{
  "version": "fixed-prompt-targets-v1",
  "task_sha256": "6292db2ef9ad7a4236c7b9f529c8350a1e87d8e2e2c992674ab84734a59ee05c",
  "targets": {
    "anger_restraint": [
      [
        "emo_Anger",
        2.5
      ],
      [
        "emo_Impatience_and_Irritability",
        1.5
      ]
    ],
    "fear_courage": [
      [
        "emo_Fear",
        2.5
      ],
      [
        "emo_Hope_Enthusiasm_Optimism",
        1.5
      ]
    ],
    "grief_gratitude": [
      [
        "emo_Sadness",
        2.0
      ],
      [
        "emo_Thankfulness_Gratitude",
        2.0
      ]
    ],
    "joy_disbelief": [
      [
        "emo_Elation",
        2.5
      ],
      [
        "emo_Thankfulness_Gratitude",
        2.0
      ]
    ],
    "amusement_embarrassment": [
      [
        "emo_Amusement",
        2.5
      ],
      [
        "emo_Embarrassment",
        1.5
      ]
    ],
    "betrayal_resolve": [
      [
        "emo_Bitterness",
        2.0
      ],
      [
        "emo_Anger",
        1.5
      ]
    ],
    "awe_tenderness": [
      [
        "emo_Awe",
        2.5
      ],
      [
        "emo_Affection",
        2.0
      ]
    ]
  },
  "target_scale": "EARS raw emotion intensity, nominal 0–4",
  "target_origin": "Human-readable scene prompt and emotional arc; no source target performance exists",
  "scope": "Identical fixed two emotion heads and targets for every model, language and reference arm",
  "limitations": "Post-hoc diagnostic of existing scenes. Frozen before new baseline scoring and continuation outputs. Approximate scene-level intended intensity, not validated ground truth. The scenes have emotional arcs; full-clip averages cannot verify sentence-level realization. These are NOT the source-audio-derived top-two emotion targets used in RL training.",
  "selection": "Never choose strongest heads from a model prediction or the speaker-reference audio",
  "mse": "mean((raw_prediction - raw_target)**2); normalized MSE = raw MSE / 16; no clipping"
}