eval_results_v2.json

{
  "metrics": {
    "A": {"accuracy": 0.400, "avg_cost": 1.000, "xRandom": 3.6, "xMaj": 1.7},
    "B": {"accuracy": 0.510, "avg_cost": 0.150, "xRandom": 4.6, "xMaj": 2.1},
    "C": {"accuracy": 0.400, "avg_cost": 1.250, "accept_rate": 0.000, "xRandom": 3.6, "xMaj": 1.7},
    "D": {"accuracy": 0.510, "avg_cost": 0.250, "accept_rate": 0.215, "mean_score": -1.515, "xRandom": 4.6, "xMaj": 2.1},
    "E": {"accuracy": 0.420, "avg_cost": 0.750, "xRandom": 3.8, "xMaj": 1.8}
  },
  "baselines": {
    "random": 0.111,
    "majority": 0.240,
    "majority_class": "final_answer"
  },
  "n": 200,
  "distribution": {
    "ask_clarification": 20, "final_answer": 48, "file_read": 27,
    "tool_call": 31, "repair": 10, "retrieval": 27,
    "file_write": 18, "verifier": 14, "BLOCKED": 5
  },
  "pareto_frontier": [{"config": "B", "cost": 0.150, "accuracy": 0.510}],
  "winner": "B",
  "conclusion": "Cheap fine-tuned proposer alone is optimal. Speculative routing fails when verifiers cannot discriminate quality.",
  "hardware": "A-D: dual A10G (a10g-largex2), E: single A10G"
}