{
  "schema": "humaneness-cfg-guidance-v1",
  "scope": "Provisional matched-wording transfer evidence, not a universal optimum",
  "guidance_one": "Fully conditional prompted generation, not emotion off",
  "comparisons": [
    {
      "model": "original15",
      "target": "anger",
      "cfg": 1.5,
      "control_score": 0.42330045007734424,
      "guided_score": 0.5628583312751166,
      "control_id": "original15/anger/canonical-v1-validation/heldout-g1",
      "guided_id": "original15/anger/canonical-v1-validation/heldout-selected",
      "evidence": "100 takes per arm; fixed wording; selected CFG versus CFG 1, not a complete scalar sweep"
    },
    {
      "model": "original15",
      "target": "whisper",
      "cfg": 2.0,
      "control_score": 0.31043992872524306,
      "guided_score": 0.5072368276889081,
      "control_id": "original15/whisper/canonical-v1-validation/heldout-g1",
      "guided_id": "original15/whisper/canonical-v1-validation/heldout-selected",
      "evidence": "100 takes per arm; fixed wording; selected CFG versus CFG 1, not a complete scalar sweep"
    },
    {
      "model": "rank4",
      "target": "anger",
      "cfg": 2.5,
      "control_score": 0.33214522007152214,
      "guided_score": 0.5406499961018478,
      "control_id": "rank4/anger/canonical-v1-validation/heldout-g1",
      "guided_id": "rank4/anger/canonical-v1-validation/heldout-selected",
      "evidence": "100 takes per arm; fixed wording; selected CFG versus CFG 1, not a complete scalar sweep"
    },
    {
      "model": "rank4",
      "target": "whisper",
      "cfg": 2,
      "control_score": 0.31962629945650806,
      "guided_score": 0.5432952782046618,
      "control_id": "rank4/whisper/canonical-v1-validation/heldout-g1",
      "guided_id": "rank4/whisper/canonical-v1-validation/heldout-selected",
      "evidence": "100 takes per arm; fixed wording; selected CFG versus CFG 1, not a complete scalar sweep"
    }
  ],
  "new_task_grid": [
    1,
    2,
    2.5,
    3
  ],
  "stress_only": [
    4,
    5
  ],
  "high_cfg_confound": "Earlier 4–5 probes also changed prompt wording",
  "future_fixed_wording_grid": [
    1,
    1.25,
    1.5,
    2,
    2.5,
    3
  ]
}