diff --git a/README.md b/README.md
index 32c182f..f6e06d1 100644
--- a/README.md
+++ b/README.md
@@ -45,3 +45,33 @@ pytest -q
 数据契约、导出命令和服务器 smoke-test 入口见
 [`docs/SFT_BASELINE.md`](docs/SFT_BASELINE.md)；公共 seam 与 RL 接入条件见
 [`docs/RL_FRAMEWORK.md`](docs/RL_FRAMEWORK.md)。
+
+## 文档与产物结构
+
+`docs/` 存放项目文档，`artifacts/generated/` 存放可由脚本重新生成的分析产物。
+
+```text
+docs/                  文档：设计说明、分析结论、阶段报告
+├── design/            设计文档（data_contract / prompt_interface）
+├── phase_reports/     阶段报告
+├── data_alignment.md
+├── prompt_length_analysis.md
+├── SFT_BASELINE.md / RL_FRAMEWORK.md / 新数据集运行说明.md
+
+artifacts/generated/   生成产物：统计结果、对齐报告等可复现输出
+├── alignment/         对齐报告（data_alignment_report.json/.md）
+└── prompt_stats/      统计产物
+    ├── canonical_id_baseline/   旧 canonical-id 统计（baseline）
+    └── choice_id/               choice-protocol 新统计
+```
+
+关键文档：
+
+| 文档 | 内容 |
+|---|---|
+| [`docs/design/data_contract.md`](docs/design/data_contract.md) | canonical category_id 数据契约（registry/corpus/target、身份策略、训练入口规则） |
+| [`docs/design/prompt_interface.md`](docs/design/prompt_interface.md) | prompt-facing choice protocol（choice id ↔ canonical id、共享解码层） |
+| [`docs/prompt_length_analysis.md`](docs/prompt_length_analysis.md) | prompt token 分析（old∽new、Stage1 收益、worst-case 2689/32768） |
+| [`docs/data_alignment.md`](docs/data_alignment.md) | dataset ↔ corpus/standard 对齐人工结论 |
+
+`docs/reports/` 为本地 gitignored 调研/smoke scratch，不入库。
diff --git a/artifacts/data_alignment_report.json b/artifacts/generated/alignment/data_alignment_report.json
similarity index 100%
rename from artifacts/data_alignment_report.json
rename to artifacts/generated/alignment/data_alignment_report.json
diff --git a/artifacts/data_alignment_report.md b/artifacts/generated/alignment/data_alignment_report.md
similarity index 100%
rename from artifacts/data_alignment_report.md
rename to artifacts/generated/alignment/data_alignment_report.md
diff --git a/artifacts/prompt_stats_finance.json b/artifacts/generated/prompt_stats/canonical_id_baseline/prompt_stats_finance.json
similarity index 100%
rename from artifacts/prompt_stats_finance.json
rename to artifacts/generated/prompt_stats/canonical_id_baseline/prompt_stats_finance.json
diff --git a/artifacts/prompt_stats_shougang.json b/artifacts/generated/prompt_stats/canonical_id_baseline/prompt_stats_shougang.json
similarity index 100%
rename from artifacts/prompt_stats_shougang.json
rename to artifacts/generated/prompt_stats/canonical_id_baseline/prompt_stats_shougang.json
diff --git a/artifacts/prompt_token_stats_finance.json b/artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_finance.json
similarity index 100%
rename from artifacts/prompt_token_stats_finance.json
rename to artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_finance.json
diff --git a/artifacts/prompt_token_stats_infra.json b/artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_infra.json
similarity index 100%
rename from artifacts/prompt_token_stats_infra.json
rename to artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_infra.json
diff --git a/artifacts/prompt_token_stats_pers_info.json b/artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_pers_info.json
similarity index 100%
rename from artifacts/prompt_token_stats_pers_info.json
rename to artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_pers_info.json
diff --git a/artifacts/prompt_token_stats_shougang.json b/artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_shougang.json
similarity index 100%
rename from artifacts/prompt_token_stats_shougang.json
rename to artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_shougang.json
diff --git a/artifacts/prompt_token_stats_summary.json b/artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_summary.json
similarity index 100%
rename from artifacts/prompt_token_stats_summary.json
rename to artifacts/generated/prompt_stats/canonical_id_baseline/prompt_token_stats_summary.json
diff --git a/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_finance.json b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_finance.json
new file mode 100644
index 0000000..c36a95a
--- /dev/null
+++ b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_finance.json
@@ -0,0 +1,139 @@
+{
+ "tokenizer": "Qwen/Qwen2.5-7B-Instruct",
+ "splits": {
+  "train": {
+   "stage1": {
+    "tokens": {
+     "rows": 403,
+     "max": 2553,
+     "p95": 2543,
+     "mean": 2535.4
+    },
+    "chars": {
+     "rows": 403,
+     "max": 5048,
+     "p95": 5033,
+     "mean": 5023.1
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 403,
+     "max": 740,
+     "p95": 560,
+     "mean": 494.0
+    },
+    "chars": {
+     "rows": 403,
+     "max": 1579,
+     "p95": 1261,
+     "mean": 1163.3
+    }
+   }
+  },
+  "val": {
+   "stage1": {
+    "tokens": {
+     "rows": 69,
+     "max": 2543,
+     "p95": 2538,
+     "mean": 2534.6
+    },
+    "chars": {
+     "rows": 69,
+     "max": 5031,
+     "p95": 5027,
+     "mean": 5020.7
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 69,
+     "max": 556,
+     "p95": 554,
+     "mean": 453.7
+    },
+    "chars": {
+     "rows": 69,
+     "max": 1254,
+     "p95": 1249,
+     "mean": 1097.1
+    }
+   }
+  },
+  "test": {
+   "stage1": {
+    "tokens": {
+     "rows": 57,
+     "max": 2539,
+     "p95": 2537,
+     "mean": 2535.3
+    },
+    "chars": {
+     "rows": 57,
+     "max": 5030,
+     "p95": 5026,
+     "mean": 5022.3
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 57,
+     "max": 560,
+     "p95": 558,
+     "mean": 495.1
+    },
+    "chars": {
+     "rows": 57,
+     "max": 1258,
+     "p95": 1257,
+     "mean": 1160.8
+    }
+   }
+  }
+ },
+ "prompt_only": {
+  "train": {
+   "stage1": {
+    "rows": 403,
+    "max": 2534,
+    "p95": 2524,
+    "mean": 2515.9
+   },
+   "stage2": {
+    "rows": 403,
+    "max": 730,
+    "p95": 550,
+    "mean": 484.0
+   }
+  },
+  "val": {
+   "stage1": {
+    "rows": 69,
+    "max": 2523,
+    "p95": 2518,
+    "mean": 2514.6
+   },
+   "stage2": {
+    "rows": 69,
+    "max": 546,
+    "p95": 544,
+    "mean": 443.7
+   }
+  },
+  "test": {
+   "stage1": {
+    "rows": 57,
+    "max": 2520,
+    "p95": 2518,
+    "mean": 2515.8
+   },
+   "stage2": {
+    "rows": 57,
+    "max": 550,
+    "p95": 548,
+    "mean": 485.1
+   }
+  }
+ }
+}
\ No newline at end of file
diff --git a/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_infra.json b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_infra.json
new file mode 100644
index 0000000..a5e119c
--- /dev/null
+++ b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_infra.json
@@ -0,0 +1,139 @@
+{
+ "tokenizer": "Qwen/Qwen2.5-7B-Instruct",
+ "splits": {
+  "train": {
+   "stage1": {
+    "tokens": {
+     "rows": 40,
+     "max": 2588,
+     "p95": 2587,
+     "mean": 2581.6
+    },
+    "chars": {
+     "rows": 40,
+     "max": 4978,
+     "p95": 4977,
+     "mean": 4966.6
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 40,
+     "max": 316,
+     "p95": 315,
+     "mean": 294.7
+    },
+    "chars": {
+     "rows": 40,
+     "max": 884,
+     "p95": 883,
+     "mean": 854.8
+    }
+   }
+  },
+  "val": {
+   "stage1": {
+    "tokens": {
+     "rows": 15,
+     "max": 2589,
+     "p95": 2583,
+     "mean": 2579.9
+    },
+    "chars": {
+     "rows": 15,
+     "max": 4973,
+     "p95": 4973,
+     "mean": 4965.1
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 15,
+     "max": 311,
+     "p95": 305,
+     "mean": 301.9
+    },
+    "chars": {
+     "rows": 15,
+     "max": 876,
+     "p95": 876,
+     "mean": 868.1
+    }
+   }
+  },
+  "test": {
+   "stage1": {
+    "tokens": {
+     "rows": 9,
+     "max": 2593,
+     "p95": 2590,
+     "mean": 2588.1
+    },
+    "chars": {
+     "rows": 9,
+     "max": 4982,
+     "p95": 4980,
+     "mean": 4976.1
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 9,
+     "max": 321,
+     "p95": 318,
+     "mean": 316.1
+    },
+    "chars": {
+     "rows": 9,
+     "max": 888,
+     "p95": 886,
+     "mean": 882.1
+    }
+   }
+  }
+ },
+ "prompt_only": {
+  "train": {
+   "stage1": {
+    "rows": 40,
+    "max": 2570,
+    "p95": 2569,
+    "mean": 2562.7
+   },
+   "stage2": {
+    "rows": 40,
+    "max": 306,
+    "p95": 305,
+    "mean": 284.7
+   }
+  },
+  "val": {
+   "stage1": {
+    "rows": 15,
+    "max": 2571,
+    "p95": 2565,
+    "mean": 2561.9
+   },
+   "stage2": {
+    "rows": 15,
+    "max": 301,
+    "p95": 295,
+    "mean": 291.9
+   }
+  },
+  "test": {
+   "stage1": {
+    "rows": 9,
+    "max": 2575,
+    "p95": 2572,
+    "mean": 2570.1
+   },
+   "stage2": {
+    "rows": 9,
+    "max": 311,
+    "p95": 308,
+    "mean": 306.1
+   }
+  }
+ }
+}
\ No newline at end of file
diff --git a/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_pers_info.json b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_pers_info.json
new file mode 100644
index 0000000..34b244a
--- /dev/null
+++ b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_pers_info.json
@@ -0,0 +1,139 @@
+{
+ "tokenizer": "Qwen/Qwen2.5-7B-Instruct",
+ "splits": {
+  "train": {
+   "stage1": {
+    "tokens": {
+     "rows": 140,
+     "max": 266,
+     "p95": 260,
+     "mean": 255.5
+    },
+    "chars": {
+     "rows": 140,
+     "max": 747,
+     "p95": 740,
+     "mean": 730.0
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 140,
+     "max": 202,
+     "p95": 196,
+     "mean": 191.3
+    },
+    "chars": {
+     "rows": 140,
+     "max": 707,
+     "p95": 700,
+     "mean": 689.0
+    }
+   }
+  },
+  "val": {
+   "stage1": {
+    "tokens": {
+     "rows": 18,
+     "max": 260,
+     "p95": 259,
+     "mean": 256.6
+    },
+    "chars": {
+     "rows": 18,
+     "max": 742,
+     "p95": 742,
+     "mean": 735.7
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 18,
+     "max": 196,
+     "p95": 195,
+     "mean": 192.7
+    },
+    "chars": {
+     "rows": 18,
+     "max": 702,
+     "p95": 702,
+     "mean": 695.6
+    }
+   }
+  },
+  "test": {
+   "stage1": {
+    "tokens": {
+     "rows": 18,
+     "max": 259,
+     "p95": 258,
+     "mean": 255.4
+    },
+    "chars": {
+     "rows": 18,
+     "max": 739,
+     "p95": 738,
+     "mean": 732.9
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 18,
+     "max": 196,
+     "p95": 195,
+     "mean": 191.9
+    },
+    "chars": {
+     "rows": 18,
+     "max": 698,
+     "p95": 698,
+     "mean": 692.3
+    }
+   }
+  }
+ },
+ "prompt_only": {
+  "train": {
+   "stage1": {
+    "rows": 140,
+    "max": 248,
+    "p95": 241,
+    "mean": 237.3
+   },
+   "stage2": {
+    "rows": 140,
+    "max": 192,
+    "p95": 186,
+    "mean": 181.3
+   }
+  },
+  "val": {
+   "stage1": {
+    "rows": 18,
+    "max": 242,
+    "p95": 241,
+    "mean": 238.6
+   },
+   "stage2": {
+    "rows": 18,
+    "max": 186,
+    "p95": 185,
+    "mean": 182.7
+   }
+  },
+  "test": {
+   "stage1": {
+    "rows": 18,
+    "max": 241,
+    "p95": 240,
+    "mean": 237.3
+   },
+   "stage2": {
+    "rows": 18,
+    "max": 186,
+    "p95": 185,
+    "mean": 181.9
+   }
+  }
+ }
+}
\ No newline at end of file
diff --git a/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_shougang.json b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_shougang.json
new file mode 100644
index 0000000..f77f4ec
--- /dev/null
+++ b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_shougang.json
@@ -0,0 +1,139 @@
+{
+ "tokenizer": "Qwen/Qwen2.5-7B-Instruct",
+ "splits": {
+  "train": {
+   "stage1": {
+    "tokens": {
+     "rows": 14521,
+     "max": 2672,
+     "p95": 2588,
+     "mean": 2582.5
+    },
+    "chars": {
+     "rows": 14521,
+     "max": 5075,
+     "p95": 4980,
+     "mean": 4968.6
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 14521,
+     "max": 376,
+     "p95": 313,
+     "mean": 293.8
+    },
+    "chars": {
+     "rows": 14521,
+     "max": 959,
+     "p95": 883,
+     "mean": 853.8
+    }
+   }
+  },
+  "val": {
+   "stage1": {
+    "tokens": {
+     "rows": 1926,
+     "max": 2626,
+     "p95": 2590,
+     "mean": 2582.5
+    },
+    "chars": {
+     "rows": 1926,
+     "max": 5016,
+     "p95": 4982,
+     "mean": 4968.7
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 1926,
+     "max": 363,
+     "p95": 316,
+     "mean": 293.3
+    },
+    "chars": {
+     "rows": 1926,
+     "max": 942,
+     "p95": 886,
+     "mean": 853.6
+    }
+   }
+  },
+  "test": {
+   "stage1": {
+    "tokens": {
+     "rows": 1946,
+     "max": 2689,
+     "p95": 2590,
+     "mean": 2582.7
+    },
+    "chars": {
+     "rows": 1946,
+     "max": 5093,
+     "p95": 4981,
+     "mean": 4968.8
+    }
+   },
+   "stage2": {
+    "tokens": {
+     "rows": 1946,
+     "max": 404,
+     "p95": 320,
+     "mean": 297.4
+    },
+    "chars": {
+     "rows": 1946,
+     "max": 981,
+     "p95": 888,
+     "mean": 858.0
+    }
+   }
+  }
+ },
+ "prompt_only": {
+  "train": {
+   "stage1": {
+    "rows": 14521,
+    "max": 2653,
+    "p95": 2569,
+    "mean": 2563.1
+   },
+   "stage2": {
+    "rows": 14521,
+    "max": 366,
+    "p95": 303,
+    "mean": 283.8
+   }
+  },
+  "val": {
+   "stage1": {
+    "rows": 1926,
+    "max": 2607,
+    "p95": 2571,
+    "mean": 2563.2
+   },
+   "stage2": {
+    "rows": 1926,
+    "max": 353,
+    "p95": 306,
+    "mean": 283.3
+   }
+  },
+  "test": {
+   "stage1": {
+    "rows": 1946,
+    "max": 2670,
+    "p95": 2572,
+    "mean": 2563.5
+   },
+   "stage2": {
+    "rows": 1946,
+    "max": 394,
+    "p95": 310,
+    "mean": 287.4
+   }
+  }
+ }
+}
\ No newline at end of file
diff --git a/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_summary.json b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_summary.json
new file mode 100644
index 0000000..52703bb
--- /dev/null
+++ b/artifacts/generated/prompt_stats/choice_id/prompt_token_stats_summary.json
@@ -0,0 +1,473 @@
+{
+ "tokenizer": "Qwen/Qwen2.5-7B-Instruct",
+ "datasets": {
+  "finance": {
+   "splits": {
+    "train": {
+     "stage1": {
+      "tokens": {
+       "rows": 403,
+       "max": 2553,
+       "p95": 2543,
+       "mean": 2535.4
+      },
+      "chars": {
+       "rows": 403,
+       "max": 5048,
+       "p95": 5033,
+       "mean": 5023.1
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 403,
+       "max": 740,
+       "p95": 560,
+       "mean": 494.0
+      },
+      "chars": {
+       "rows": 403,
+       "max": 1579,
+       "p95": 1261,
+       "mean": 1163.3
+      }
+     }
+    },
+    "val": {
+     "stage1": {
+      "tokens": {
+       "rows": 69,
+       "max": 2543,
+       "p95": 2538,
+       "mean": 2534.6
+      },
+      "chars": {
+       "rows": 69,
+       "max": 5031,
+       "p95": 5027,
+       "mean": 5020.7
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 69,
+       "max": 556,
+       "p95": 554,
+       "mean": 453.7
+      },
+      "chars": {
+       "rows": 69,
+       "max": 1254,
+       "p95": 1249,
+       "mean": 1097.1
+      }
+     }
+    },
+    "test": {
+     "stage1": {
+      "tokens": {
+       "rows": 57,
+       "max": 2539,
+       "p95": 2537,
+       "mean": 2535.3
+      },
+      "chars": {
+       "rows": 57,
+       "max": 5030,
+       "p95": 5026,
+       "mean": 5022.3
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 57,
+       "max": 560,
+       "p95": 558,
+       "mean": 495.1
+      },
+      "chars": {
+       "rows": 57,
+       "max": 1258,
+       "p95": 1257,
+       "mean": 1160.8
+      }
+     }
+    }
+   },
+   "vs_legacy_train": {
+    "stage1": {
+     "old_mean": 5940.8,
+     "new_mean": 2535.4,
+     "old_p95": 5950,
+     "new_p95": 2543,
+     "old_max": 5959,
+     "new_max": 2553,
+     "p95_reduction_pct": 57.3,
+     "max_reduction_pct": 57.2
+    },
+    "stage2": {
+     "old_mean": 542.9,
+     "new_mean": 494.0,
+     "old_p95": 603,
+     "new_p95": 560,
+     "old_max": 794,
+     "new_max": 740,
+     "p95_reduction_pct": 7.1,
+     "max_reduction_pct": 6.8
+    }
+   },
+   "train_stage2_gt_position_hist": null
+  },
+  "infra": {
+   "splits": {
+    "train": {
+     "stage1": {
+      "tokens": {
+       "rows": 40,
+       "max": 2588,
+       "p95": 2587,
+       "mean": 2581.6
+      },
+      "chars": {
+       "rows": 40,
+       "max": 4978,
+       "p95": 4977,
+       "mean": 4966.6
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 40,
+       "max": 316,
+       "p95": 315,
+       "mean": 294.7
+      },
+      "chars": {
+       "rows": 40,
+       "max": 884,
+       "p95": 883,
+       "mean": 854.8
+      }
+     }
+    },
+    "val": {
+     "stage1": {
+      "tokens": {
+       "rows": 15,
+       "max": 2589,
+       "p95": 2583,
+       "mean": 2579.9
+      },
+      "chars": {
+       "rows": 15,
+       "max": 4973,
+       "p95": 4973,
+       "mean": 4965.1
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 15,
+       "max": 311,
+       "p95": 305,
+       "mean": 301.9
+      },
+      "chars": {
+       "rows": 15,
+       "max": 876,
+       "p95": 876,
+       "mean": 868.1
+      }
+     }
+    },
+    "test": {
+     "stage1": {
+      "tokens": {
+       "rows": 9,
+       "max": 2593,
+       "p95": 2590,
+       "mean": 2588.1
+      },
+      "chars": {
+       "rows": 9,
+       "max": 4982,
+       "p95": 4980,
+       "mean": 4976.1
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 9,
+       "max": 321,
+       "p95": 318,
+       "mean": 316.1
+      },
+      "chars": {
+       "rows": 9,
+       "max": 888,
+       "p95": 886,
+       "mean": 882.1
+      }
+     }
+    }
+   },
+   "vs_legacy_train": {
+    "stage1": {
+     "old_mean": 5066.6,
+     "new_mean": 2581.6,
+     "old_p95": 5073,
+     "new_p95": 2587,
+     "old_max": 5074,
+     "new_max": 2588,
+     "p95_reduction_pct": 49.0,
+     "max_reduction_pct": 49.0
+    },
+    "stage2": {
+     "old_mean": 326.7,
+     "new_mean": 294.7,
+     "old_p95": 347,
+     "new_p95": 315,
+     "old_max": 348,
+     "new_max": 316,
+     "p95_reduction_pct": 9.2,
+     "max_reduction_pct": 9.2
+    }
+   },
+   "train_stage2_gt_position_hist": null
+  },
+  "pers_info": {
+   "splits": {
+    "train": {
+     "stage1": {
+      "tokens": {
+       "rows": 140,
+       "max": 266,
+       "p95": 260,
+       "mean": 255.5
+      },
+      "chars": {
+       "rows": 140,
+       "max": 747,
+       "p95": 740,
+       "mean": 730.0
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 140,
+       "max": 202,
+       "p95": 196,
+       "mean": 191.3
+      },
+      "chars": {
+       "rows": 140,
+       "max": 707,
+       "p95": 700,
+       "mean": 689.0
+      }
+     }
+    },
+    "val": {
+     "stage1": {
+      "tokens": {
+       "rows": 18,
+       "max": 260,
+       "p95": 259,
+       "mean": 256.6
+      },
+      "chars": {
+       "rows": 18,
+       "max": 742,
+       "p95": 742,
+       "mean": 735.7
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 18,
+       "max": 196,
+       "p95": 195,
+       "mean": 192.7
+      },
+      "chars": {
+       "rows": 18,
+       "max": 702,
+       "p95": 702,
+       "mean": 695.6
+      }
+     }
+    },
+    "test": {
+     "stage1": {
+      "tokens": {
+       "rows": 18,
+       "max": 259,
+       "p95": 258,
+       "mean": 255.4
+      },
+      "chars": {
+       "rows": 18,
+       "max": 739,
+       "p95": 738,
+       "mean": 732.9
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 18,
+       "max": 196,
+       "p95": 195,
+       "mean": 191.9
+      },
+      "chars": {
+       "rows": 18,
+       "max": 698,
+       "p95": 698,
+       "mean": 692.3
+      }
+     }
+    }
+   },
+   "vs_legacy_train": {
+    "stage1": {
+     "old_mean": 512.3,
+     "new_mean": 255.5,
+     "old_p95": 517,
+     "new_p95": 260,
+     "old_max": 523,
+     "new_max": 266,
+     "p95_reduction_pct": 49.7,
+     "max_reduction_pct": 49.1
+    },
+    "stage2": {
+     "old_mean": 225.6,
+     "new_mean": 191.3,
+     "old_p95": 232,
+     "new_p95": 196,
+     "old_max": 236,
+     "new_max": 202,
+     "p95_reduction_pct": 15.5,
+     "max_reduction_pct": 14.4
+    }
+   },
+   "train_stage2_gt_position_hist": null
+  },
+  "shougang": {
+   "splits": {
+    "train": {
+     "stage1": {
+      "tokens": {
+       "rows": 14521,
+       "max": 2672,
+       "p95": 2588,
+       "mean": 2582.5
+      },
+      "chars": {
+       "rows": 14521,
+       "max": 5075,
+       "p95": 4980,
+       "mean": 4968.6
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 14521,
+       "max": 376,
+       "p95": 313,
+       "mean": 293.8
+      },
+      "chars": {
+       "rows": 14521,
+       "max": 959,
+       "p95": 883,
+       "mean": 853.8
+      }
+     }
+    },
+    "val": {
+     "stage1": {
+      "tokens": {
+       "rows": 1926,
+       "max": 2626,
+       "p95": 2590,
+       "mean": 2582.5
+      },
+      "chars": {
+       "rows": 1926,
+       "max": 5016,
+       "p95": 4982,
+       "mean": 4968.7
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 1926,
+       "max": 363,
+       "p95": 316,
+       "mean": 293.3
+      },
+      "chars": {
+       "rows": 1926,
+       "max": 942,
+       "p95": 886,
+       "mean": 853.6
+      }
+     }
+    },
+    "test": {
+     "stage1": {
+      "tokens": {
+       "rows": 1946,
+       "max": 2689,
+       "p95": 2590,
+       "mean": 2582.7
+      },
+      "chars": {
+       "rows": 1946,
+       "max": 5093,
+       "p95": 4981,
+       "mean": 4968.8
+      }
+     },
+     "stage2": {
+      "tokens": {
+       "rows": 1946,
+       "max": 404,
+       "p95": 320,
+       "mean": 297.4
+      },
+      "chars": {
+       "rows": 1946,
+       "max": 981,
+       "p95": 888,
+       "mean": 858.0
+      }
+     }
+    }
+   },
+   "vs_legacy_train": {
+    "stage1": {
+     "old_mean": 5067.1,
+     "new_mean": 2582.5,
+     "old_p95": 5073,
+     "new_p95": 2588,
+     "old_max": 5157,
+     "new_max": 2672,
+     "p95_reduction_pct": 49.0,
+     "max_reduction_pct": 48.2
+    },
+    "stage2": {
+     "old_mean": 326.0,
+     "new_mean": 293.8,
+     "old_p95": 345,
+     "new_p95": 313,
+     "old_max": 408,
+     "new_max": 376,
+     "p95_reduction_pct": 9.3,
+     "max_reduction_pct": 7.8
+    }
+   },
+   "train_stage2_gt_position_hist": null
+  }
+ }
+}
\ No newline at end of file
diff --git a/docs/data_alignment.md b/docs/data_alignment.md
new file mode 100644
index 0000000..34cfc3e
--- /dev/null
+++ b/docs/data_alignment.md
@@ -0,0 +1,67 @@
+# 数据对齐结论（dataset ↔ corpus/standard alignment）
+
+> 人类维护的结论性文档。机器可复现的完整报告（含逐 dataset 覆盖率矩阵与
+> UNKNOWN 明细）见 `artifacts/generated/alignment/data_alignment_report.{json,md}`，
+> 由 `python -m script.analysis.analyze_dataset_corpus_alignment` 再生产。
+> 本文件的内容更新需人工判断，不随脚本自动更新。
+
+## 1. dataset ↔ 对应 corpus/standard
+
+两种覆盖率口径（均来自生成报告，精确匹配/空白归一化）：
+
+- **unique leaf coverage**：数据集唯一 leaf 中能在语料命中的比例（leaf 维度）；
+- **sample leaf_exact coverage**：样本级精确匹配率（样本维度）。
+
+| dataset | 对齐 source | unique leaf coverage | sample leaf_exact coverage | 备注 |
+|---|---|---|---|---|
+| finance | `corpus:finance` / `financial_standards_dict.json` | 20/25（80%） | 534/568（94.0%） | 语料 220 leaf ≫ 标注 25；fs 是 3 段路径（L1-L2-leaf），数据集是 4 级，**不把 join(level_1..4) 当身份** |
+| infra | `standard:guanji_dict.json` | 4/4（100%） | 64/64（100%） | infra ⊂ shougang，共享 code 空间 |
+| pers_info | `standard:education_dict.json` | 4/18（≈22%） | 35/176（19.9%） | 单级分类（仅 level_4）；14/18 leaf 无任何语料定义 |
+| shougang | `standard:guanji_dict.json` | 192/193（≈99.5%） | 18393/19415（94.7%） | 样本口径下未匹配的 1022 条即 `——` 占位符（structural skip）；code 唯一、字母↔L1 100% |
+
+> 注意：finance/shougang 的样本级匹配率（94.0%/94.7%）与 unique leaf
+> 覆盖（80%/≈99.5%）分母不同（样本数 vs leaf 数），不要混读。
+
+训练推荐 leaf level 全部为 **level_4**。
+
+## 2. 现状数字（canonical 解析后，2026-08）
+
+| dataset | 样本 | 完整 4 级路径 | level_4 leaf 数 | canonical resolved | trainable |
+|---|---|---|---|---|---|
+| finance | 568 | 381 | 25 | 531 | 529 |
+| infra | 64 | 64 | 4 | 64 | 64 |
+| pers_info | 176 | 0（单级） | 18 | 176 | 176 |
+| shougang | 19,415 | 19,415 | 193 | 18,393 | 18,393 |
+
+finance 有 2 条 resolved 记录落在原始 split 边界之外（数据管线事实，未静默补入）；
+shougang `——` 占位符 1,022 条（5.3%）作为 structural skip 不进训练。
+
+## 3. 需要记住的 schema 问题（人工视角）
+
+- **finance**：187/568 样本 level_3 为空而 level_4 完整（路径不全，属数据事实）。
+  2026-08-19 已修复 `"经营 管理"`→`"经营管理"` whitespace 变体（6 条），只读守护
+  test 见 `tests/task/test_label_whitespace_variants.py`。
+- **finance 命名噪声**：`交易清金额信息`（疑似错别字 vs `交易清结算信息`）、
+  `基本信息（公开`（截断）等——标注侧定名前不改动。
+- **cross-corpus leaf 冲突**：同名 leaf 出现在多个 standard（如 shougang
+  `供应商管理数据` 在 guanji 与 shanghai_fta；finance `基本信息` 在
+  financial_standards 与 education）——leaf 字符串不能作为身份，必须用
+  canonical category_id。
+- **pers_info**：单级分类 + 语料覆盖低（4/18），target 允许 description 缺失。
+- **corpus:finance**：leaf-only（不带 path/code），历史遗留；正式 corpus 已
+  改为 `CorpusCategory`（category_id/name/description/descriptions/examples）。
+- **数据层级 vs data_level 字段**：data_level 分布与分类深度不符，语义未定
+  （仅作 provenance，不解释为敏感级别）。
+
+## 4. 仍需人工确认（摘要）
+
+finance 5 个 leaf 无语料定义 / guanji code 数字段与 level_2/3 的对应 /
+pers_info 14/18 语料来源 / education code 无层级可对照 / data_level 语义 /
+shougang corpus 1 条 malformed 条目（`"nan"`）。完整清单见报告 §4.9。
+
+## 5. 相关文档
+
+- 身份策略与训练规则：`docs/design/data_contract.md`
+- 历史设计记录：`docs/phase_reports/stage2_contract_design.md`、
+  `docs/phase_reports/stage3c_report.md`
+- 数据维护 SOP：`docs/新数据集运行说明.md`
diff --git a/docs/design/data_contract.md b/docs/design/data_contract.md
new file mode 100644
index 0000000..a7ac65f
--- /dev/null
+++ b/docs/design/data_contract.md
@@ -0,0 +1,96 @@
+# Canonical Data Contract
+
+Status: **frozen** (Stage 3C/3B). The canonical identity is the single source of
+truth for registry, corpus, ground truth, SFT labels, evaluation and reward.
+
+## 1. One identity: canonical `category_id`
+
+- **`target.category_id` is the only canonical semantic ground-truth identity**
+  across SFT and RL. It is unique within one dataset's `LeafRegistry`.
+  Model-facing supervision/actions use derived choice ids (see
+  `docs/design/prompt_interface.md`) — the model never generates canonical ids.
+- `classification.level_1..level_4` are **provenance only** — never read as a
+  label, never fallback. There is deliberately **no `level_4` fallback**.
+- Prompt-facing **choice ids are a derived, decoupled view** (see
+  `docs/design/prompt_interface.md`); they never leak back into canonical
+  records or reward semantics.
+
+## 2. Canonical sample schema (`data/<dataset>/canonical/all.json`)
+
+Input sample keeps the original fields (`classification`, `metadata`,
+`data_level`, `label_status`, …) untouched as provenance. The resolver
+appends:
+
+```jsonc
+"resolution_status": "resolved",            // resolved | missing_leaf | path_mismatch | ...
+"target": {
+  "leaf_level": "level_4",
+  "leaf_name": "合同归并",                    // display name; not guaranteed unique
+  "category_id": "B1-2",                     // the ONLY training identity (registry-unique)
+  "category_path": ["生产数据域", "生产合同（订单）", "合同归并"]   // provenance, not an ID
+}
+```
+
+## 3. Canonical corpus schema (`cfg/task/corpus/<dataset>.corpus.json`)
+
+```jsonc
+{ "categories": [ {
+  "category_id": "A1-1-1",
+  "name": "科研设备预约管理",
+  "description": "…",            // primary description; may be empty
+  "descriptions": ["…"],          // extra description documents (semantically distinct from examples)
+  "path": ["研发数据域", "科研设备管理", "科研设备预约管理"],   // may be empty
+  "code": "A1-1-1",               // may be empty; opaque stable id, digit groups are NOT level_2/3
+  "examples": ["设备预约信息", "审核表"]
+} ] }
+```
+
+- A category may have multiple descriptions/examples or none.
+- **Production invariant**: the corpus must cover every registry category so
+  Stage 2 can always resolve candidates by `category_id` (no registry fallback,
+  validated by `require_corpus_covers_registry`).
+
+## 4. Leaf registry (`cfg/task/registry/<dataset>.registry.json`)
+
+- The `LeafRegistry` is the **complete leaf universe** of the classification
+  standard/corpus, built from the corpus — **never derived from training
+  samples** (samples may cover only a subset).
+- Per-dataset registry sizes: finance **233** / infra **233** / shougang **233**
+  / pers_info **18**; corpus sizes match.
+
+## 5. Per-dataset identity strategy
+
+| dataset | strategy | category_id shape |
+|---|---|---|
+| shougang | code (guanji) | `A1-1-1`, `B1-2`; letter↔level_1 100% (A/B/C); digits are opaque ordinals |
+| infra | code, `registry_source=shougang` | shares shougang codes (infra ⊂ shougang) |
+| finance | path (L1/L2/L4) | `finance:业务.账户信息.基本信息`; level_3 is provenance, not part of identity |
+| pers_info | path (level_4 only) | `pers_info:学籍管理信息`; single-level labels |
+
+## 6. Training entry rule (SFT + RL)
+
+Only records with `resolution_status == "resolved"` **and** `target.category_id ∈
+LeafRegistry` **and** `target.leaf_name == registry.get(category_id).name`
+enter training; violations fail fast (no silent skip). Split membership follows
+the original `data/<dataset>/{train,val,test}.json` by record id. `canonical`
+resolved / trainable counts (trainable = resolved within split boundaries):
+
+| dataset | canonical resolved | trainable | outside splits |
+|---|---|---|---|
+| finance | 531 | 529 | 2 |
+| infra | 64 | 64 | 0 |
+| pers_info | 176 | 176 | 0 |
+| shougang | 18,393 | 18,393 | 0 |
+
+## 7. Consumers & files
+
+- Contracts: `src/agent/task/contracts.py` (`LeafCategory`, `CorpusCategory`,
+  `SampleTarget`, `LeafRegistry`, `TaskConfig`)
+- Resolver: `src/agent/task/resolver.py`, identity: `src/agent/task/identity.py`,
+  dataset config: `src/agent/task/dataset_config.py`
+- Shared training helpers: `src/agent/training/common.py`
+  (`canonical_target`, `build_candidates`, `require_corpus*`)
+- Data-side alignment facts: `docs/data_alignment.md`; generated report:
+  `artifacts/generated/alignment/data_alignment_report.{json,md}`
+- Historical design records: `docs/phase_reports/stage2_contract_design.md`,
+  `docs/phase_reports/stage3c_report.md`
diff --git a/docs/design/prompt_interface.md b/docs/design/prompt_interface.md
new file mode 100644
index 0000000..feb7ce6
--- /dev/null
+++ b/docs/design/prompt_interface.md
@@ -0,0 +1,61 @@
+# Prompt Interface (choice protocol)
+
+Status: **frozen** (Phase 6 / PR #8, merged on master). The model never sees or
+generates canonical `category_id`; it speaks a compact **choice protocol** that
+is decoded to canonical ids at the LLM boundary.
+
+## 1. Why a choice protocol
+
+- Canonical ids are internal identity (`finance:业务.账户信息.基本信息`) — long,
+  verbose, and **must not become model actions** (generated ids would be
+  brittle and leak internal taxonomy structure).
+- Choice ids are compact (`"1".."N"`), deterministic, and provide a **compact
+  action representation** (they do not reduce the number of categories). This
+  is a **decoupling/identity** decision, not only a token reduction one (see
+  `docs/prompt_length_analysis.md` §6).
+
+## 2. Mapping (`src/agent/task/prompt_choices.py`)
+
+`PromptChoiceRegistry` maps canonical `category_id` ↔ global choice id
+(`"1".."N"` following `LeafRegistry.categories` order, stable across runs)
+plus a **display name** = shortest unique path suffix (leaf name when unique,
+parent-qualified otherwise). Choice ids exist only inside prompts/model outputs;
+they are decoded back immediately and never written to canonical records.
+
+## 3. Stage 1 (retrieve 5 candidates over the full registry)
+
+- Prompt: catalog of `[choice_id, display_name]` pairs for the whole registry.
+- Model output: `{"candidates": ["3", "7", "12", "19", "24"]}` (global choice ids).
+- Checks (shared, never raise on model output): JSON/schema → exactly 5 →
+  unique → choice_id in catalog → decode to canonical `category_id` tuple.
+
+## 4. Stage 2 (pick one answer from the 5-candidate bundle)
+
+- Prompt: candidate bundle with **local** ids `"1".."5"` in candidate order +
+  display name + corpus description/examples.
+- Model output: `{"answer": "2"}` (local bundle id).
+- Checks (shared): JSON/schema → answer ∈ `"1".."5"` (strict, no coercion) →
+  positional decode against the canonical candidates.
+
+## 5. Shared decode layer (Phase 7 / PR #9)
+
+`agent.task.parser.check_stage1_choices` / `check_stage2_choices` →
+`ChoiceParseResult` (format/constraint validity + decoded canonical ids +
+`canonical_view()`). **Both evaluation and reward consume this single
+implementation** — no second copy of choice validation:
+
+- `agent.evaluation.evaluate_stage1_choices` / `evaluate_stage2_choices`
+- `agent.training.rl.reward_stage1_choices` / `reward_stage2_choices` /
+  `reward_for_choice_result` (decode first, then the unchanged reward table)
+
+Stage-2 answer — `docs/design/data_contract.md` §6 — ground truth is NOT
+required to be among the candidates (a Stage-1 recall miss is a legal state).
+
+## 6. Files
+
+- `src/agent/task/prompt_choices.py` — registry + encode/decode primitives
+- `src/agent/task/parser.py` — shared never-raise choice-aware checks
+- `src/agent/task/prompts.py` — prompt builders (`build_stage1_prompt`,
+  `build_stage2_prompt`, `stage1_answer`, `stage2_answer`)
+- Exports/layout facts: `docs/SFT_BASELINE.md`, `docs/RL_FRAMEWORK.md`;
+  historical design: `docs/phase_reports/stage2_contract_design.md`
diff --git a/artifacts/stage2_contract_design.md b/docs/phase_reports/stage2_contract_design.md
similarity index 98%
rename from artifacts/stage2_contract_design.md
rename to docs/phase_reports/stage2_contract_design.md
index aca7b71..056c029 100644
--- a/artifacts/stage2_contract_design.md
+++ b/docs/phase_reports/stage2_contract_design.md
@@ -1,6 +1,6 @@
 # 阶段 2：最小统一数据 contract 设计
 
-> 依据：`artifacts/data_alignment_report.md/.json`（阶段 1）。
+> 依据：`artifacts/generated/alignment/data_alignment_report.md/.json`（阶段 1）。
 > 原则：不自动修复报告中的 UNKNOWN 数据问题；`classification.level_1~level_4`
 > 原样保留作为 provenance；训练 ground truth 最终只依赖 `target.category_id`。
 
diff --git a/artifacts/stage3c_report.md b/docs/phase_reports/stage3c_report.md
similarity index 96%
rename from artifacts/stage3c_report.md
rename to docs/phase_reports/stage3c_report.md
index 9b33e2a..72be8a8 100644
--- a/artifacts/stage3c_report.md
+++ b/docs/phase_reports/stage3c_report.md
@@ -13,7 +13,7 @@
 - `script/verl/sft/prompt_stats.py` — prompt 长度统计（字符级 + 可选 tokenizer）
 - `tests/sft/test_sft_canonical_e2e.py` — 真实数据端到端（CI 无 data/ 时 skip）
 - `tests/sft/test_prompt_stats.py`、`tests/sft/fixtures/canonical/`
-- `artifacts/prompt_stats_{finance,shougang}.json`、`artifacts/prompt_token_stats_{finance,shougang}.json`
+- `artifacts/generated/prompt_stats/legacy/prompt_stats_{finance,shougang}.json`、`artifacts/generated/prompt_stats/legacy/prompt_token_stats_{finance,shougang}.json`
 
 ## 2. 新 SFT 数据流
 
diff --git a/artifacts/stage4a_report.md b/docs/phase_reports/stage4a_report.md
similarity index 99%
rename from artifacts/stage4a_report.md
rename to docs/phase_reports/stage4a_report.md
index 5940330..5c07104 100644
--- a/artifacts/stage4a_report.md
+++ b/docs/phase_reports/stage4a_report.md
@@ -20,7 +20,7 @@
 - `src/agent/training/rl/__init__.py` / `sample.py` / `reward.py` / `dataset.py` — RL 数据 / reward / VeRL parquet 契约
 - `script/verl/rl/__init__.py` / `export.py` / `validate.py` — CLI 适配器
 - `tests/rl/test_rl_parser.py` / `test_rl_reward.py` / `test_rl_dataset.py` / `test_rl_canonical_e2e.py` / `test_rl_verl_compat.py`
-- `artifacts/stage4a_report.md`（本文件）
+- `docs/phase_reports/stage4a_report.md`（本文件）
 
 ## 2. RL sample schema
 
diff --git a/docs/prompt_length_analysis.md b/docs/prompt_length_analysis.md
new file mode 100644
index 0000000..c5e01a0
--- /dev/null
+++ b/docs/prompt_length_analysis.md
@@ -0,0 +1,83 @@
+# Prompt Length Analysis (Phase 9)
+
+Recomputed on the **Phase 8 real SFT/RL prompts** (choice protocol) against the
+legacy canonical-id baseline, using the `Qwen/Qwen2.5-7B-Instruct` tokenizer
+(snapshot `a09a3545`, tokenizer.json only). Measurement = token count of the
+full chat-template-rendered conversation (identical semantics to the legacy
+`prompt_stats.py` baseline; p95 uses the legacy definition
+`sorted[int(0.95·n)-1]`). Old stats are kept untouched as the legacy baseline.
+
+Generated stats: `artifacts/generated/prompt_stats/` (canonical_id_baseline/ =
+old canonical-id stats, choice_id/ = new choice-protocol stats).
+
+## 1. Old vs new p95 / max (train split, tokens)
+
+### p95
+| dataset | stage | old p95 | new p95 | reduction % |
+|---|---|---|---|---|
+| finance | stage1 | 5950 | 2543 | **57.3%** |
+| finance | stage2 | 603 | 560 | 7.1% |
+| infra | stage1 | 5073 | 2587 | 49.0% |
+| infra | stage2 | 347 | 315 | 9.2% |
+| pers_info | stage1 | 517 | 260 | 49.7% |
+| pers_info | stage2 | 232 | 196 | 15.5% |
+| shougang | stage1 | 5073 | 2588 | 49.0% |
+| shougang | stage2 | 345 | 313 | 9.3% |
+
+### max
+| dataset | stage | old max | new max | reduction % |
+|---|---|---|---|---|
+| finance | stage1 | 5959 | 2553 | **57.2%** |
+| finance | stage2 | 794 | 740 | 6.8% |
+| infra | stage1 | 5074 | 2588 | 49.0% |
+| infra | stage2 | 348 | 316 | 9.2% |
+| pers_info | stage1 | 523 | 266 | 49.1% |
+| pers_info | stage2 | 236 | 202 | 14.4% |
+| shougang | stage1 | 5157 | 2672 | 48.2% |
+| shougang | stage2 | 408 | 376 | 7.8% |
+
+## 2. Stage 1 is the main token-saving source
+
+Per-record combined mean (train): finance 6483.7 → 3029.4 tokens (**−53.3%**,
+abs. saving ≈ 3454); infra/shougang 5393 → 2876 (**−46.7%**); pers_info 737.9 →
+446.8 (−39.5%). ~99% of the absolute saving comes from **Stage 1**
+(finance/infra/shougang ≈ 98–99%), where the full registry (finance/infra/
+shougang 233 categories; pers_info 18) used to be rendered with verbose
+canonical ids and is now a compact `[choice_id, display_name]` catalog. Stage 2 (the 5-candidate bundle) was
+already compact, so it gains only 7–15%.
+
+## 3. finance vs shougang: which is longest now — and a reversal
+
+- **shougang** is now the longest (global max 2689, test stage1; prompt-only
+  2670) vs finance (max 2553).
+- This **reverses the old baseline** (finance 5959 > shougang 5157). Reasons:
+  finance's old ids were long (`finance:客户.个人….`) so choice ids save 57%;
+  shougang's old ids were short (`A1-1-1`) so they save 49%. The remaining
+  longest rows are driven by **long metadata text** in the data (shougang's
+  longest row carries a 17-item enumerated `field_description`; finance's
+  longest row has tiny metadata), not by the registry.
+
+## 4. Worst case vs context
+
+Global worst case = **2689 tokens** (shougang test, Stage 1) ≈ **8.2% of the
+32768 context** (prompt-only 2670 ≈ 8.15%). The new prompts are safely below
+the context limit with a large margin — the old worst case was ~6000 tokens
+(~18%), so the choice protocol roughly halves the headroom consumed.
+
+## 5. Baselines & reproducibility
+
+`artifacts/generated/prompt_stats/canonical_id_baseline/` keeps the old
+canonical-id stats (`prompt_token_stats_*.json`, `prompt_stats_*.json`,
+`prompt_token_stats_summary.json`) unchanged for regressions and comparison;
+`artifacts/generated/prompt_stats/choice_id/` holds the new choice-protocol
+stats. Regenerate (reproduces the tables above) with
+`python -m script.verl.sft.prompt_stats --dataset-dir data/sft/<ds> --model
+Qwen/Qwen2.5-7B-Instruct --report <out>.json`.
+
+## 6. Note: PromptChoice is not only about length
+
+The choice protocol exists primarily to **decouple the internal canonical
+`category_id` from model actions** — ids must never be generated as actions
+(fragile, leaks taxonomy). Token savings (≈50% Stage 1) are a welcome side
+effect. See `docs/design/prompt_interface.md` and
+`docs/design/data_contract.md`.
diff --git a/script/analysis/analyze_dataset_corpus_alignment.py b/script/analysis/analyze_dataset_corpus_alignment.py
index 22ce188..16d591d 100644
--- a/script/analysis/analyze_dataset_corpus_alignment.py
+++ b/script/analysis/analyze_dataset_corpus_alignment.py
@@ -7,8 +7,8 @@
   - standards:    data/knowledge/standards_map/*.json
 
 Writes:
-  - artifacts/data_alignment_report.json
-  - artifacts/data_alignment_report.md
+  - artifacts/generated/alignment/data_alignment_report.json
+  - artifacts/generated/alignment/data_alignment_report.md
 
 The script never modifies data/ or src/.
 """
@@ -25,7 +25,7 @@
 
 PROJECT_ROOT = Path(__file__).resolve().parents[2]
 DEFAULT_DATA_DIR = PROJECT_ROOT / "data"
-DEFAULT_OUT_DIR = PROJECT_ROOT / "artifacts"
+DEFAULT_OUT_DIR = PROJECT_ROOT / "artifacts" / "generated" / "alignment"
 
 CLASSIFICATION_FIELDS = ("level_1", "level_2", "level_3", "level_4")
 CODE_RE = re.compile(r"[A-Za-z]+\d*(?:-\d+)*")
@@ -1314,15 +1314,15 @@ def main(argv: list[str] | None = None) -> int:
     parser = argparse.ArgumentParser(
         description=(
             "Analyze dataset <-> corpus alignment (exact/normalized matching only) "
-            "and write artifacts/data_alignment_report.json + .md"
+            "and write artifacts/generated/alignment/data_alignment_report.json + .md"
         )
     )
     parser.add_argument("--overwrite", action="store_true", help="Overwrite outputs")
     args = parser.parse_args(argv)
 
     build_report(overwrite=args.overwrite)
-    print("wrote: artifacts/data_alignment_report.json")
-    print("wrote: artifacts/data_alignment_report.md")
+    print("wrote: artifacts/generated/alignment/data_alignment_report.json")
+    print("wrote: artifacts/generated/alignment/data_alignment_report.md")
     return 0
 
 
diff --git a/src/agent/task/dataset_config.py b/src/agent/task/dataset_config.py
index 647c4b2..c00a672 100644
--- a/src/agent/task/dataset_config.py
+++ b/src/agent/task/dataset_config.py
@@ -1,6 +1,6 @@
 """Per-dataset configuration for the canonical training target.
 
-Decision basis: artifacts/data_alignment_report (stage 1). Each dataset gets
+Decision basis: artifacts/generated/alignment/data_alignment_report (stage 1). Each dataset gets
 an explicit category_id strategy instead of a single global rule:
 
 - shougang: guanji codes are stable and unique (233 codes, letter prefix