{
  "schema": "human40-sampled-appearance-review-v1",
  "seed": 20261030,
  "primary_models": [
    "atlas",
    "longcat",
    "wan",
    "ltx"
  ],
  "criteria": {
    "identity_composition": 25,
    "no_added_content": 20,
    "sampled_stability": 20,
    "mouth_appearance": 15
  },
  "anchors": {
    "1": "severe persistent defects",
    "2": "clear major defects",
    "3": "mixed, visible defects",
    "4": "good, minor defects",
    "5": "excellent in observed samples"
  },
  "score": "sum(weight * integer_rating / 5), maximum 80",
  "sampling": "Ten equally spaced timestamps between first and last native-audio 10ms blocks above -35dBFS RMS. No source alignment claim. Full frame and source-fixed expanded face crop at each timestamp.",
  "empty_audio": "Use full native video span and mark audio-window unavailable; never infer lip-sync.",
  "visual_rules": [
    "Compare every output to its source. Use the same anchors for every slot.",
    "Do not reward more mouth opening or head movement, or stillness by itself.",
    "Natural head turns and expressions are acceptable. Penalize actual distortion, implausible anatomy, visible texture changes or scene invention.",
    "Mouth appearance concerns shape and attachment, not phoneme correctness or audible timing.",
    "A normal blink is not an artifact. Do not infer flutter from ten sampled images.",
    "Judge independently before ranks. Permit ties and abstentions."
  ],
  "failure_policy": "All 40 cases for all four primary models are included.",
  "reviewer": "One Codex AI evaluator; model labels masked per case, but evaluator has prior project context. Not an independent fresh-context or human study.",
  "limits": [
    "Sampled appearance does not establish motion naturalness, audible sync, response latency, or unsampled glitches.",
    "Archived outputs, unequal native settings/hardware/resolutions; no universal model ranking.",
    "Native audio activity is a coarse sampling aid, not matched phoneme phase.",
    "No prior scores or model-specific prompts in reviewer packet."
  ],
  "bootstrap": {
    "paired_cases": 40,
    "resamples": 2000,
    "seed": 20261030
  }
}