{
  "aggregate": [
    {
      "model": "sft13",
      "condition": "prompt_only",
      "samples": 35,
      "ce": {
        "n": 35,
        "mean": 6.274169363294329,
        "std": 0.17617152166441002
      },
      "genuineness": {
        "n": 35,
        "mean": 4.256933106694903,
        "std": 0.17384309897292455
      },
      "wer": {
        "n": 35,
        "mean": 0.04358086567662996,
        "std": 0.16170332195427864
      }
    },
    {
      "model": "sft13",
      "condition": "reference",
      "samples": 35,
      "ce": {
        "n": 35,
        "mean": 6.242923477717809,
        "std": 0.1959923598956535
      },
      "genuineness": {
        "n": 35,
        "mean": 4.143336769512722,
        "std": 0.25421066884355803
      },
      "wer": {
        "n": 35,
        "mean": 0.021181270556851938,
        "std": 0.05334843136181323
      }
    },
    {
      "model": "rl13",
      "condition": "prompt_only",
      "samples": 35,
      "ce": {
        "n": 35,
        "mean": 6.341001279013497,
        "std": 0.13653024074129377
      },
      "genuineness": {
        "n": 35,
        "mean": 4.322741838863918,
        "std": 0.18768877536678139
      },
      "wer": {
        "n": 35,
        "mean": 0.036893608894247906,
        "std": 0.15982573985043483
      }
    },
    {
      "model": "rl13",
      "condition": "reference",
      "samples": 35,
      "ce": {
        "n": 35,
        "mean": 6.3142798832484655,
        "std": 0.1658511720878949
      },
      "genuineness": {
        "n": 35,
        "mean": 4.275581911631993,
        "std": 0.24532359587634955
      },
      "wer": {
        "n": 35,
        "mean": 0.01797475586757579,
        "std": 0.040464944886961486
      }
    }
  ],
  "warning": "Learned proxy predictions, not human ratings. No certified speaker-similarity or emotion-target ground truth on these scripted cases.",
  "scope": "70 fixed cases per model; no input identifiers, transcripts or raw recordings redistributed"
}