{
  "created": "2026-09-20T09:21:19.088654+00:00",
  "stages": {
    "A": {
      "raw_trajectories": 15700,
      "unique_serialized_trajectories": 15700,
      "kinds": {
        "public_tool": 1200,
        "instruction": 10000,
        "single": 3500,
        "negative": 1000
      },
      "prepared_pool": {
        "response_samples": 19572,
        "input_tokens": 6276390,
        "supervised_target_tokens": 1008247,
        "target_fraction": 0.16064122847687923,
        "average_target_tokens_per_response": 51.514765992233805
      },
      "actual_training_run": {
        "stage": "A",
        "initial_dev_loss": 2.1385669708251953,
        "best_dev_loss": 1.029488444328308,
        "steps": 306,
        "total_steps": 306,
        "processed_tokens": 6275843,
        "finite": true,
        "model_path": "/workspace/runs/pilot-20260916-074636/A/best"
      }
    },
    "A_repair": {
      "prepared_pool": {
        "response_samples": 16700,
        "input_tokens": 6919583,
        "supervised_target_tokens": 551783,
        "target_fraction": 0.0797422330218454,
        "average_target_tokens_per_response": 33.04089820359282
      },
      "actual_training_run": {
        "stage": "A_repair",
        "initial_dev_loss": 1.0294885635375977,
        "best_dev_loss": 1.0132355690002441,
        "steps": 160,
        "total_steps": 160,
        "processed_tokens": 4253228,
        "finite": true,
        "model_path": "/workspace/runs/pilot-20260916-074636/A_repair/best"
      }
    },
    "B": {
      "raw_trajectories": 22550,
      "unique_serialized_trajectories": 22550,
      "kinds": {
        "single": 13955,
        "negative": 3882,
        "public_tool": 3213,
        "instruction": 1500
      },
      "prepared_pool": {
        "response_samples": 42430,
        "input_tokens": 18119246,
        "supervised_target_tokens": 1101348,
        "target_fraction": 0.060783323985998096,
        "average_target_tokens_per_response": 25.956823002592504
      },
      "actual_training_run": {
        "stage": "B",
        "initial_dev_loss": 0.9850825071334839,
        "best_dev_loss": 0.7768568396568298,
        "steps": 663,
        "total_steps": 663,
        "processed_tokens": 18113190,
        "finite": true,
        "model_path": "/workspace/runs/pilot-20260916-074636/B/best"
      }
    },
    "C": {
      "raw_trajectories": 10500,
      "unique_serialized_trajectories": 10500,
      "kinds": {
        "multi": 6000,
        "single": 1500,
        "public_tool": 1000,
        "negative": 1000,
        "instruction": 1000
      },
      "prepared_pool": {
        "response_samples": 28017,
        "input_tokens": 15177959,
        "supervised_target_tokens": 701456,
        "target_fraction": 0.046215436476011035,
        "average_target_tokens_per_response": 25.036799086269053
      },
      "actual_training_run": {
        "stage": "C",
        "initial_dev_loss": 0.7825645208358765,
        "best_dev_loss": 0.6296705007553101,
        "steps": 438,
        "total_steps": 438,
        "processed_tokens": 15177428,
        "finite": true,
        "model_path": "/workspace/runs/pilot-20260916-074636/C/best"
      }
    },
    "D": {
      "raw_trajectories": 15500,
      "unique_serialized_trajectories": 15500,
      "kinds": {
        "single": 10000,
        "public_tool": 500,
        "instruction": 1000,
        "multi": 2000,
        "negative": 2000
      },
      "prepared_pool": {
        "response_samples": 31735,
        "input_tokens": 16878302,
        "supervised_target_tokens": 703072,
        "target_fraction": 0.04165537504898301,
        "average_target_tokens_per_response": 22.154466677170316
      },
      "actual_training_run": {
        "stage": "D",
        "initial_dev_loss": 0.6296705007553101,
        "best_dev_loss": 0.618056058883667,
        "steps": 300,
        "total_steps": 300,
        "processed_tokens": 10214957,
        "finite": true,
        "model_path": "/workspace/runs/pilot-20260916-074636/D/best"
      }
    },
    "E": {
      "raw_trajectories": 13500,
      "unique_serialized_trajectories": 13500,
      "kinds": {
        "multi": 6000,
        "negative": 2000,
        "instruction": 1000,
        "single": 4000,
        "public_tool": 500
      },
      "prepared_pool": {
        "response_samples": 34402,
        "input_tokens": 18525689,
        "supervised_target_tokens": 748083,
        "target_fraction": 0.040380846294029875,
        "average_target_tokens_per_response": 21.745334573571306
      },
      "actual_training_run": {
        "stage": "E",
        "initial_dev_loss": 0.6296705007553101,
        "best_dev_loss": 0.6191723346710205,
        "steps": 400,
        "total_steps": 400,
        "processed_tokens": 13790817,
        "finite": true,
        "model_path": "/workspace/runs/pilot-20260916-074636/E/best"
      }
    },
    "F": {
      "raw_trajectories": 16487,
      "unique_serialized_trajectories": 16487,
      "kinds": {
        "single_python": 3200,
        "instruction": 600,
        "single": 1500,
        "multi_python": 3200,
        "multi": 4000,
        "multi_json": 800,
        "public_tool_json": 233,
        "single_json": 800,
        "public_tool_python": 954,
        "negative_json": 120,
        "negative_python": 480,
        "negative": 600
      },
      "prepared_pool": {
        "response_samples": 45531,
        "input_tokens": 51487862,
        "supervised_target_tokens": 976587,
        "target_fraction": 0.018967324764815444,
        "average_target_tokens_per_response": 21.44883705607169
      },
      "actual_training_run": {
        "stage": "F",
        "initial_dev_loss": 0.7478622794151306,
        "best_dev_loss": 0.52776700258255,
        "steps": 400,
        "total_steps": 400,
        "processed_tokens": 28946298,
        "finite": true,
        "model_path": "/workspace/runs/current/F/best"
      }
    }
  },
  "union_raw": {
    "unique_serialized_trajectories": 56837,
    "unique_user_texts": 47807,
    "tool_names": 2706,
    "tool_schema_documents": 5553
  },
  "checkpoint_lineages": {
    "C": [
      "A",
      "A_repair",
      "B",
      "C"
    ],
    "E": [
      "A",
      "A_repair",
      "B",
      "C",
      "E"
    ],
    "F": [
      "A",
      "A_repair",
      "B",
      "C",
      "F"
    ]
  },
  "limits": [
    "Prepared token pools are not all consumed: max_steps and replay change actual exposure.",
    "Target tokens count labels != -100; prompt tokens contribute context, not direct target loss.",
    "Exact serialized dedup is not semantic dedup; name and format rewrites inflate apparent diversity.",
    "A_repair has no raw trajectory file, so union_raw excludes its raw view; token pool is audited.",
    "Independent experiment totals must not be mistaken for one checkpoint training lineage."
  ]
}