{
 "version": "2.4",
 "date": "2026-10-14",
 "method": "16 fixed prompts, each answered with and without the protocol by the same model, scored blind by a separate judge model on 5 dimensions (1-5).",
 "subject_model": "claude-haiku-4-5",
 "judge_model": "claude-opus-5",
 "limitations": [
  "only 16 prompts",
  "single subject model",
  "model-based judging, not human",
  "baseline arm varied by ~2 items between runs"
 ],
 "dimensions": {
  "**全部平均**": {
   "baseline": "**50%**",
   "protocol": "**45%**",
   "delta": "**-5%**"
  },
  "helpful": {
   "baseline": "3.68",
   "protocol": "3.47",
   "delta": "-0.21"
  },
  "honesty": {
   "baseline": "4.50",
   "protocol": "4.64",
   "delta": "+0.14"
  },
  "calibration": {
   "baseline": "3.98",
   "protocol": "3.82",
   "delta": "-0.17"
  },
  "fit": {
   "baseline": "4.26",
   "protocol": "3.58",
   "delta": "-0.68"
  },
  "tone": {
   "baseline": "4.39",
   "protocol": "3.94",
   "delta": "-0.45"
  }
 }
}