Rules 3.1/4.1/5.1: rebuilt packs and re-pilots, hindsight capture as every board's headline, the registered Qwen3-4B training run (part 4)

training/4b/eval_checkpoints.json

@@ -0,0 +1,990 @@
+[
+ {
+  "checkpoint_step": 25,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7496,
+  "reward_std": 0.2566,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.86,
+  "rate_ok": 0.82,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.04,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.02,
+  "rate_cycle_cap": 0.04,
+  "rate_end_short": 0.04,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 245.4,
+  "ci95": [
+   0.6772,
+   0.8171
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.7897,
+    "ci95": [
+     0.6391,
+     0.9152
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.7038,
+    "ci95": [
+     0.5893,
+     0.8058
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.761,
+    "ci95": [
+     0.6545,
+     0.8549
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.7386,
+    "ci95": [
+     0.5992,
+     0.8538
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.7378,
+    "ci95": [
+     0.557,
+     0.889
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7628,
+    "ci95": [
+     0.6605,
+     0.8541
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 50,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7022,
+  "reward_std": 0.2929,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.84,
+  "rate_ok": 0.8,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.06,
+  "rate_sell_before_charge": 0.1,
+  "rate_rate": 0.0,
+  "rate_cycle_cap": 0.0,
+  "rate_end_short": 0.04,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 240.7,
+  "ci95": [
+   0.6202,
+   0.7801
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.7227,
+    "ci95": [
+     0.5708,
+     0.8569
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6519,
+    "ci95": [
+     0.5214,
+     0.7706
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.7472,
+    "ci95": [
+     0.5798,
+     0.8831
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8179,
+    "ci95": [
+     0.7122,
+     0.9096
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6148,
+    "ci95": [
+     0.4753,
+     0.7505
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.6547,
+    "ci95": [
+     0.5101,
+     0.7899
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 75,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7406,
+  "reward_std": 0.2475,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.9,
+  "rate_ok": 0.82,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.04,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.0,
+  "rate_cycle_cap": 0.02,
+  "rate_end_short": 0.08,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 247.2,
+  "ci95": [
+   0.6696,
+   0.8062
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.8845,
+    "ci95": [
+     0.8384,
+     0.929
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6522,
+    "ci95": [
+     0.5332,
+     0.7626
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.6707,
+    "ci95": [
+     0.5066,
+     0.8176
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.7711,
+    "ci95": [
+     0.6687,
+     0.8582
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6458,
+    "ci95": [
+     0.4667,
+     0.8084
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7593,
+    "ci95": [
+     0.651,
+     0.8552
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 100,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7932,
+  "reward_std": 0.1912,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.96,
+  "rate_ok": 0.94,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.0,
+  "rate_cycle_cap": 0.0,
+  "rate_end_short": 0.02,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 244.9,
+  "ci95": [
+   0.7373,
+   0.8445
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.884,
+    "ci95": [
+     0.8292,
+     0.9338
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6879,
+    "ci95": [
+     0.5812,
+     0.7867
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.8214,
+    "ci95": [
+     0.748,
+     0.8854
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8104,
+    "ci95": [
+     0.7399,
+     0.8728
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.7283,
+    "ci95": [
+     0.5657,
+     0.8567
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.8087,
+    "ci95": [
+     0.7253,
+     0.8833
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 125,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7609,
+  "reward_std": 0.2299,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.9,
+  "rate_ok": 0.86,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.02,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.0,
+  "rate_cycle_cap": 0.04,
+  "rate_end_short": 0.04,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 247.2,
+  "ci95": [
+   0.6965,
+   0.8232
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.879,
+    "ci95": [
+     0.8235,
+     0.9312
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6832,
+    "ci95": [
+     0.5649,
+     0.795
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.7108,
+    "ci95": [
+     0.5777,
+     0.8318
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.7464,
+    "ci95": [
+     0.6527,
+     0.8258
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6993,
+    "ci95": [
+     0.5488,
+     0.8363
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7984,
+    "ci95": [
+     0.6922,
+     0.8928
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 150,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7952,
+  "reward_std": 0.2318,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.86,
+  "rate_ok": 0.84,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.02,
+  "rate_sell_before_charge": 0.06,
+  "rate_rate": 0.02,
+  "rate_cycle_cap": 0.04,
+  "rate_end_short": 0.02,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 243.0,
+  "ci95": [
+   0.7252,
+   0.8555
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.9051,
+    "ci95": [
+     0.8656,
+     0.9441
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6451,
+    "ci95": [
+     0.5016,
+     0.7771
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.8624,
+    "ci95": [
+     0.8183,
+     0.9052
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8381,
+    "ci95": [
+     0.7883,
+     0.8803
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.7073,
+    "ci95": [
+     0.4957,
+     0.879
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.8017,
+    "ci95": [
+     0.6924,
+     0.8959
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 175,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7909,
+  "reward_std": 0.2135,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.9,
+  "rate_ok": 0.82,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.02,
+  "rate_rate": 0.02,
+  "rate_cycle_cap": 0.06,
+  "rate_end_short": 0.08,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 244.6,
+  "ci95": [
+   0.7282,
+   0.8478
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.8961,
+    "ci95": [
+     0.8547,
+     0.9365
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6675,
+    "ci95": [
+     0.5435,
+     0.7825
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.8257,
+    "ci95": [
+     0.7409,
+     0.894
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8402,
+    "ci95": [
+     0.7749,
+     0.8912
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6929,
+    "ci95": [
+     0.5217,
+     0.837
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7971,
+    "ci95": [
+     0.6999,
+     0.8863
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 200,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.737,
+  "reward_std": 0.2746,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.88,
+  "rate_ok": 0.82,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.08,
+  "rate_rate": 0.02,
+  "rate_cycle_cap": 0.02,
+  "rate_end_short": 0.06,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 247.4,
+  "ci95": [
+   0.6579,
+   0.8099
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.8541,
+    "ci95": [
+     0.7971,
+     0.9088
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6072,
+    "ci95": [
+     0.4594,
+     0.7462
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.7645,
+    "ci95": [
+     0.5953,
+     0.8923
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8043,
+    "ci95": [
+     0.7465,
+     0.8568
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.5699,
+    "ci95": [
+     0.3371,
+     0.7872
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7599,
+    "ci95": [
+     0.6438,
+     0.8686
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 225,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7521,
+  "reward_std": 0.259,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.92,
+  "rate_ok": 0.86,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.04,
+  "rate_cycle_cap": 0.0,
+  "rate_end_short": 0.06,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 242.8,
+  "ci95": [
+   0.6762,
+   0.8209
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.8807,
+    "ci95": [
+     0.8337,
+     0.9266
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6327,
+    "ci95": [
+     0.5031,
+     0.756
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.7483,
+    "ci95": [
+     0.5672,
+     0.8846
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.827,
+    "ci95": [
+     0.7741,
+     0.8742
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6002,
+    "ci95": [
+     0.3752,
+     0.8056
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7627,
+    "ci95": [
+     0.6507,
+     0.8639
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 250,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.778,
+  "reward_std": 0.2312,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.92,
+  "rate_ok": 0.84,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.02,
+  "rate_cycle_cap": 0.02,
+  "rate_end_short": 0.08,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 241.2,
+  "ci95": [
+   0.7093,
+   0.8393
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.8936,
+    "ci95": [
+     0.8452,
+     0.9396
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6504,
+    "ci95": [
+     0.5329,
+     0.761
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.8042,
+    "ci95": [
+     0.6598,
+     0.8999
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8368,
+    "ci95": [
+     0.7738,
+     0.8916
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6727,
+    "ci95": [
+     0.4617,
+     0.8508
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7802,
+    "ci95": [
+     0.6823,
+     0.869
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 275,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7669,
+  "reward_std": 0.2417,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.9,
+  "rate_ok": 0.82,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.04,
+  "rate_cycle_cap": 0.02,
+  "rate_end_short": 0.08,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 242.1,
+  "ci95": [
+   0.6967,
+   0.8314
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.8936,
+    "ci95": [
+     0.8452,
+     0.9396
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6504,
+    "ci95": [
+     0.5329,
+     0.761
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.7615,
+    "ci95": [
+     0.5954,
+     0.8872
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8387,
+    "ci95": [
+     0.7746,
+     0.8932
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6142,
+    "ci95": [
+     0.4007,
+     0.81
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7802,
+    "ci95": [
+     0.6823,
+     0.869
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 300,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.7743,
+  "reward_std": 0.2441,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.9,
+  "rate_ok": 0.82,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.04,
+  "rate_cycle_cap": 0.02,
+  "rate_end_short": 0.08,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 240.5,
+  "ci95": [
+   0.7036,
+   0.838
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.8938,
+    "ci95": [
+     0.8454,
+     0.9396
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6773,
+    "ci95": [
+     0.5578,
+     0.7883
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.7508,
+    "ci95": [
+     0.5686,
+     0.8872
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8388,
+    "ci95": [
+     0.7748,
+     0.8933
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6002,
+    "ci95": [
+     0.3752,
+     0.8056
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.8024,
+    "ci95": [
+     0.7093,
+     0.8867
+    ]
+   }
+  }
+ },
+ {
+  "checkpoint_step": 325,
+  "n_tasks": 50,
+  "n": 50,
+  "reward_mean": 0.772,
+  "reward_std": 0.2416,
+  "valid_rate": 1.0,
+  "tool_call_rate": 1.0,
+  "feasible_rate": 0.9,
+  "rate_ok": 0.82,
+  "rate_all_idle": 0.0,
+  "rate_overfill": 0.0,
+  "rate_sell_before_charge": 0.04,
+  "rate_rate": 0.04,
+  "rate_cycle_cap": 0.02,
+  "rate_end_short": 0.08,
+  "rate_malformed": 0.0,
+  "rate_truncated": 0.0,
+  "completion_len_mean": 243.3,
+  "ci95": [
+   0.7013,
+   0.8371
+  ],
+  "by_start": {
+   "empty": {
+    "n": 18,
+    "reward_mean": 0.9057,
+    "ci95": [
+     0.8635,
+     0.946
+    ]
+   },
+   "full": {
+    "n": 19,
+    "reward_mean": 0.6524,
+    "ci95": [
+     0.5351,
+     0.7625
+    ]
+   },
+   "half": {
+    "n": 13,
+    "reward_mean": 0.7615,
+    "ci95": [
+     0.5954,
+     0.8872
+    ]
+   }
+  },
+  "by_curriculum": {
+   "C0": {
+    "n": 17,
+    "reward_mean": 0.8555,
+    "ci95": [
+     0.7998,
+     0.9039
+    ]
+   },
+   "C1": {
+    "n": 10,
+    "reward_mean": 0.6142,
+    "ci95": [
+     0.4007,
+     0.81
+    ]
+   },
+   "C2": {
+    "n": 23,
+    "reward_mean": 0.7788,
+    "ci95": [
+     0.6806,
+     0.8679
+    ]
+   }
+  }
+ }
+]

training/4b/summary.json

@@ -0,0 +1,902 @@
+{
+ "n_tasks": 200,
+ "label": "registered run (docs/prereg/v1-train-4b.md)",
+ "model": "Qwen/Qwen3-4B",
+ "revision": "1cfa9a7208912126459214e8b04321603b3df60c",
+ "bars": {
+  "sampled": {
+   "role": "primary bar and secondary bar",
+   "checkpoint_step": 330,
+   "reward_mean": 0.8041,
+   "ci95": [
+    0.7729,
+    0.8335
+   ],
+   "valid_rate": 1.0,
+   "greedy_rule_ci95": [
+    0.7259,
+    0.7998
+   ],
+   "cis_apart": false,
+   "valid_ge_98": true,
+   "primary_met": false,
+   "paired_minus_greedy": {
+    "mean": 0.04,
+    "ci95": [
+     -0.0018,
+     0.0815
+    ],
+    "win_rate": 0.3
+   },
+   "secondary_met": false
+  },
+  "greedy": {
+   "role": "reported (secondary decoding)",
+   "checkpoint_step": 330,
+   "reward_mean": 0.8078,
+   "ci95": [
+    0.7777,
+    0.8363
+   ],
+   "valid_rate": 1.0,
+   "greedy_rule_ci95": [
+    0.7259,
+    0.7998
+   ],
+   "cis_apart": false,
+   "valid_ge_98": true,
+   "primary_met": false,
+   "paired_minus_greedy": {
+    "mean": 0.0437,
+    "ci95": [
+     0.0025,
+     0.0852
+    ],
+    "win_rate": 0.295
+   },
+   "secondary_met": true
+  }
+ },
+ "rows": [
+  {
+   "model": "Qwen3-4B base (zero-shot)",
+   "decoding": "sampled",
+   "checkpoint_step": 0,
+   "reward_mean": 0.0605,
+   "ci95": [
+    0.0479,
+    0.0734
+   ],
+   "valid_rate": 0.355,
+   "feasible_rate": 0.265,
+   "rate_ok": 0.0,
+   "rate_all_idle": 0.265,
+   "rate_overfill": 0.0,
+   "rate_sell_before_charge": 0.09,
+   "rate_end_short": 0.0,
+   "rate_rate": 0.0,
+   "rate_cycle_cap": 0.0,
+   "rate_malformed": 0.165,
+   "rate_truncated": 0.48,
+   "completion_len_mean": 403.5,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.0679,
+     "ci95": [
+      0.0429,
+      0.0929
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.0546,
+     "ci95": [
+      0.0349,
+      0.0761
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.0606,
+     "ci95": [
+      0.0411,
+      0.0822
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.0571,
+     "ci95": [
+      0.0381,
+      0.0762
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.0569,
+     "ci95": [
+      0.03,
+      0.085
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.0661,
+     "ci95": [
+      0.0451,
+      0.0872
+     ]
+    }
+   }
+  },
+  {
+   "model": "Qwen3-4B base (zero-shot)",
+   "decoding": "greedy",
+   "checkpoint_step": 0,
+   "reward_mean": 0.029,
+   "ci95": [
+    0.02,
+    0.039
+   ],
+   "valid_rate": 0.155,
+   "feasible_rate": 0.135,
+   "rate_ok": 0.0,
+   "rate_all_idle": 0.135,
+   "rate_overfill": 0.0,
+   "rate_sell_before_charge": 0.02,
+   "rate_end_short": 0.0,
+   "rate_rate": 0.0,
+   "rate_cycle_cap": 0.0,
+   "rate_malformed": 0.055,
+   "rate_truncated": 0.79,
+   "completion_len_mean": 461.0,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.025,
+     "ci95": [
+      0.0071,
+      0.0429
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.0225,
+     "ci95": [
+      0.0085,
+      0.0366
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.0384,
+     "ci95": [
+      0.0219,
+      0.0575
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.0214,
+     "ci95": [
+      0.0095,
+      0.0357
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.02,
+     "ci95": [
+      0.005,
+      0.04
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.0421,
+     "ci95": [
+      0.0237,
+      0.0605
+     ]
+    }
+   }
+  },
+  {
+   "model": "4B SFT-1500 (step 0)",
+   "decoding": "sampled",
+   "checkpoint_step": 0,
+   "reward_mean": 0.6661,
+   "ci95": [
+    0.6199,
+    0.712
+   ],
+   "valid_rate": 0.99,
+   "feasible_rate": 0.7,
+   "rate_ok": 0.645,
+   "rate_all_idle": 0.0,
+   "rate_overfill": 0.14,
+   "rate_sell_before_charge": 0.055,
+   "rate_end_short": 0.055,
+   "rate_rate": 0.03,
+   "rate_cycle_cap": 0.065,
+   "rate_malformed": 0.005,
+   "rate_truncated": 0.005,
+   "completion_len_mean": 260.4,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.6882,
+     "ci95": [
+      0.5863,
+      0.785
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.6256,
+     "ci95": [
+      0.5514,
+      0.6996
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.6886,
+     "ci95": [
+      0.612,
+      0.7608
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.6409,
+     "ci95": [
+      0.5614,
+      0.7158
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.5674,
+     "ci95": [
+      0.4482,
+      0.6807
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.7459,
+     "ci95": [
+      0.6865,
+      0.8004
+     ]
+    }
+   }
+  },
+  {
+   "model": "4B SFT-1500 (step 0)",
+   "decoding": "greedy",
+   "checkpoint_step": 0,
+   "reward_mean": 0.6473,
+   "ci95": [
+    0.599,
+    0.6947
+   ],
+   "valid_rate": 1.0,
+   "feasible_rate": 0.66,
+   "rate_ok": 0.615,
+   "rate_all_idle": 0.0,
+   "rate_overfill": 0.18,
+   "rate_sell_before_charge": 0.075,
+   "rate_end_short": 0.045,
+   "rate_rate": 0.04,
+   "rate_cycle_cap": 0.045,
+   "rate_malformed": 0.0,
+   "rate_truncated": 0.0,
+   "completion_len_mean": 260.0,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.6612,
+     "ci95": [
+      0.5558,
+      0.7601
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.6108,
+     "ci95": [
+      0.5289,
+      0.6908
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.672,
+     "ci95": [
+      0.5976,
+      0.743
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.6384,
+     "ci95": [
+      0.5613,
+      0.7119
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.543,
+     "ci95": [
+      0.4299,
+      0.6525
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.712,
+     "ci95": [
+      0.6397,
+      0.7807
+     ]
+    }
+   }
+  },
+  {
+   "model": "4B SFT + GRPO (final)",
+   "decoding": "sampled",
+   "checkpoint_step": 330,
+   "reward_mean": 0.8041,
+   "ci95": [
+    0.7729,
+    0.8335
+   ],
+   "valid_rate": 1.0,
+   "feasible_rate": 0.905,
+   "rate_ok": 0.86,
+   "rate_all_idle": 0.0,
+   "rate_overfill": 0.025,
+   "rate_sell_before_charge": 0.04,
+   "rate_end_short": 0.045,
+   "rate_rate": 0.015,
+   "rate_cycle_cap": 0.015,
+   "rate_malformed": 0.0,
+   "rate_truncated": 0.0,
+   "completion_len_mean": 245.2,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.8256,
+     "ci95": [
+      0.7714,
+      0.8735
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.8033,
+     "ci95": [
+      0.7492,
+      0.85
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.7883,
+     "ci95": [
+      0.7319,
+      0.8398
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.8745,
+     "ci95": [
+      0.8492,
+      0.8988
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.6852,
+     "ci95": [
+      0.5935,
+      0.7716
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.7888,
+     "ci95": [
+      0.7349,
+      0.8373
+     ]
+    }
+   }
+  },
+  {
+   "model": "4B SFT + GRPO (final)",
+   "decoding": "greedy",
+   "checkpoint_step": 330,
+   "reward_mean": 0.8078,
+   "ci95": [
+    0.7777,
+    0.8363
+   ],
+   "valid_rate": 1.0,
+   "feasible_rate": 0.91,
+   "rate_ok": 0.87,
+   "rate_all_idle": 0.0,
+   "rate_overfill": 0.02,
+   "rate_sell_before_charge": 0.03,
+   "rate_end_short": 0.04,
+   "rate_rate": 0.02,
+   "rate_cycle_cap": 0.02,
+   "rate_malformed": 0.0,
+   "rate_truncated": 0.0,
+   "completion_len_mean": 245.7,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.8195,
+     "ci95": [
+      0.7649,
+      0.8687
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.8178,
+     "ci95": [
+      0.7742,
+      0.8562
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.7891,
+     "ci95": [
+      0.729,
+      0.8428
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.8707,
+     "ci95": [
+      0.8441,
+      0.8962
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.6876,
+     "ci95": [
+      0.6032,
+      0.7691
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.8016,
+     "ci95": [
+      0.7476,
+      0.8501
+     ]
+    }
+   }
+  },
+  {
+   "model": "1.7B arm 2c (step 330)",
+   "decoding": "sampled",
+   "checkpoint_step": 330,
+   "reward_mean": 0.7197,
+   "ci95": [
+    0.6811,
+    0.7557
+   ],
+   "valid_rate": 1.0,
+   "feasible_rate": 0.925,
+   "rate_ok": 0.835,
+   "rate_all_idle": 0.0,
+   "rate_overfill": 0.03,
+   "rate_sell_before_charge": 0.025,
+   "rate_end_short": 0.09,
+   "rate_rate": 0.02,
+   "rate_cycle_cap": 0.0,
+   "rate_malformed": 0.0,
+   "rate_truncated": 0.0,
+   "completion_len_mean": 248.3,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.7344,
+     "ci95": [
+      0.6667,
+      0.7989
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.7535,
+     "ci95": [
+      0.6999,
+      0.801
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.6755,
+     "ci95": [
+      0.6048,
+      0.7437
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.7639,
+     "ci95": [
+      0.7111,
+      0.8131
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.6255,
+     "ci95": [
+      0.5277,
+      0.7163
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.7204,
+     "ci95": [
+      0.6616,
+      0.7751
+     ]
+    }
+   }
+  },
+  {
+   "model": "1.7B arm 2c (step 330)",
+   "decoding": "greedy",
+   "checkpoint_step": 330,
+   "reward_mean": 0.7194,
+   "ci95": [
+    0.6813,
+    0.7574
+   ],
+   "valid_rate": 0.99,
+   "feasible_rate": 0.915,
+   "rate_ok": 0.825,
+   "rate_all_idle": 0.0,
+   "rate_overfill": 0.015,
+   "rate_sell_before_charge": 0.035,
+   "rate_end_short": 0.09,
+   "rate_rate": 0.025,
+   "rate_cycle_cap": 0.0,
+   "rate_malformed": 0.005,
+   "rate_truncated": 0.005,
+   "completion_len_mean": 252.5,
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.7534,
+     "ci95": [
+      0.6895,
+      0.8133
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.7308,
+     "ci95": [
+      0.6693,
+      0.7836
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.6823,
+     "ci95": [
+      0.6088,
+      0.7517
+     ]
+    }
+   },
+   "by_curriculum": {
+    "C0": {
+     "n": 84,
+     "reward_mean": 0.7612,
+     "ci95": [
+      0.7055,
+      0.8124
+     ]
+    },
+    "C1": {
+     "n": 40,
+     "reward_mean": 0.6165,
+     "ci95": [
+      0.5183,
+      0.7093
+     ]
+    },
+    "C2": {
+     "n": 76,
+     "reward_mean": 0.7274,
+     "ci95": [
+      0.6677,
+      0.7834
+     ]
+    }
+   }
+  },
+  {
+   "model": "rule:greedy",
+   "reward_mean": 0.764,
+   "ci95": [
+    0.7259,
+    0.7998
+   ],
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.7805,
+     "ci95": [
+      0.7049,
+      0.8502
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.7963,
+     "ci95": [
+      0.7421,
+      0.8451
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.72,
+     "ci95": [
+      0.6527,
+      0.7841
+     ]
+    }
+   }
+  },
+  {
+   "model": "rule:threshold",
+   "reward_mean": 0.6371,
+   "ci95": [
+    0.6008,
+    0.6726
+   ],
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.59,
+     "ci95": [
+      0.5087,
+      0.6689
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.6956,
+     "ci95": [
+      0.6427,
+      0.7484
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.6164,
+     "ci95": [
+      0.5617,
+      0.6704
+     ]
+    }
+   }
+  },
+  {
+   "model": "rule:clock",
+   "reward_mean": 0.4979,
+   "ci95": [
+    0.4598,
+    0.5355
+   ],
+   "by_start": {
+    "empty": {
+     "n": 56,
+     "reward_mean": 0.4649,
+     "ci95": [
+      0.3917,
+      0.5372
+     ]
+    },
+    "half": {
+     "n": 71,
+     "reward_mean": 0.5267,
+     "ci95": [
+      0.4613,
+      0.5924
+     ]
+    },
+    "full": {
+     "n": 73,
+     "reward_mean": 0.4952,
+     "ci95": [
+      0.4354,
+      0.5568
+     ]
+    }
+   }
+  }
+ ],
+ "paired": {
+  "trained_sampled_minus_greedy": {
+   "mean": 0.04,
+   "ci95": [
+    -0.0018,
+    0.0815
+   ],
+   "win_rate": 0.3
+  },
+  "trained_sampled_minus_threshold": {
+   "mean": 0.1669,
+   "ci95": [
+    0.1189,
+    0.2148
+   ],
+   "win_rate": 0.685
+  },
+  "trained_sampled_minus_arm2c_1.7b": {
+   "mean": 0.0844,
+   "ci95": [
+    0.0516,
+    0.1179
+   ],
+   "win_rate": 0.645
+  },
+  "trained_sampled_minus_sft": {
+   "mean": 0.1379,
+   "ci95": [
+    0.0917,
+    0.1853
+   ],
+   "win_rate": 0.53
+  },
+  "trained_sampled_minus_base": {
+   "mean": 0.7435,
+   "ci95": [
+    0.71,
+    0.7759
+   ],
+   "win_rate": 0.98
+  },
+  "trained_greedy_minus_greedy": {
+   "mean": 0.0437,
+   "ci95": [
+    0.0025,
+    0.0852
+   ],
+   "win_rate": 0.295
+  },
+  "trained_greedy_minus_threshold": {
+   "mean": 0.1706,
+   "ci95": [
+    0.1234,
+    0.2175
+   ],
+   "win_rate": 0.695
+  },
+  "trained_greedy_minus_arm2c_1.7b": {
+   "mean": 0.0884,
+   "ci95": [
+    0.0549,
+    0.1221
+   ],
+   "win_rate": 0.635
+  },
+  "trained_greedy_minus_sft": {
+   "mean": 0.1605,
+   "ci95": [
+    0.1147,
+    0.2074
+   ],
+   "win_rate": 0.51
+  },
+  "trained_greedy_minus_base": {
+   "mean": 0.7788,
+   "ci95": [
+    0.7455,
+    0.8108
+   ],
+   "win_rate": 0.98
+  },
+  "sft_sampled_minus_greedy": {
+   "mean": -0.0979,
+   "ci95": [
+    -0.155,
+    -0.0401
+   ],
+   "win_rate": 0.36
+  },
+  "sft_sampled_minus_threshold": {
+   "mean": 0.029,
+   "ci95": [
+    -0.0326,
+    0.0905
+   ],
+   "win_rate": 0.555
+  },
+  "sft_sampled_minus_arm2c_1.7b": {
+   "mean": -0.0536,
+   "ci95": [
+    -0.1031,
+    -0.0053
+   ],
+   "win_rate": 0.515
+  },
+  "sft_sampled_minus_base": {
+   "mean": 0.6056,
+   "ci95": [
+    0.5559,
+    0.6547
+   ],
+   "win_rate": 0.895
+  },
+  "sft_greedy_minus_greedy": {
+   "mean": -0.1168,
+   "ci95": [
+    -0.1792,
+    -0.0557
+   ],
+   "win_rate": 0.38
+  },
+  "sft_greedy_minus_threshold": {
+   "mean": 0.0101,
+   "ci95": [
+    -0.0552,
+    0.0754
+   ],
+   "win_rate": 0.525
+  },
+  "sft_greedy_minus_arm2c_1.7b": {
+   "mean": -0.0721,
+   "ci95": [
+    -0.122,
+    -0.0232
+   ],
+   "win_rate": 0.465
+  },
+  "sft_greedy_minus_base": {
+   "mean": 0.6183,
+   "ci95": [
+    0.569,
+    0.6663
+   ],
+   "win_rate": 0.905
+  }
+ }
+}

training/arms.json

@@ -3,7 +3,8 @@
   "arm": "arm1",
   "label": "arm 1: SFT-200 warm start, then GRPO 280 steps",
   "recipe": "arm 1: SFT-200 warm start, then GRPO 280 steps",
-  "pre_registered": true,
+  "pre_registered": false,
+  "registration": "bars set in the training plan, but the learning rate changed in the run: exploratory",
   "model": "Qwen/Qwen3-1.7B",
   "tasks": 200,
   "split": "eval_2025",
@@ -150,6 +151,7 @@
   "label": "arm 2 (exploratory, not pre-registered)",
   "recipe": "arm 2: SoC-aware SFT-1500 warm start, then GRPO 75 steps",
   "pre_registered": false,
+  "registration": "none: exploratory",
   "model": "Qwen/Qwen3-1.7B",
   "tasks": 200,
   "split": "eval_2025",
@@ -302,6 +304,7 @@
   "label": "arm 2b (exploratory continuation)",
   "recipe": "arm 2b: arm 2, then 105 more GRPO steps from its step 75",
   "pre_registered": false,
+  "registration": "none: exploratory",
   "model": "Qwen/Qwen3-1.7B",
   "tasks": 200,
   "split": "eval_2025",
@@ -443,6 +446,7 @@
   "label": "arm 2c (exploratory continuation)",
   "recipe": "arm 2c: arm 2b, then 150 more GRPO steps from its step 180",
   "pre_registered": false,
+  "registration": "none: exploratory",
   "model": "Qwen/Qwen3-1.7B",
   "tasks": 200,
   "split": "eval_2025",
@@ -578,5 +582,281 @@
     "win_rate": 0.215
    }
   }
+ },
+ {
+  "arm": "4b",
+  "label": "registered run (docs/prereg/v1-train-4b.md)",
+  "recipe": "Qwen3-4B, registered run: SoC-aware SFT-1500 warm start, then GRPO 330 steps as one schedule",
+  "pre_registered": true,
+  "registration": "the registered run: plan, schedule and bars committed before any step",
+  "model": "Qwen/Qwen3-4B",
+  "tasks": 200,
+  "split": "eval_2025",
+  "tier": "foresight",
+  "harness": "oneshot (path B)",
+  "base": {
+   "reward_mean": 0.0605,
+   "ci95": [
+    0.0479,
+    0.0734
+   ],
+   "valid_rate": 0.355
+  },
+  "final": {
+   "model": "4B SFT + GRPO (final)",
+   "checkpoint_step": 330,
+   "reward_mean": 0.8041,
+   "ci95": [
+    0.7729,
+    0.8335
+   ],
+   "valid_rate": 1.0,
+   "feasible_rate": 0.905
+  },
+  "rules": {
+   "greedy": {
+    "reward_mean": 0.764,
+    "ci95": [
+     0.7259,
+     0.7998
+    ]
+   },
+   "threshold": {
+    "reward_mean": 0.6371,
+    "ci95": [
+     0.6008,
+     0.6726
+    ]
+   },
+   "clock": {
+    "reward_mean": 0.4979,
+    "ci95": [
+     0.4598,
+     0.5355
+    ]
+   }
+  },
+  "bars": [
+   {
+    "bar": "beats the greedy rule: 95% CIs apart, valid >= 98%",
+    "rule": "greedy",
+    "rule_reward": 0.764,
+    "rule_ci95": [
+     0.7259,
+     0.7998
+    ],
+    "met": false
+   },
+   {
+    "bar": "beats the threshold rule: 95% CIs apart, valid >= 98%",
+    "rule": "threshold",
+    "rule_reward": 0.6371,
+    "rule_ci95": [
+     0.6008,
+     0.6726
+    ],
+    "met": true
+   }
+  ],
+  "paired": {
+   "trained_sampled_minus_greedy": {
+    "mean": 0.04,
+    "ci95": [
+     -0.0018,
+     0.0815
+    ],
+    "win_rate": 0.3
+   },
+   "trained_sampled_minus_threshold": {
+    "mean": 0.1669,
+    "ci95": [
+     0.1189,
+     0.2148
+    ],
+    "win_rate": 0.685
+   },
+   "trained_sampled_minus_arm2c_1.7b": {
+    "mean": 0.0844,
+    "ci95": [
+     0.0516,
+     0.1179
+    ],
+    "win_rate": 0.645
+   },
+   "trained_sampled_minus_sft": {
+    "mean": 0.1379,
+    "ci95": [
+     0.0917,
+     0.1853
+    ],
+    "win_rate": 0.53
+   },
+   "trained_sampled_minus_base": {
+    "mean": 0.7435,
+    "ci95": [
+     0.71,
+     0.7759
+    ],
+    "win_rate": 0.98
+   },
+   "trained_greedy_minus_greedy": {
+    "mean": 0.0437,
+    "ci95": [
+     0.0025,
+     0.0852
+    ],
+    "win_rate": 0.295
+   },
+   "trained_greedy_minus_threshold": {
+    "mean": 0.1706,
+    "ci95": [
+     0.1234,
+     0.2175
+    ],
+    "win_rate": 0.695
+   },
+   "trained_greedy_minus_arm2c_1.7b": {
+    "mean": 0.0884,
+    "ci95": [
+     0.0549,
+     0.1221
+    ],
+    "win_rate": 0.635
+   },
+   "trained_greedy_minus_sft": {
+    "mean": 0.1605,
+    "ci95": [
+     0.1147,
+     0.2074
+    ],
+    "win_rate": 0.51
+   },
+   "trained_greedy_minus_base": {
+    "mean": 0.7788,
+    "ci95": [
+     0.7455,
+     0.8108
+    ],
+    "win_rate": 0.98
+   },
+   "sft_sampled_minus_greedy": {
+    "mean": -0.0979,
+    "ci95": [
+     -0.155,
+     -0.0401
+    ],
+    "win_rate": 0.36
+   },
+   "sft_sampled_minus_threshold": {
+    "mean": 0.029,
+    "ci95": [
+     -0.0326,
+     0.0905
+    ],
+    "win_rate": 0.555
+   },
+   "sft_sampled_minus_arm2c_1.7b": {
+    "mean": -0.0536,
+    "ci95": [
+     -0.1031,
+     -0.0053
+    ],
+    "win_rate": 0.515
+   },
+   "sft_sampled_minus_base": {
+    "mean": 0.6056,
+    "ci95": [
+     0.5559,
+     0.6547
+    ],
+    "win_rate": 0.895
+   },
+   "sft_greedy_minus_greedy": {
+    "mean": -0.1168,
+    "ci95": [
+     -0.1792,
+     -0.0557
+    ],
+    "win_rate": 0.38
+   },
+   "sft_greedy_minus_threshold": {
+    "mean": 0.0101,
+    "ci95": [
+     -0.0552,
+     0.0754
+    ],
+    "win_rate": 0.525
+   },
+   "sft_greedy_minus_arm2c_1.7b": {
+    "mean": -0.0721,
+    "ci95": [
+     -0.122,
+     -0.0232
+    ],
+    "win_rate": 0.465
+   },
+   "sft_greedy_minus_base": {
+    "mean": 0.6183,
+    "ci95": [
+     0.569,
+     0.6663
+    ],
+    "win_rate": 0.905
+   }
+  },
+  "registered_bars": {
+   "sampled": {
+    "role": "primary bar and secondary bar",
+    "checkpoint_step": 330,
+    "reward_mean": 0.8041,
+    "ci95": [
+     0.7729,
+     0.8335
+    ],
+    "valid_rate": 1.0,
+    "greedy_rule_ci95": [
+     0.7259,
+     0.7998
+    ],
+    "cis_apart": false,
+    "valid_ge_98": true,
+    "primary_met": false,
+    "paired_minus_greedy": {
+     "mean": 0.04,
+     "ci95": [
+      -0.0018,
+      0.0815
+     ],
+     "win_rate": 0.3
+    },
+    "secondary_met": false
+   },
+   "greedy": {
+    "role": "reported (secondary decoding)",
+    "checkpoint_step": 330,
+    "reward_mean": 0.8078,
+    "ci95": [
+     0.7777,
+     0.8363
+    ],
+    "valid_rate": 1.0,
+    "greedy_rule_ci95": [
+     0.7259,
+     0.7998
+    ],
+    "cis_apart": false,
+    "valid_ge_98": true,
+    "primary_met": false,
+    "paired_minus_greedy": {
+     "mean": 0.0437,
+     "ci95": [
+      0.0025,
+      0.0852
+     ],
+     "win_rate": 0.295
+    },
+    "secondary_met": true
+   }
+  }
  }
 ]