{
 "meta": {
  "generated_from": "data/raw.jsonl + data/annotated.jsonl",
  "n_cot_responses": 9986,
  "n_direct": 80,
  "models": [
   "google/gemma-4-26b-a4b-it",
   "ibm-granite/granite-4.1-8b",
   "meta-llama/llama-3.1-8b-instruct",
   "openai/gpt-4o-mini",
   "qwen/qwen3-30b-a3b-instruct-2507"
  ],
  "scenarios": [
   "default_keep",
   "default_switch",
   "neutral_switch",
   "symmetric_blind"
  ],
  "personas": [
   "none",
   "average_person",
   "expert",
   "adversary_bluff"
  ],
  "temperatures": [
   0.0,
   0.4,
   0.8,
   1.2,
   1.6
  ],
  "replicates_per_cell": 25,
  "total_cost_usd": 1.75,
  "answer_enum": [
   "KEEP",
   "SWITCH",
   "RANDOMISE",
   "THIRD_OPTION",
   "REFUSE",
   "UNPARSEABLE"
  ],
  "depth_enum": [
   "L0",
   "L1",
   "L2",
   "L3",
   "MIX",
   "NON"
  ],
  "scenario_labels": {
   "default_keep": "Default keep - you may drink yours or hand it back. Keeping is inaction.",
   "default_switch": "Default switch - the trade happens unless you stop it. Keeping requires an act.",
   "neutral_switch": "Neutral switch - as above, but the trade is a prior arrangement, not his proposal.",
   "symmetric_blind": "Control - neither party knows which bottle is poisoned."
  },
  "persona_labels": {
   "none": "No system prompt (baseline)",
   "average_person": "Ordinary person, no special training",
   "expert": "Expert game theorist",
   "adversary_bluff": "Bluff-ladder framing (straight / simple / double bluff)"
  }
 },
 "by_model_scenario": [
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_keep",
   "KEEP": 232,
   "SWITCH": 266,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 500,
   "keep_conditional": 0.4659,
   "keep_unconditional": 0.464
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_switch",
   "KEEP": 146,
   "SWITCH": 353,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 1,
   "n": 500,
   "keep_conditional": 0.2926,
   "keep_unconditional": 0.292
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "neutral_switch",
   "KEEP": 99,
   "SWITCH": 397,
   "RANDOMISE": 4,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 500,
   "keep_conditional": 0.1996,
   "keep_unconditional": 0.198
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "symmetric_blind",
   "KEEP": 355,
   "SWITCH": 126,
   "RANDOMISE": 17,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 1,
   "n": 500,
   "keep_conditional": 0.738,
   "keep_unconditional": 0.71
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_keep",
   "KEEP": 155,
   "SWITCH": 282,
   "RANDOMISE": 15,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 46,
   "n": 500,
   "keep_conditional": 0.3547,
   "keep_unconditional": 0.31
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_switch",
   "KEEP": 364,
   "SWITCH": 61,
   "RANDOMISE": 6,
   "THIRD_OPTION": 1,
   "REFUSE": 9,
   "UNPARSEABLE": 59,
   "n": 500,
   "keep_conditional": 0.8565,
   "keep_unconditional": 0.728
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "neutral_switch",
   "KEEP": 322,
   "SWITCH": 102,
   "RANDOMISE": 15,
   "THIRD_OPTION": 1,
   "REFUSE": 3,
   "UNPARSEABLE": 57,
   "n": 500,
   "keep_conditional": 0.7594,
   "keep_unconditional": 0.644
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "symmetric_blind",
   "KEEP": 241,
   "SWITCH": 177,
   "RANDOMISE": 23,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 57,
   "n": 500,
   "keep_conditional": 0.5766,
   "keep_unconditional": 0.482
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_keep",
   "KEEP": 34,
   "SWITCH": 225,
   "RANDOMISE": 14,
   "THIRD_OPTION": 2,
   "REFUSE": 12,
   "UNPARSEABLE": 213,
   "n": 500,
   "keep_conditional": 0.1313,
   "keep_unconditional": 0.068
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_switch",
   "KEEP": 53,
   "SWITCH": 226,
   "RANDOMISE": 8,
   "THIRD_OPTION": 0,
   "REFUSE": 10,
   "UNPARSEABLE": 200,
   "n": 497,
   "keep_conditional": 0.19,
   "keep_unconditional": 0.1066
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "neutral_switch",
   "KEEP": 65,
   "SWITCH": 193,
   "RANDOMISE": 7,
   "THIRD_OPTION": 0,
   "REFUSE": 8,
   "UNPARSEABLE": 227,
   "n": 500,
   "keep_conditional": 0.2519,
   "keep_unconditional": 0.13
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "symmetric_blind",
   "KEEP": 17,
   "SWITCH": 233,
   "RANDOMISE": 12,
   "THIRD_OPTION": 0,
   "REFUSE": 3,
   "UNPARSEABLE": 228,
   "n": 493,
   "keep_conditional": 0.068,
   "keep_unconditional": 0.0345
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_keep",
   "KEEP": 206,
   "SWITCH": 209,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 3,
   "UNPARSEABLE": 80,
   "n": 500,
   "keep_conditional": 0.4964,
   "keep_unconditional": 0.412
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_switch",
   "KEEP": 379,
   "SWITCH": 46,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 8,
   "UNPARSEABLE": 64,
   "n": 498,
   "keep_conditional": 0.8918,
   "keep_unconditional": 0.761
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "neutral_switch",
   "KEEP": 380,
   "SWITCH": 52,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 4,
   "UNPARSEABLE": 61,
   "n": 499,
   "keep_conditional": 0.8796,
   "keep_unconditional": 0.7615
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "symmetric_blind",
   "KEEP": 265,
   "SWITCH": 150,
   "RANDOMISE": 8,
   "THIRD_OPTION": 0,
   "REFUSE": 4,
   "UNPARSEABLE": 72,
   "n": 499,
   "keep_conditional": 0.6386,
   "keep_unconditional": 0.5311
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_keep",
   "KEEP": 15,
   "SWITCH": 409,
   "RANDOMISE": 63,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 11,
   "n": 500,
   "keep_conditional": 0.0354,
   "keep_unconditional": 0.03
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_switch",
   "KEEP": 387,
   "SWITCH": 39,
   "RANDOMISE": 59,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 13,
   "n": 500,
   "keep_conditional": 0.9085,
   "keep_unconditional": 0.774
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "neutral_switch",
   "KEEP": 281,
   "SWITCH": 132,
   "RANDOMISE": 73,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 14,
   "n": 500,
   "keep_conditional": 0.6804,
   "keep_unconditional": 0.562
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "symmetric_blind",
   "KEEP": 21,
   "SWITCH": 399,
   "RANDOMISE": 67,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 11,
   "n": 500,
   "keep_conditional": 0.05,
   "keep_unconditional": 0.042
  }
 ],
 "by_model_scenario_persona": [
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_keep",
   "persona": "none",
   "KEEP": 29,
   "SWITCH": 95,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.2339,
   "keep_unconditional": 0.232
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_keep",
   "persona": "average_person",
   "KEEP": 104,
   "SWITCH": 21,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.832,
   "keep_unconditional": 0.832
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_keep",
   "persona": "expert",
   "KEEP": 61,
   "SWITCH": 63,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.4919,
   "keep_unconditional": 0.488
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_keep",
   "persona": "adversary_bluff",
   "KEEP": 38,
   "SWITCH": 87,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.304,
   "keep_unconditional": 0.304
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_switch",
   "persona": "none",
   "KEEP": 19,
   "SWITCH": 105,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 1,
   "n": 125,
   "keep_conditional": 0.1532,
   "keep_unconditional": 0.152
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_switch",
   "persona": "average_person",
   "KEEP": 53,
   "SWITCH": 72,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.424,
   "keep_unconditional": 0.424
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_switch",
   "persona": "expert",
   "KEEP": 42,
   "SWITCH": 83,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.336,
   "keep_unconditional": 0.336
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "default_switch",
   "persona": "adversary_bluff",
   "KEEP": 32,
   "SWITCH": 93,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.256,
   "keep_unconditional": 0.256
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "neutral_switch",
   "persona": "none",
   "KEEP": 10,
   "SWITCH": 115,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.08,
   "keep_unconditional": 0.08
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "neutral_switch",
   "persona": "average_person",
   "KEEP": 42,
   "SWITCH": 83,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.336,
   "keep_unconditional": 0.336
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "neutral_switch",
   "persona": "expert",
   "KEEP": 9,
   "SWITCH": 114,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.0732,
   "keep_unconditional": 0.072
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "neutral_switch",
   "persona": "adversary_bluff",
   "KEEP": 38,
   "SWITCH": 85,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.3089,
   "keep_unconditional": 0.304
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "symmetric_blind",
   "persona": "none",
   "KEEP": 67,
   "SWITCH": 42,
   "RANDOMISE": 15,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 1,
   "n": 125,
   "keep_conditional": 0.6147,
   "keep_unconditional": 0.536
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "symmetric_blind",
   "persona": "average_person",
   "KEEP": 125,
   "SWITCH": 0,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 1.0,
   "keep_unconditional": 1.0
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "symmetric_blind",
   "persona": "expert",
   "KEEP": 122,
   "SWITCH": 1,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.9919,
   "keep_unconditional": 0.976
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "scenario": "symmetric_blind",
   "persona": "adversary_bluff",
   "KEEP": 41,
   "SWITCH": 83,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.3306,
   "keep_unconditional": 0.328
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_keep",
   "persona": "none",
   "KEEP": 60,
   "SWITCH": 54,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 9,
   "n": 125,
   "keep_conditional": 0.5263,
   "keep_unconditional": 0.48
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_keep",
   "persona": "average_person",
   "KEEP": 26,
   "SWITCH": 87,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 11,
   "n": 125,
   "keep_conditional": 0.2301,
   "keep_unconditional": 0.208
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_keep",
   "persona": "expert",
   "KEEP": 42,
   "SWITCH": 66,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 15,
   "n": 125,
   "keep_conditional": 0.3889,
   "keep_unconditional": 0.336
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_keep",
   "persona": "adversary_bluff",
   "KEEP": 27,
   "SWITCH": 75,
   "RANDOMISE": 11,
   "THIRD_OPTION": 1,
   "REFUSE": 0,
   "UNPARSEABLE": 11,
   "n": 125,
   "keep_conditional": 0.2647,
   "keep_unconditional": 0.216
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_switch",
   "persona": "none",
   "KEEP": 93,
   "SWITCH": 16,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 14,
   "n": 125,
   "keep_conditional": 0.8532,
   "keep_unconditional": 0.744
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_switch",
   "persona": "average_person",
   "KEEP": 99,
   "SWITCH": 11,
   "RANDOMISE": 0,
   "THIRD_OPTION": 1,
   "REFUSE": 5,
   "UNPARSEABLE": 9,
   "n": 125,
   "keep_conditional": 0.9,
   "keep_unconditional": 0.792
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_switch",
   "persona": "expert",
   "KEEP": 91,
   "SWITCH": 13,
   "RANDOMISE": 4,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 16,
   "n": 125,
   "keep_conditional": 0.875,
   "keep_unconditional": 0.728
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "default_switch",
   "persona": "adversary_bluff",
   "KEEP": 81,
   "SWITCH": 21,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 20,
   "n": 125,
   "keep_conditional": 0.7941,
   "keep_unconditional": 0.648
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "neutral_switch",
   "persona": "none",
   "KEEP": 83,
   "SWITCH": 30,
   "RANDOMISE": 0,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 10,
   "n": 125,
   "keep_conditional": 0.7345,
   "keep_unconditional": 0.664
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "neutral_switch",
   "persona": "average_person",
   "KEEP": 83,
   "SWITCH": 26,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 13,
   "n": 125,
   "keep_conditional": 0.7615,
   "keep_unconditional": 0.664
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "neutral_switch",
   "persona": "expert",
   "KEEP": 74,
   "SWITCH": 26,
   "RANDOMISE": 5,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 19,
   "n": 125,
   "keep_conditional": 0.74,
   "keep_unconditional": 0.592
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "neutral_switch",
   "persona": "adversary_bluff",
   "KEEP": 82,
   "SWITCH": 20,
   "RANDOMISE": 8,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 15,
   "n": 125,
   "keep_conditional": 0.8039,
   "keep_unconditional": 0.656
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "symmetric_blind",
   "persona": "none",
   "KEEP": 87,
   "SWITCH": 17,
   "RANDOMISE": 4,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 17,
   "n": 125,
   "keep_conditional": 0.8365,
   "keep_unconditional": 0.696
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "symmetric_blind",
   "persona": "average_person",
   "KEEP": 63,
   "SWITCH": 45,
   "RANDOMISE": 5,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 11,
   "n": 125,
   "keep_conditional": 0.5833,
   "keep_unconditional": 0.504
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "symmetric_blind",
   "persona": "expert",
   "KEEP": 67,
   "SWITCH": 37,
   "RANDOMISE": 7,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 14,
   "n": 125,
   "keep_conditional": 0.6442,
   "keep_unconditional": 0.536
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "scenario": "symmetric_blind",
   "persona": "adversary_bluff",
   "KEEP": 24,
   "SWITCH": 78,
   "RANDOMISE": 7,
   "THIRD_OPTION": 1,
   "REFUSE": 0,
   "UNPARSEABLE": 15,
   "n": 125,
   "keep_conditional": 0.2353,
   "keep_unconditional": 0.192
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_keep",
   "persona": "none",
   "KEEP": 6,
   "SWITCH": 80,
   "RANDOMISE": 3,
   "THIRD_OPTION": 0,
   "REFUSE": 3,
   "UNPARSEABLE": 33,
   "n": 125,
   "keep_conditional": 0.0698,
   "keep_unconditional": 0.048
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_keep",
   "persona": "average_person",
   "KEEP": 6,
   "SWITCH": 41,
   "RANDOMISE": 0,
   "THIRD_OPTION": 2,
   "REFUSE": 6,
   "UNPARSEABLE": 70,
   "n": 125,
   "keep_conditional": 0.1277,
   "keep_unconditional": 0.048
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_keep",
   "persona": "expert",
   "KEEP": 10,
   "SWITCH": 69,
   "RANDOMISE": 7,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 38,
   "n": 125,
   "keep_conditional": 0.1266,
   "keep_unconditional": 0.08
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_keep",
   "persona": "adversary_bluff",
   "KEEP": 12,
   "SWITCH": 35,
   "RANDOMISE": 4,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 72,
   "n": 125,
   "keep_conditional": 0.2553,
   "keep_unconditional": 0.096
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_switch",
   "persona": "none",
   "KEEP": 9,
   "SWITCH": 38,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 4,
   "UNPARSEABLE": 74,
   "n": 125,
   "keep_conditional": 0.1915,
   "keep_unconditional": 0.072
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_switch",
   "persona": "average_person",
   "KEEP": 10,
   "SWITCH": 78,
   "RANDOMISE": 3,
   "THIRD_OPTION": 0,
   "REFUSE": 5,
   "UNPARSEABLE": 28,
   "n": 124,
   "keep_conditional": 0.1136,
   "keep_unconditional": 0.0806
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_switch",
   "persona": "expert",
   "KEEP": 13,
   "SWITCH": 64,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 44,
   "n": 124,
   "keep_conditional": 0.1688,
   "keep_unconditional": 0.1048
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "default_switch",
   "persona": "adversary_bluff",
   "KEEP": 21,
   "SWITCH": 46,
   "RANDOMISE": 3,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 54,
   "n": 124,
   "keep_conditional": 0.3134,
   "keep_unconditional": 0.1694
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "neutral_switch",
   "persona": "none",
   "KEEP": 10,
   "SWITCH": 54,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 3,
   "UNPARSEABLE": 57,
   "n": 125,
   "keep_conditional": 0.1562,
   "keep_unconditional": 0.08
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "neutral_switch",
   "persona": "average_person",
   "KEEP": 14,
   "SWITCH": 72,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 3,
   "UNPARSEABLE": 36,
   "n": 125,
   "keep_conditional": 0.1628,
   "keep_unconditional": 0.112
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "neutral_switch",
   "persona": "expert",
   "KEEP": 26,
   "SWITCH": 39,
   "RANDOMISE": 6,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 52,
   "n": 125,
   "keep_conditional": 0.4,
   "keep_unconditional": 0.208
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "neutral_switch",
   "persona": "adversary_bluff",
   "KEEP": 15,
   "SWITCH": 28,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 82,
   "n": 125,
   "keep_conditional": 0.3488,
   "keep_unconditional": 0.12
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "symmetric_blind",
   "persona": "none",
   "KEEP": 6,
   "SWITCH": 78,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 40,
   "n": 125,
   "keep_conditional": 0.0714,
   "keep_unconditional": 0.048
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "symmetric_blind",
   "persona": "average_person",
   "KEEP": 1,
   "SWITCH": 49,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 65,
   "n": 118,
   "keep_conditional": 0.02,
   "keep_unconditional": 0.0085
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "symmetric_blind",
   "persona": "expert",
   "KEEP": 2,
   "SWITCH": 72,
   "RANDOMISE": 6,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 44,
   "n": 125,
   "keep_conditional": 0.027,
   "keep_unconditional": 0.016
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "scenario": "symmetric_blind",
   "persona": "adversary_bluff",
   "KEEP": 8,
   "SWITCH": 34,
   "RANDOMISE": 4,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 79,
   "n": 125,
   "keep_conditional": 0.1905,
   "keep_unconditional": 0.064
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_keep",
   "persona": "none",
   "KEEP": 36,
   "SWITCH": 70,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 17,
   "n": 125,
   "keep_conditional": 0.3396,
   "keep_unconditional": 0.288
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_keep",
   "persona": "average_person",
   "KEEP": 41,
   "SWITCH": 67,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 16,
   "n": 125,
   "keep_conditional": 0.3796,
   "keep_unconditional": 0.328
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_keep",
   "persona": "expert",
   "KEEP": 55,
   "SWITCH": 46,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 23,
   "n": 125,
   "keep_conditional": 0.5446,
   "keep_unconditional": 0.44
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_keep",
   "persona": "adversary_bluff",
   "KEEP": 74,
   "SWITCH": 26,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 24,
   "n": 125,
   "keep_conditional": 0.74,
   "keep_unconditional": 0.592
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_switch",
   "persona": "none",
   "KEEP": 92,
   "SWITCH": 20,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 12,
   "n": 125,
   "keep_conditional": 0.8214,
   "keep_unconditional": 0.736
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_switch",
   "persona": "average_person",
   "KEEP": 108,
   "SWITCH": 3,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 6,
   "UNPARSEABLE": 8,
   "n": 125,
   "keep_conditional": 0.973,
   "keep_unconditional": 0.864
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_switch",
   "persona": "expert",
   "KEEP": 96,
   "SWITCH": 5,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 23,
   "n": 125,
   "keep_conditional": 0.9505,
   "keep_unconditional": 0.768
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "default_switch",
   "persona": "adversary_bluff",
   "KEEP": 83,
   "SWITCH": 18,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 21,
   "n": 123,
   "keep_conditional": 0.8218,
   "keep_unconditional": 0.6748
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "neutral_switch",
   "persona": "none",
   "KEEP": 103,
   "SWITCH": 10,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 11,
   "n": 125,
   "keep_conditional": 0.9115,
   "keep_unconditional": 0.824
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "neutral_switch",
   "persona": "average_person",
   "KEEP": 101,
   "SWITCH": 15,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 7,
   "n": 125,
   "keep_conditional": 0.8707,
   "keep_unconditional": 0.808
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "neutral_switch",
   "persona": "expert",
   "KEEP": 93,
   "SWITCH": 10,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 21,
   "n": 125,
   "keep_conditional": 0.9029,
   "keep_unconditional": 0.744
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "neutral_switch",
   "persona": "adversary_bluff",
   "KEEP": 83,
   "SWITCH": 17,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 22,
   "n": 124,
   "keep_conditional": 0.83,
   "keep_unconditional": 0.6694
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "symmetric_blind",
   "persona": "none",
   "KEEP": 44,
   "SWITCH": 58,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 22,
   "n": 125,
   "keep_conditional": 0.4314,
   "keep_unconditional": 0.352
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "symmetric_blind",
   "persona": "average_person",
   "KEEP": 66,
   "SWITCH": 45,
   "RANDOMISE": 2,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 10,
   "n": 125,
   "keep_conditional": 0.5946,
   "keep_unconditional": 0.528
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "symmetric_blind",
   "persona": "expert",
   "KEEP": 76,
   "SWITCH": 26,
   "RANDOMISE": 5,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 16,
   "n": 124,
   "keep_conditional": 0.7451,
   "keep_unconditional": 0.6129
  },
  {
   "model": "openai/gpt-4o-mini",
   "scenario": "symmetric_blind",
   "persona": "adversary_bluff",
   "KEEP": 79,
   "SWITCH": 21,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 24,
   "n": 125,
   "keep_conditional": 0.79,
   "keep_unconditional": 0.632
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_keep",
   "persona": "none",
   "KEEP": 0,
   "SWITCH": 125,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.0,
   "keep_unconditional": 0.0
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_keep",
   "persona": "average_person",
   "KEEP": 0,
   "SWITCH": 125,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.0,
   "keep_unconditional": 0.0
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_keep",
   "persona": "expert",
   "KEEP": 0,
   "SWITCH": 125,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.0,
   "keep_unconditional": 0.0
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_keep",
   "persona": "adversary_bluff",
   "KEEP": 15,
   "SWITCH": 34,
   "RANDOMISE": 63,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 11,
   "n": 125,
   "keep_conditional": 0.3061,
   "keep_unconditional": 0.12
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_switch",
   "persona": "none",
   "KEEP": 125,
   "SWITCH": 0,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 1.0,
   "keep_unconditional": 1.0
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_switch",
   "persona": "average_person",
   "KEEP": 125,
   "SWITCH": 0,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 1.0,
   "keep_unconditional": 1.0
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_switch",
   "persona": "expert",
   "KEEP": 115,
   "SWITCH": 9,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 1,
   "n": 125,
   "keep_conditional": 0.9274,
   "keep_unconditional": 0.92
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "default_switch",
   "persona": "adversary_bluff",
   "KEEP": 22,
   "SWITCH": 30,
   "RANDOMISE": 59,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 12,
   "n": 125,
   "keep_conditional": 0.4231,
   "keep_unconditional": 0.176
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "neutral_switch",
   "persona": "none",
   "KEEP": 114,
   "SWITCH": 7,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 4,
   "n": 125,
   "keep_conditional": 0.9421,
   "keep_unconditional": 0.912
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "neutral_switch",
   "persona": "average_person",
   "KEEP": 110,
   "SWITCH": 13,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 2,
   "n": 125,
   "keep_conditional": 0.8943,
   "keep_unconditional": 0.88
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "neutral_switch",
   "persona": "expert",
   "KEEP": 44,
   "SWITCH": 81,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.352,
   "keep_unconditional": 0.352
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "neutral_switch",
   "persona": "adversary_bluff",
   "KEEP": 13,
   "SWITCH": 31,
   "RANDOMISE": 73,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 8,
   "n": 125,
   "keep_conditional": 0.2955,
   "keep_unconditional": 0.104
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "symmetric_blind",
   "persona": "none",
   "KEEP": 0,
   "SWITCH": 125,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.0,
   "keep_unconditional": 0.0
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "symmetric_blind",
   "persona": "average_person",
   "KEEP": 8,
   "SWITCH": 117,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.064,
   "keep_unconditional": 0.064
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "symmetric_blind",
   "persona": "expert",
   "KEEP": 0,
   "SWITCH": 125,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 125,
   "keep_conditional": 0.0,
   "keep_unconditional": 0.0
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "scenario": "symmetric_blind",
   "persona": "adversary_bluff",
   "KEEP": 13,
   "SWITCH": 32,
   "RANDOMISE": 67,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 11,
   "n": 125,
   "keep_conditional": 0.2889,
   "keep_unconditional": 0.104
  }
 ],
 "by_model_temperature": [
  {
   "model": "google/gemma-4-26b-a4b-it",
   "temperature": 0.0,
   "KEEP": 165,
   "SWITCH": 223,
   "RANDOMISE": 11,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 1,
   "n": 400,
   "keep_conditional": 0.4253,
   "keep_unconditional": 0.4125
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "temperature": 0.4,
   "KEEP": 163,
   "SWITCH": 232,
   "RANDOMISE": 4,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 1,
   "n": 400,
   "keep_conditional": 0.4127,
   "keep_unconditional": 0.4075
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "temperature": 0.8,
   "KEEP": 159,
   "SWITCH": 240,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.3985,
   "keep_unconditional": 0.3975
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "temperature": 1.2,
   "KEEP": 168,
   "SWITCH": 229,
   "RANDOMISE": 3,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.4232,
   "keep_unconditional": 0.42
  },
  {
   "model": "google/gemma-4-26b-a4b-it",
   "temperature": 1.6,
   "KEEP": 177,
   "SWITCH": 218,
   "RANDOMISE": 4,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.4481,
   "keep_unconditional": 0.4425
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "temperature": 0.0,
   "KEEP": 304,
   "SWITCH": 94,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 2,
   "n": 400,
   "keep_conditional": 0.7638,
   "keep_unconditional": 0.76
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "temperature": 0.4,
   "KEEP": 237,
   "SWITCH": 162,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 1,
   "n": 400,
   "keep_conditional": 0.594,
   "keep_unconditional": 0.5925
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "temperature": 0.8,
   "KEEP": 236,
   "SWITCH": 163,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.5915,
   "keep_unconditional": 0.59
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "temperature": 1.2,
   "KEEP": 233,
   "SWITCH": 163,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 2,
   "n": 400,
   "keep_conditional": 0.5884,
   "keep_unconditional": 0.5825
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "temperature": 1.6,
   "KEEP": 72,
   "SWITCH": 40,
   "RANDOMISE": 57,
   "THIRD_OPTION": 4,
   "REFUSE": 13,
   "UNPARSEABLE": 214,
   "n": 400,
   "keep_conditional": 0.6429,
   "keep_unconditional": 0.18
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "temperature": 0.0,
   "KEEP": 27,
   "SWITCH": 236,
   "RANDOMISE": 1,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 135,
   "n": 399,
   "keep_conditional": 0.1027,
   "keep_unconditional": 0.0677
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "temperature": 0.4,
   "KEEP": 47,
   "SWITCH": 290,
   "RANDOMISE": 5,
   "THIRD_OPTION": 0,
   "REFUSE": 2,
   "UNPARSEABLE": 56,
   "n": 400,
   "keep_conditional": 0.1395,
   "keep_unconditional": 0.1175
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "temperature": 0.8,
   "KEEP": 64,
   "SWITCH": 276,
   "RANDOMISE": 5,
   "THIRD_OPTION": 1,
   "REFUSE": 5,
   "UNPARSEABLE": 49,
   "n": 400,
   "keep_conditional": 0.1882,
   "keep_unconditional": 0.16
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "temperature": 1.2,
   "KEEP": 28,
   "SWITCH": 70,
   "RANDOMISE": 25,
   "THIRD_OPTION": 0,
   "REFUSE": 15,
   "UNPARSEABLE": 261,
   "n": 399,
   "keep_conditional": 0.2857,
   "keep_unconditional": 0.0702
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "temperature": 1.6,
   "KEEP": 3,
   "SWITCH": 5,
   "RANDOMISE": 5,
   "THIRD_OPTION": 1,
   "REFUSE": 11,
   "UNPARSEABLE": 367,
   "n": 392,
   "keep_conditional": 0.375,
   "keep_unconditional": 0.0077
  },
  {
   "model": "openai/gpt-4o-mini",
   "temperature": 0.0,
   "KEEP": 281,
   "SWITCH": 119,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.7025,
   "keep_unconditional": 0.7025
  },
  {
   "model": "openai/gpt-4o-mini",
   "temperature": 0.4,
   "KEEP": 287,
   "SWITCH": 113,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.7175,
   "keep_unconditional": 0.7175
  },
  {
   "model": "openai/gpt-4o-mini",
   "temperature": 0.8,
   "KEEP": 293,
   "SWITCH": 107,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.7325,
   "keep_unconditional": 0.7325
  },
  {
   "model": "openai/gpt-4o-mini",
   "temperature": 1.2,
   "KEEP": 310,
   "SWITCH": 90,
   "RANDOMISE": 0,
   "THIRD_OPTION": 0,
   "REFUSE": 0,
   "UNPARSEABLE": 0,
   "n": 400,
   "keep_conditional": 0.775,
   "keep_unconditional": 0.775
  },
  {
   "model": "openai/gpt-4o-mini",
   "temperature": 1.6,
   "KEEP": 59,
   "SWITCH": 28,
   "RANDOMISE": 13,
   "THIRD_OPTION": 0,
   "REFUSE": 19,
   "UNPARSEABLE": 277,
   "n": 396,
   "keep_conditional": 0.6782,
   "keep_unconditional": 0.149
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "temperature": 0.0,
   "KEEP": 135,
   "SWITCH": 205,
   "RANDOMISE": 50,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 9,
   "n": 400,
   "keep_conditional": 0.3971,
   "keep_unconditional": 0.3375
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "temperature": 0.4,
   "KEEP": 149,
   "SWITCH": 187,
   "RANDOMISE": 49,
   "THIRD_OPTION": 1,
   "REFUSE": 1,
   "UNPARSEABLE": 13,
   "n": 400,
   "keep_conditional": 0.4435,
   "keep_unconditional": 0.3725
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "temperature": 0.8,
   "KEEP": 144,
   "SWITCH": 193,
   "RANDOMISE": 49,
   "THIRD_OPTION": 1,
   "REFUSE": 0,
   "UNPARSEABLE": 13,
   "n": 400,
   "keep_conditional": 0.4273,
   "keep_unconditional": 0.36
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "temperature": 1.2,
   "KEEP": 138,
   "SWITCH": 198,
   "RANDOMISE": 54,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 9,
   "n": 400,
   "keep_conditional": 0.4107,
   "keep_unconditional": 0.345
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "temperature": 1.6,
   "KEEP": 138,
   "SWITCH": 196,
   "RANDOMISE": 60,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 5,
   "n": 400,
   "keep_conditional": 0.4132,
   "keep_unconditional": 0.345
  }
 ],
 "by_temperature": [
  {
   "temperature": 0.0,
   "KEEP": 912,
   "SWITCH": 877,
   "RANDOMISE": 62,
   "THIRD_OPTION": 0,
   "REFUSE": 1,
   "UNPARSEABLE": 147,
   "n": 1999,
   "keep_conditional": 0.5098,
   "keep_unconditional": 0.4562,
   "median_output_tokens": 142,
   "hit_token_cap": 0.051
  },
  {
   "temperature": 0.4,
   "KEEP": 883,
   "SWITCH": 984,
   "RANDOMISE": 58,
   "THIRD_OPTION": 1,
   "REFUSE": 3,
   "UNPARSEABLE": 71,
   "n": 2000,
   "keep_conditional": 0.473,
   "keep_unconditional": 0.4415,
   "median_output_tokens": 163.5,
   "hit_token_cap": 0.052
  },
  {
   "temperature": 0.8,
   "KEEP": 896,
   "SWITCH": 979,
   "RANDOMISE": 56,
   "THIRD_OPTION": 2,
   "REFUSE": 5,
   "UNPARSEABLE": 62,
   "n": 2000,
   "keep_conditional": 0.4779,
   "keep_unconditional": 0.448,
   "median_output_tokens": 193.0,
   "hit_token_cap": 0.048
  },
  {
   "temperature": 1.2,
   "KEEP": 877,
   "SWITCH": 750,
   "RANDOMISE": 83,
   "THIRD_OPTION": 0,
   "REFUSE": 17,
   "UNPARSEABLE": 272,
   "n": 1999,
   "keep_conditional": 0.539,
   "keep_unconditional": 0.4387,
   "median_output_tokens": 280,
   "hit_token_cap": 0.1761
  },
  {
   "temperature": 1.6,
   "KEEP": 449,
   "SWITCH": 487,
   "RANDOMISE": 139,
   "THIRD_OPTION": 5,
   "REFUSE": 45,
   "UNPARSEABLE": 863,
   "n": 1988,
   "keep_conditional": 0.4797,
   "keep_unconditional": 0.2259,
   "median_output_tokens": 1889.5,
   "hit_token_cap": 0.495
  }
 ],
 "depth_by_persona": [
  {
   "persona": "none",
   "L0": 0.265,
   "L1": 0.551,
   "L2": 0.0395,
   "L3": 0.0015,
   "MIX": 0.0075,
   "NON": 0.1355,
   "n": 2000
  },
  {
   "persona": "average_person",
   "L0": 0.3102,
   "L1": 0.5492,
   "L2": 0.0176,
   "L3": 0.001,
   "MIX": 0.0065,
   "NON": 0.1155,
   "n": 1992
  },
  {
   "persona": "expert",
   "L0": 0.1471,
   "L1": 0.5521,
   "L2": 0.1126,
   "L3": 0.003,
   "MIX": 0.025,
   "NON": 0.1602,
   "n": 1998
  },
  {
   "persona": "adversary_bluff",
   "L0": 0.0075,
   "L1": 0.0421,
   "L2": 0.5105,
   "L3": 0.2485,
   "MIX": 0.0175,
   "NON": 0.1738,
   "n": 1996
  }
 ],
 "depth_by_model": [
  {
   "model": "google/gemma-4-26b-a4b-it",
   "L0": 0.2325,
   "L1": 0.43,
   "L2": 0.2185,
   "L3": 0.1075,
   "MIX": 0.009,
   "NON": 0.0025,
   "excluded_from_pooled": false
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "L0": 0.1865,
   "L1": 0.4675,
   "L2": 0.1685,
   "L3": 0.0455,
   "MIX": 0.027,
   "NON": 0.105,
   "excluded_from_pooled": false
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "L0": 0.1779,
   "L1": 0.2905,
   "L2": 0.0834,
   "L3": 0.093,
   "MIX": 0.0161,
   "NON": 0.3392,
   "excluded_from_pooled": false
  },
  {
   "model": "openai/gpt-4o-mini",
   "L0": 0.1328,
   "L1": 0.506,
   "L2": 0.2094,
   "L3": 0.008,
   "MIX": 0.0045,
   "NON": 0.1393,
   "excluded_from_pooled": false
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "L0": 0.748,
   "L1": 0.0295,
   "L2": 0.041,
   "L3": 0.067,
   "MIX": 0.101,
   "NON": 0.0135,
   "excluded_from_pooled": true
  }
 ],
 "flip_rate": [
  {
   "model": "google/gemma-4-26b-a4b-it",
   "keep_default_keep": 0.4659,
   "keep_default_switch": 0.2926,
   "flip": 0.1733
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "keep_default_keep": 0.3547,
   "keep_default_switch": 0.8565,
   "flip": -0.5018
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "keep_default_keep": 0.1313,
   "keep_default_switch": 0.19,
   "flip": -0.0587
  },
  {
   "model": "openai/gpt-4o-mini",
   "keep_default_keep": 0.4964,
   "keep_default_switch": 0.8918,
   "flip": -0.3954
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "keep_default_keep": 0.0354,
   "keep_default_switch": 0.9085,
   "flip": -0.8731
  }
 ],
 "determinism": [
  {
   "model": "google/gemma-4-26b-a4b-it",
   "unanimous_cells": 11,
   "cells": 80,
   "share": 0.1375
  },
  {
   "model": "ibm-granite/granite-4.1-8b",
   "unanimous_cells": 8,
   "cells": 69,
   "share": 0.1159
  },
  {
   "model": "meta-llama/llama-3.1-8b-instruct",
   "unanimous_cells": 12,
   "cells": 49,
   "share": 0.2449
  },
  {
   "model": "openai/gpt-4o-mini",
   "unanimous_cells": 8,
   "cells": 69,
   "share": 0.1159
  },
  {
   "model": "qwen/qwen3-30b-a3b-instruct-2507",
   "unanimous_cells": 49,
   "cells": 70,
   "share": 0.7
  }
 ],
 "human_data": {
  "main": {
   "condition": "one-on-one, time to think",
   "keep_pct": 90,
   "n": 150
  },
  "replications": [
   {
    "asker": "second",
    "keep_pct": 80
   },
   {
    "asker": "third",
    "keep_pct": 95
   }
  ],
  "variations": [
   {
    "condition": "told he has won 3 times before",
    "switch_pct": 20,
    "n": 50
   },
   {
    "condition": "asked ~20 people at once",
    "switch_pct": 30,
    "n": 60,
    "note": "3 groups of 20"
   },
   {
    "condition": "little time to think (Monty Hall pattern-match)",
    "switch_pct": 10,
    "n": 5,
    "note": "very small n, anecdotal"
   }
  ],
  "caveat": "Sample skews young (15-25) and mathematically able. Collected over ~18 months; figures are approximate."
 },
 "informal_frontier": {
  "model": "Claude Opus",
  "keep_pct": 90,
  "n": 10,
  "note": "informal, not part of the main run"
 }
}