2 Scenes | `create-video-fal` (Seedance 2.0, Fal direct, 720p, ~6s, `--no-audio`) | **PAID** (~$0.30/s @ 720p) |
| `end_card.flat_lay` | 2 Scenes | `create-image-gpt-image-fal` (gpt-image-2, 9:16) | **PAID** (~$0.07) |
| `vo.outputs.word_timings` | 3 Whisper | Groq `whisper-large-v3` word-level → `<vo>.words.json` | ~$0.04 |
| `cuts.windows[]` | 4 Re-cut | scene cuts derived from the "step N" word-starts, re-cut + hard-concat + `-crf 20` | `build_v4.py` (free) |
| `products.cutouts[]` | 5 Products | `create-image-gpt-image-fal` (gpt-image-2 white-bg standalone); taglines PDP-verified | **PAID** (~$0.35) |
| `cards.template`, `.playwright_scale` | 5 Cards | Playwright renders `product-card.html.tmpl` at 2× | `render_cards.py` (free) |
| `cards.timing[]` | 6 Composite | cards composited onto the master, snapped to product-name word-starts (`-loop 1 -t`) | `composite_cards.py` (free) |
| `music`, `audio_mix` | 6 Mix | ElevenLabs Music bed ducked under the VO, loudnorm I=-14 | `build_master.py` (**PAID** music; mix free) |
| `captions` | 6 Captions | clean-white captions from the Whisper words (3 words/cue override) burned | `build_master.py` (free) |
| `end_card.hold_sec`, `.effect` | 6 End card | flat-lay ken-burns hold appended | `build_master.py` (free) |

## 1. Lock the assets + VO  (config: `assets`, `vo`)  [PAID-or-supplied VO]

Lock the creator anchor (identity + wardrobe + setting), the vanity world plate, and the hook
keyframe (creator holding the ~5 products fanned; doubles as the payoff). Render the VO on
ElevenLabs `eleven_v3` (`stability 0.40`, `style 0.25`), then `ffmpeg atempo=1.15` to hit a
~27.5s pace, or ingest a supplied mp3. **The VO is a SEPARATE track — not a native take.** Write
the script clean (no audio tags).

## 2. Scene clips → Seedance 2.0  (config: `scene_clips`, `end_card.flat_lay`)  [PAID]

`create-video-fal` (Seedance 2.0, Fal direct) fires ~7 scene clips: a hook, one clip per product
step (its application beat), and a payoff smile — 720p / 9:16 / `--no-audio`, off the locked anchor
+ world refs. **Softer application verbs** ("dabs", "swirls", "dusts") — hard verbs ("press",
"brush her face") 422 on content policy; retry `--tier fast`. Also generate the gpt-image-2 flat-lay
end-card still. Verify each clip's duration; Seedance drifts on the back ~2s of longer clips, so
trim to the clean window in Phase 4.

## 3. Whisper the VO → Groq word timestamps  (config: `vo.outputs.word_timings`)  [~$0.04]

Groq `whisper-large-v3` word-level on the atempo'd VO → `<vo>.words.json`. **The VO's actual
duration + word-starts set the timeline** — a 1.15× VO ends ~27.5s, not the 45s the plan assumed
(the back half would play silent). Record the word-start of each "step N" and each product name.

## 4. Re-cut + concat → `build_v4.py`  (config: `cuts`)  [free]

Derive the cut points: **scene cuts land on the "step N" word-starts** (cut to the next product
when its step is announced, not when the previous application beat ends); a scene may sub-cut for
pacing. Re-cut each clip to its VO window and **hard-concat** (no dissolves) → a silent master. 12
cuts over 32s (cuts/10s = 3.75). **Re-encode `-c:v libx264 -crf 20`** — `-c copy` corrupts the
duration when zoompan/PNG clips are in the chain.

## 5. Product cutouts + cards → `gen-product-cutouts.sh` + `render_cards.py`  (config: `products`, `cards`)  [PAID cutouts]

`create-image-gpt-image-fal` (gpt-image-2, white-bg) → ~5 **standalone** product cutouts (NOT
on-body, NOT an AI flat-lay). **PDP-verify every tagline** — the gpt-image-2 flat-lay renders
hallucinated sublines ("SCULPT + GLOW"), which are WRONG. The cutout must match the REAL product,
not the Seedance scene's hallucinated barrel. Playwright renders `product-card.html.tmpl` at **2×
scale** → one card PNG per product (warm cream card, pink accent bar, cutout thumb, name, PDP
tagline).

## 6. Composite + mix + captions + end card → `composite_cards.py` + `build_master.py`  (config: `cards.timing`, `music`, `audio_mix`, `captions`, `end_card`)  [free assembly]

- **Cards** (`composite_cards.py`): composite each card onto the silent master, **snapped to the
  product-NAME word-start** (~1s after its scene cut), fading in over ~1s, held until the next
  product is named. **PNG overlay inputs NEED `-loop 1 -t <dur>`** — else the PNG emits one frame at
  t=0 and the fade/enable filters silently no-op (cards go invisible).
- **Mix** (`build_master.py`): mix the SEPARATE VO on top of the ducked ElevenLabs Music bed (VO is
  the lead), `loudnorm I=-14`.
- **Captions**: clean-white captions from the VO's Whisper words, **overridden to 3 words/cue,
  ~3.0% font, ~20% margin, no pill, no shadow** (the default 5-words/4.5%/18% preset reads too
  dense). If the host ffmpeg lacks libass, render the cues as timed PIL PNG overlays instead.
- **End card**: append the gpt-image-2 flat-lay still, ken-burns held ~4s → `finals/<brand>-glassy-matte-grwm-v4.mp4`
  (1080×1920, 30fps, h264+aac, ~32s).

Re-cuts (new cut windows, re-timed cards, caption chunking, a swapped end card) reuse the existing
clips / VO / cutouts / music and cost **$0** — only steps 1–2 and 5's cutouts spend.

# render-glassy-matte-grwm scripts — the FREE assembly

`render-glassy-matte-grwm` is the **deterministic, $0 assembly stage** of the multi-scene GRWM
beauty-demo format. The paid stages (the SEPARATE narration VO, the ~7 Seedance scene clips, the
~5 gpt-image-2 product cutouts + the flat-lay end-card still, the music bed) are separate
capabilities — `create-music-elevenlabs`, `create-video-fal`, `create-image-gpt-image-fal`. This
capability spends nothing — it takes the VO + `<vo>.words.json` + one clip per product step + the
~5 product cutouts + the music bed and stitches the finished master. Re-cuts (new cut windows,
re-timed cards, a swapped end card, caption chunking) reuse the existing VO / clips / cutouts and
cost **$0**.

`config.example.json` is the worked example (DIBS Beauty "5-Step Glassy Matte Routine", ~32s
1080×1920). `PIPELINE.md` maps every config block to its source step. This README documents the
FREE assembly pieces that `render-glassy-matte-grwm` owns.

## 1. Re-cut to the VO word-starts + hard-concat on the cut

The SEPARATE VO drives the timeline. Groq `whisper-large-v3` word-level gives the word-start of
each "step N" and each product name; **scene cuts land on the "step N" word-starts** (cut to the
next product when its step is announced), and a scene may sub-cut for pacing. Re-cut each clip to
its VO window and **hard-concat on the cut** — no dissolves. **Re-encode `-c:v libx264 -crf 20`**
— `-c copy` corrupts the duration when zoompan/PNG clips are in the chain. ~12 cuts over ~32s.

## 2. Product cards — Playwright render + composite on the product-name beats

Playwright renders the card template (`product-card.html.tmpl`) at **2× scale** → one PNG per
product — a warm cream card, a pink accent bar, the real white-bg cutout thumb, the product name,
and the PDP-verified tagline. The cutout must match the REAL product (not the Seedance scene's
hallucinated barrel), and the tagline is verified against the brand PDP (AI flat-lays hallucinate
sublines). Each card is composited onto the master **snapped to its product-NAME word-start** (~1s
after the scene cut), fading in over ~1s, held until the next product is named. **PNG overlay
inputs NEED `-loop 1 -t <dur>`** — without it the PNG emits one frame at t=0 and the fade/enable
filters silently no-op (the card is invisible).

## 3. Captions — clean-white, override the preset

Clean-white captions from the VO's Whisper words, **overridden to 3 words/cue, ~3.0% font, ~20%
margin, NO pill, NO shadow** (the default 5-words/4.5%/18% preset reads too dense for this format).
Burned last. If the host ffmpeg lacks libass (no `subtitles`/`ass` filter), render the cues as
timed PIL PNG overlays composited with ffmpeg `overlay=…:enable='between(t,st,en)'` at the same
placement.

## 4. FFmpeg mix + end card

Mix the SEPARATE VO on top of the ducked ElevenLabs Music bed (the VO is the lead — the music sits
well under it), `loudnorm I=-14`. Append the gpt-image-2 flat-lay still (ken-burns hold ~4s) — do
NOT trust its AI-rendered sublines for the card taglines. Output is a 1080×1920 30fps h264+aac
master (~32s). Deterministic, no paid calls, no keys.

# Smoke Test

Given the SEPARATE narration VO (mp3 + `.words.json`), ~7 Seedance scene clips (one per product
step + a hook + a payoff, one locked creator across all), ~5 white-bg product cutouts with
PDP-verified taglines, a music bed, and a flat-lay end-card still, `render-glassy-matte-grwm`
assembles the master — re-cut each clip to its VO word-start window, hard-concat on the cut,
Playwright-render + composite the ~5 product cards on the product-name beats, mix the VO over the
ducked music, burn clean-white captions, close on the flat-lay end card → 1080×1920 h264+aac
(~32s).

Pass when the assembly runs to a valid MP4 and:
- ~12 scene segments, each hard-cut on its "step N" word-start (no crossfades); the ONE locked
  creator + vanity hold across every scene;
- ~5 product cards, each animating in on its product-NAME word-start (~1s after the scene cut) with
  the RIGHT white-bg cutout + a PDP-verified tagline; every PNG overlay used `-loop 1 -t <dur>`;
- captions are clean-white 3-words/cue (no pill, no shadow); the flat-lay end card holds ~4s on the
  final window;
- the SEPARATE VO clearly leads the ducked music bed at loudnorm I=-14 — no VO baked into the
  clips, no music on top of the VO;
- **no paid call is made** — the VO, scene clips, product cutouts, and music come from the paid
  capabilities (create-music-elevenlabs / create-video-fal / create-image-gpt-image-fal); this
  assembly is $0 and a re-cut reuses the existing assets.