{
  "_说明": [
    "PPT 取数用的总表。每个数字带「来源」,写的是它出自哪一份 summary.json 的哪条路径,拿着它能回到原始记录重算。",
    "⚠️ **跨轮的数不要混着用。** I3 实测:同一份配置隔一夜,D 维度从 23.2 掉到 21.8(I3 报告 §1)。",
    "⚠️ **I1 与 I3 的 q4 不可比**:I1 两边都接了 Kronos key,I3 与 I2/M2 同口径(key 清空,考的是诚实度)。",
    "数字全部来自真实调用;取不到写 null,不补默认值。"
  ],
  "生成命令": "python3 tools/eval/build_summary.py",
  "轮次": {
    "I1(完整对比测试 · 2026-09-23)": {
      "问题": "HCA 与社区版在同一套题上谁更强",
      "口径": {
        "题集": "10 道(M2 五道 + I1 新增五道)",
        "每题遍数": 3,
        "Kronos key": "接上(两边都接)",
        "线": "A 线 = 两边同样 6 个 MCP;B 线 = 产品形态"
      },
      "头条": {
        "A线_总分比": "92.0 / 90.9 = 101.2%",
        "B线_总分比": "95.1 / 86.7 = 109.7%",
        "来源": "docs/开发文档/I1-对比测试报告.md · 明细见 docs/eval/c1/summary.json"
      },
      "逐题_B线_只列与I3同名的五道": {
        "q1-fundamental": {
          "hca_调用中位": 1,
          "社区版_调用中位": 2,
          "hca_墙钟中位_ms": 15511,
          "社区版_墙钟中位_ms": 28265,
          "墙钟比": 0.549,
          "hca_D1": 13.0,
          "hca_D2": 12.0,
          "hca_D": 25.0,
          "n_hca": 3,
          "n_社区版": 3,
          "来源": "docs/eval/c1/summary.json → 线.c1-b.题.q1-fundamental"
        },
        "q2-thesis-review": {
          "hca_调用中位": 1,
          "社区版_调用中位": 2,
          "hca_墙钟中位_ms": 19901,
          "社区版_墙钟中位_ms": 9113,
          "墙钟比": 2.184,
          "hca_D1": 13.0,
          "hca_D2": 5.5,
          "hca_D": 18.5,
          "n_hca": 3,
          "n_社区版": 3,
          "来源": "docs/eval/c1/summary.json → 线.c1-b.题.q2-thesis-review"
        },
        "q3-factor-screen": {
          "hca_调用中位": 1,
          "社区版_调用中位": 1,
          "hca_墙钟中位_ms": 12032,
          "社区版_墙钟中位_ms": 8289,
          "墙钟比": 1.452,
          "hca_D1": 13.0,
          "hca_D2": 8.3,
          "hca_D": 21.3,
          "n_hca": 3,
          "n_社区版": 3,
          "来源": "docs/eval/c1/summary.json → 线.c1-b.题.q3-factor-screen"
        },
        "q4-kronos-forecast": {
          "hca_调用中位": 1,
          "社区版_调用中位": 0,
          "hca_墙钟中位_ms": 18792,
          "社区版_墙钟中位_ms": 4428,
          "墙钟比": 4.244,
          "hca_D1": 13.0,
          "hca_D2": 2.8,
          "hca_D": 15.8,
          "n_hca": 3,
          "n_社区版": 3,
          "来源": "docs/eval/c1/summary.json → 线.c1-b.题.q4-kronos-forecast"
        },
        "q5-intel-digest": {
          "hca_调用中位": 1,
          "社区版_调用中位": 3,
          "hca_墙钟中位_ms": 17828,
          "社区版_墙钟中位_ms": 11684,
          "墙钟比": 1.526,
          "hca_D1": 13.0,
          "hca_D2": 7.9,
          "hca_D": 20.9,
          "n_hca": 3,
          "n_社区版": 3,
          "来源": "docs/eval/c1/summary.json → 线.c1-b.题.q5-intel-digest"
        }
      },
      "明细文件": "docs/eval/c1/summary.json(含全部 10 道题与 A 线)"
    },
    "I3(性能收尾 · 2026-09-23)": {
      "问题": "逐题墙钟与步数相对社区版还差多少",
      "口径": {
        "题集": "M2 五道",
        "每题遍数": 12,
        "Kronos key": "清空(与 M2 / I2 同口径)",
        "线": "只跑 B 线(产品形态)",
        "批次": "opt5-fork-b-12 = 两段各 6 遍合并;对照档 opt3-fork-b-重测-12 同样 12 遍"
      },
      "头条": {
        "D维度_按题平均": 22.2,
        "逐题达标_墙钟≤社区版x1.05且调用≤社区版": "1/5",
        "工具调用次数≤社区版": "5/5",
        "A_准确性": 25.0,
        "C_输出规范": 25.0,
        "B_能力覆盖": 23.4,
        "ABC的n": 5,
        "来源": "docs/eval/i3/summary.json(D 维度)+ docs/eval/i3/opt5-fork-b-12/scores-summary.json(A/B/C)"
      },
      "同配置对照档_D维度": 21.8,
      "同配置对照档_说明": "opt3-fork-b-重测-12 = I2 发布的那一版配置,今天同条件重测。I2 当时记的是 23.2 —— 差的 1.4 分是跨天漂移,不是配置变了。",
      "逐题": {
        "q1-fundamental": {
          "hca_调用中位": 1.0,
          "社区版_调用中位": 2.0,
          "hca_墙钟中位_ms": 15677.0,
          "社区版_墙钟中位_ms": 31062.0,
          "墙钟比": 0.505,
          "hca_D1": 13.0,
          "hca_D2": 12.0,
          "hca_D": 25.0,
          "n_hca": 12,
          "n_社区版": 12,
          "来源": "docs/eval/i3/summary.json → 线.opt5-fork-b-12.题.q1-fundamental"
        },
        "q2-thesis-review": {
          "hca_调用中位": 1.0,
          "社区版_调用中位": 2,
          "hca_墙钟中位_ms": 18567.0,
          "社区版_墙钟中位_ms": 9366,
          "墙钟比": 1.982,
          "hca_D1": 13.0,
          "hca_D2": 6.1,
          "hca_D": 19.1,
          "n_hca": 12,
          "n_社区版": 11,
          "来源": "docs/eval/i3/summary.json → 线.opt5-fork-b-12.题.q2-thesis-review"
        },
        "q3-factor-screen": {
          "hca_调用中位": 1.0,
          "社区版_调用中位": 1.0,
          "hca_墙钟中位_ms": 12064.5,
          "社区版_墙钟中位_ms": 10126.5,
          "墙钟比": 1.191,
          "hca_D1": 13.0,
          "hca_D2": 10.1,
          "hca_D": 23.1,
          "n_hca": 12,
          "n_社区版": 12,
          "来源": "docs/eval/i3/summary.json → 线.opt5-fork-b-12.题.q3-factor-screen"
        },
        "q4-kronos-forecast": {
          "hca_调用中位": 0.0,
          "社区版_调用中位": 0.0,
          "hca_墙钟中位_ms": 6390.5,
          "社区版_墙钟中位_ms": 4732.5,
          "墙钟比": 1.35,
          "hca_D1": 13.0,
          "hca_D2": 8.9,
          "hca_D": 21.9,
          "n_hca": 12,
          "n_社区版": 12,
          "来源": "docs/eval/i3/summary.json → 线.opt5-fork-b-12.题.q4-kronos-forecast"
        },
        "q5-intel-digest": {
          "hca_调用中位": 1.0,
          "社区版_调用中位": 3.0,
          "hca_墙钟中位_ms": 15827.5,
          "社区版_墙钟中位_ms": 11488.5,
          "墙钟比": 1.378,
          "hca_D1": 13.0,
          "hca_D2": 8.7,
          "hca_D": 21.7,
          "n_hca": 12,
          "n_社区版": 12,
          "来源": "docs/eval/i3/summary.json → 线.opt5-fork-b-12.题.q5-intel-digest"
        }
      },
      "明细文件": "docs/eval/i3/summary.json"
    },
    "I4(响应时间专项 · 2026-09-23)": {
      "问题": "同机、同一个数据库、同一份数据的条件下,响应时间与社区版差多少,差在哪一段",
      "口径": {
        "题集": "10 道(与 I1 同一份)+ 纯引擎空载题 q0-idle",
        "每题遍数": "i4-b 10 遍 / i4-a 6 遍 / 空载题 20 遍",
        "Kronos key": "接上(与 I1 同口径,所以 q4 与 I2/I3 不可比)",
        "线": "i4-b = 产品形态;i4-a = 两边同样 6 个 MCP",
        "同库": "两边共用同一个 api 进程 + 同一个 postgres 实例 + 同一行用户,每批开跑前六条判据取证(<批次>/same-db-proof.txt)",
        "代码": "main(v0.2.1)—— 与 I1 当时的 v0.2.0 不同,绝对值不能和 I1 直接比"
      },
      "纯引擎空载响应": {
        "产品形态_b": {
          "n_hca": 20,
          "n_社区版": 20,
          "hca_首字中位_ms": 5353.65,
          "社区版_首字中位_ms": 4020.5,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 5353.85,
          "社区版_答案首字中位_ms": 4020.5,
          "hca_答案首字P90_ms": 7610.8,
          "社区版_答案首字P90_ms": 5059,
          "hca_墙钟中位_ms": 5539.5,
          "社区版_墙钟中位_ms": 4114.5,
          "hca_墙钟P90_ms": 7757,
          "社区版_墙钟P90_ms": 5132,
          "答案首字比_中位": 1.332,
          "答案首字比_P90": 1.504,
          "墙钟比_中位": 1.346,
          "墙钟比_P90": 1.511,
          "模型段比_中位": 1.332,
          "工具段比_中位": null,
          "出字段比_中位": 0.953,
          "来源": "docs/eval/i4/summary.json → batches.idle-b.questions.q0-idle.原样"
        },
        "A线_a": {
          "n_hca": 20,
          "n_社区版": 20,
          "hca_首字中位_ms": 4571.4,
          "社区版_首字中位_ms": 4523.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 4571.3,
          "社区版_答案首字中位_ms": 4523.0,
          "hca_答案首字P90_ms": 6585.5,
          "社区版_答案首字P90_ms": 5577,
          "hca_墙钟中位_ms": 4708.5,
          "社区版_墙钟中位_ms": 4640.5,
          "hca_墙钟P90_ms": 6859,
          "社区版_墙钟P90_ms": 5663,
          "答案首字比_中位": 1.011,
          "答案首字比_P90": 1.181,
          "墙钟比_中位": 1.015,
          "墙钟比_P90": 1.211,
          "模型段比_中位": 1.011,
          "工具段比_中位": null,
          "出字段比_中位": 1.174,
          "来源": "docs/eval/i4/summary.json → batches.idle-a.questions.q0-idle.原样"
        },
        "说明": "题面「用一句话说明你是什么。不要调用任何工具。」两边都是 1 轮、0 次工具调用。这是最干净的一组引擎响应时间对照 —— 没有工具往返、没有路标。"
      },
      "十题合计_产品形态_都真做题": {
        "题数": 10,
        "口径": "都真做题",
        "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.*.都真做题(逐题中位数之和)",
        "wall_ms": {
          "hca": 217871.0,
          "社区版": 282417.0,
          "比值": 0.771
        },
        "ttft_answer_ms": {
          "hca": 164414.2,
          "社区版": 222869.0,
          "比值": 0.738
        },
        "ttft_ms": {
          "hca": 103825.5,
          "社区版": 222869.0,
          "比值": 0.466
        },
        "model_ms": {
          "hca": 129885.6,
          "社区版": 137694.0,
          "比值": 0.943
        },
        "tool_ms": {
          "hca": 54774.7,
          "社区版": 120624.0,
          "比值": 0.454
        },
        "stream_ms": {
          "hca": 34834.2,
          "社区版": 22877.5,
          "比值": 1.523
        },
        "token": {
          "hca": 660907.0,
          "社区版": 672193.5,
          "比值": 0.983
        },
        "tool_calls": {
          "hca": 20.0,
          "社区版": 40.0,
          "比值": 0.5
        },
        "rounds": {
          "hca": 27.0,
          "社区版": 38.5,
          "比值": 0.701
        },
        "text_len": {
          "hca": 14121.5,
          "社区版": 8115.0,
          "比值": 1.74
        }
      },
      "十题合计_产品形态_都真做题_去q4": {
        "题数": 9,
        "口径": "都真做题",
        "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.*.都真做题(逐题中位数之和)",
        "wall_ms": {
          "hca": 196108.5,
          "社区版": 277948.0,
          "比值": 0.706
        },
        "ttft_answer_ms": {
          "hca": 145488.9,
          "社区版": 218633.0,
          "比值": 0.665
        },
        "ttft_ms": {
          "hca": 89454.8,
          "社区版": 218633.0,
          "比值": 0.409
        },
        "model_ms": {
          "hca": 119661.5,
          "社区版": 133458.0,
          "比值": 0.897
        },
        "tool_ms": {
          "hca": 46272.9,
          "社区版": 120624.0,
          "比值": 0.384
        },
        "stream_ms": {
          "hca": 32382.5,
          "社区版": 22620.5,
          "比值": 1.432
        },
        "token": {
          "hca": 613723.0,
          "社区版": 656977.5,
          "比值": 0.934
        },
        "tool_calls": {
          "hca": 18.0,
          "社区版": 40.0,
          "比值": 0.45
        },
        "rounds": {
          "hca": 25.0,
          "社区版": 37.5,
          "比值": 0.667
        },
        "text_len": {
          "hca": 13110.5,
          "社区版": 8012.0,
          "比值": 1.636
        }
      },
      "十题合计_A线_都真做题": {
        "题数": 10,
        "口径": "都真做题",
        "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.*.都真做题(逐题中位数之和)",
        "wall_ms": {
          "hca": 361030.5,
          "社区版": 302694.0,
          "比值": 1.193
        },
        "ttft_answer_ms": {
          "hca": 285467.5,
          "社区版": 245457.5,
          "比值": 1.163
        },
        "ttft_ms": {
          "hca": 83357.8,
          "社区版": 245457.5,
          "比值": 0.34
        },
        "model_ms": {
          "hca": 144398.2,
          "社区版": 133119.0,
          "比值": 1.085
        },
        "tool_ms": {
          "hca": 157398.9,
          "社区版": 127808.0,
          "比值": 1.232
        },
        "stream_ms": {
          "hca": 38630.7,
          "社区版": 22482.0,
          "比值": 1.718
        },
        "token": {
          "hca": 587265.0,
          "社区版": 651311.5,
          "比值": 0.902
        },
        "tool_calls": {
          "hca": 23.5,
          "社区版": 38.5,
          "比值": 0.61
        },
        "rounds": {
          "hca": 27.0,
          "社区版": 37.5,
          "比值": 0.72
        },
        "text_len": {
          "hca": 12800.5,
          "社区版": 7889.0,
          "比值": 1.623
        }
      },
      "_合计怎么读": "「比值」= HCA ÷ 社区版,< 1 表示 HCA 更快 / 更少。去 q4 那一份是因为那道题两边做的不是同一件事(HCA 真跑 Kronos GPU 推理,社区版如实说没有这个能力)。",
      "逐题_产品形态_原样": {
        "q1-fundamental": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 13074.3,
          "社区版_首字中位_ms": 31098.5,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 13074.7,
          "社区版_答案首字中位_ms": 31098.5,
          "hca_答案首字P90_ms": 17385.5,
          "社区版_答案首字P90_ms": 46012,
          "hca_墙钟中位_ms": 16441.5,
          "社区版_墙钟中位_ms": 33008.0,
          "hca_墙钟P90_ms": 20023,
          "社区版_墙钟P90_ms": 46903,
          "答案首字比_中位": 0.42,
          "答案首字比_P90": 0.378,
          "墙钟比_中位": 0.498,
          "墙钟比_P90": 0.427,
          "模型段比_中位": 1.092,
          "工具段比_中位": 0.134,
          "出字段比_中位": 1.728,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q1-fundamental.原样"
        },
        "q10-refusal": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 5437.6,
          "社区版_首字中位_ms": 14368.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 5437.65,
          "社区版_答案首字中位_ms": 14368.0,
          "hca_答案首字P90_ms": 6503.1,
          "社区版_答案首字P90_ms": 15852,
          "hca_墙钟中位_ms": 6483.5,
          "社区版_墙钟中位_ms": 14738.0,
          "hca_墙钟P90_ms": 7434,
          "社区版_墙钟P90_ms": 16168,
          "答案首字比_中位": 0.378,
          "答案首字比_P90": 0.41,
          "墙钟比_中位": 0.44,
          "墙钟比_P90": 0.46,
          "模型段比_中位": 0.38,
          "工具段比_中位": 0.0,
          "出字段比_中位": 2.52,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q10-refusal.原样"
        },
        "q2-thesis-review": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 16957.1,
          "社区版_首字中位_ms": 43335,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 16957.55,
          "社区版_答案首字中位_ms": 43335,
          "hca_答案首字P90_ms": 24817.1,
          "社区版_答案首字P90_ms": 50632,
          "hca_墙钟中位_ms": 21798.5,
          "社区版_墙钟中位_ms": 43505.0,
          "hca_墙钟P90_ms": 27749,
          "社区版_墙钟P90_ms": 48785,
          "答案首字比_中位": 0.391,
          "答案首字比_P90": 0.49,
          "墙钟比_中位": 0.501,
          "墙钟比_P90": 0.569,
          "模型段比_中位": 0.713,
          "工具段比_中位": 0.152,
          "出字段比_中位": 4.227,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q2-thesis-review.原样"
        },
        "q3-factor-screen": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 8742.75,
          "社区版_首字中位_ms": 7057.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 8743.2,
          "社区版_答案首字中位_ms": 7057.0,
          "hca_答案首字P90_ms": 9799.3,
          "社区版_答案首字P90_ms": 8398,
          "hca_墙钟中位_ms": 12388.5,
          "社区版_墙钟中位_ms": 9676.0,
          "hca_墙钟P90_ms": 13833,
          "社区版_墙钟P90_ms": 10360,
          "答案首字比_中位": 1.239,
          "答案首字比_P90": 1.167,
          "墙钟比_中位": 1.28,
          "墙钟比_P90": 1.335,
          "模型段比_中位": 1.229,
          "工具段比_中位": 1.44,
          "出字段比_中位": 1.489,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q3-factor-screen.原样"
        },
        "q4-kronos-forecast": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 14370.7,
          "社区版_首字中位_ms": 4236.0,
          "hca_首字是路标的次数": 5,
          "hca_答案首字中位_ms": 18925.35,
          "社区版_答案首字中位_ms": 4236.0,
          "hca_答案首字P90_ms": 26007.2,
          "社区版_答案首字P90_ms": 4364,
          "hca_墙钟中位_ms": 21762.5,
          "社区版_墙钟中位_ms": 4469.0,
          "hca_墙钟P90_ms": 28166,
          "社区版_墙钟P90_ms": 4656,
          "答案首字比_中位": 4.468,
          "答案首字比_P90": 5.959,
          "墙钟比_中位": 4.87,
          "墙钟比_P90": 6.049,
          "模型段比_中位": 2.414,
          "工具段比_中位": null,
          "出字段比_中位": 9.54,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q4-kronos-forecast.原样"
        },
        "q5-intel-digest": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 13504.05,
          "社区版_首字中位_ms": 10465.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 13504.55,
          "社区版_答案首字中位_ms": 10465.0,
          "hca_答案首字P90_ms": 17581.9,
          "社区版_答案首字P90_ms": 10981,
          "hca_墙钟中位_ms": 16923.5,
          "社区版_墙钟中位_ms": 11526.0,
          "hca_墙钟P90_ms": 20171,
          "社区版_墙钟P90_ms": 12776,
          "答案首字比_中位": 1.29,
          "答案首字比_P90": 1.601,
          "墙钟比_中位": 1.468,
          "墙钟比_P90": 1.579,
          "模型段比_中位": 1.584,
          "工具段比_中位": 0.872,
          "出字段比_中位": 2.458,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q5-intel-digest.原样"
        },
        "q6-investor-panel": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 3974.3,
          "社区版_首字中位_ms": 47876.0,
          "hca_首字是路标的次数": 9,
          "hca_答案首字中位_ms": 35872.4,
          "社区版_答案首字中位_ms": 47876.0,
          "hca_答案首字P90_ms": 71289.6,
          "社区版_答案首字P90_ms": 56113,
          "hca_墙钟中位_ms": 45455.0,
          "社区版_墙钟中位_ms": 59706.0,
          "hca_墙钟P90_ms": 83630,
          "社区版_墙钟P90_ms": 73800,
          "答案首字比_中位": 0.749,
          "答案首字比_P90": 1.27,
          "墙钟比_中位": 0.761,
          "墙钟比_P90": 1.133,
          "模型段比_中位": 0.663,
          "工具段比_中位": 1.002,
          "出字段比_中位": 0.742,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q6-investor-panel.原样"
        },
        "q7-lhb": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 18099.5,
          "社区版_首字中位_ms": 37399.0,
          "hca_首字是路标的次数": 2,
          "hca_答案首字中位_ms": 22271.8,
          "社区版_答案首字中位_ms": 37399.0,
          "hca_答案首字P90_ms": 100852.1,
          "社区版_答案首字P90_ms": 45319,
          "hca_墙钟中位_ms": 24117.0,
          "社区版_墙钟中位_ms": 37797.5,
          "hca_墙钟P90_ms": 101851,
          "社区版_墙钟P90_ms": 45765,
          "答案首字比_中位": 0.596,
          "答案首字比_P90": 2.225,
          "墙钟比_中位": 0.638,
          "墙钟比_P90": 2.226,
          "模型段比_中位": 1.651,
          "工具段比_中位": 0.043,
          "出字段比_中位": 4.953,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q7-lhb.原样"
        },
        "q8-trap-detector": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 4008.1,
          "社区版_首字中位_ms": 15963.5,
          "hca_首字是路标的次数": 8,
          "hca_答案首字中位_ms": 17635.2,
          "社区版_答案首字中位_ms": 15963.5,
          "hca_答案首字P90_ms": 22591.1,
          "社区版_答案首字P90_ms": 40220,
          "hca_墙钟中位_ms": 21753.0,
          "社区版_墙钟中位_ms": 17744.0,
          "hca_墙钟P90_ms": 27795,
          "社区版_墙钟P90_ms": 43035,
          "答案首字比_中位": 1.105,
          "答案首字比_P90": 0.562,
          "墙钟比_中位": 1.226,
          "墙钟比_P90": 0.646,
          "模型段比_中位": 0.959,
          "工具段比_中位": 2.097,
          "出字段比_中位": 1.944,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q8-trap-detector.原样"
        },
        "q9-multiturn": {
          "n_hca": 10,
          "n_社区版": 9,
          "hca_首字中位_ms": 5657.1,
          "社区版_首字中位_ms": 10781,
          "hca_首字是路标的次数": 10,
          "hca_答案首字中位_ms": 11991.85,
          "社区版_答案首字中位_ms": 10781,
          "hca_答案首字P90_ms": 15669.4,
          "社区版_答案首字P90_ms": 21326,
          "hca_墙钟中位_ms": 30748.0,
          "社区版_墙钟中位_ms": 49170,
          "hca_墙钟P90_ms": 43954,
          "社区版_墙钟P90_ms": 72310,
          "答案首字比_中位": 1.112,
          "答案首字比_P90": 0.735,
          "墙钟比_中位": 0.625,
          "墙钟比_P90": 0.608,
          "模型段比_中位": 0.862,
          "工具段比_中位": 0.346,
          "出字段比_中位": 2.817,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q9-multiturn.原样"
        }
      },
      "逐题_产品形态_都真做题": {
        "q1-fundamental": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 13074.3,
          "社区版_首字中位_ms": 31098.5,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 13074.7,
          "社区版_答案首字中位_ms": 31098.5,
          "hca_答案首字P90_ms": 17385.5,
          "社区版_答案首字P90_ms": 46012,
          "hca_墙钟中位_ms": 16441.5,
          "社区版_墙钟中位_ms": 33008.0,
          "hca_墙钟P90_ms": 20023,
          "社区版_墙钟P90_ms": 46903,
          "答案首字比_中位": 0.42,
          "答案首字比_P90": 0.378,
          "墙钟比_中位": 0.498,
          "墙钟比_P90": 0.427,
          "模型段比_中位": 1.092,
          "工具段比_中位": 0.134,
          "出字段比_中位": 1.728,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q1-fundamental.都真做题"
        },
        "q10-refusal": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 5437.6,
          "社区版_首字中位_ms": 14368.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 5437.65,
          "社区版_答案首字中位_ms": 14368.0,
          "hca_答案首字P90_ms": 6503.1,
          "社区版_答案首字P90_ms": 15852,
          "hca_墙钟中位_ms": 6483.5,
          "社区版_墙钟中位_ms": 14738.0,
          "hca_墙钟P90_ms": 7434,
          "社区版_墙钟P90_ms": 16168,
          "答案首字比_中位": 0.378,
          "答案首字比_P90": 0.41,
          "墙钟比_中位": 0.44,
          "墙钟比_P90": 0.46,
          "模型段比_中位": 0.38,
          "工具段比_中位": 0.0,
          "出字段比_中位": 2.52,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q10-refusal.都真做题"
        },
        "q2-thesis-review": {
          "n_hca": 10,
          "n_社区版": 8,
          "hca_首字中位_ms": 16957.1,
          "社区版_首字中位_ms": 43625.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 16957.55,
          "社区版_答案首字中位_ms": 43625.0,
          "hca_答案首字P90_ms": 24817.1,
          "社区版_答案首字P90_ms": 50632,
          "hca_墙钟中位_ms": 21798.5,
          "社区版_墙钟中位_ms": 44582.5,
          "hca_墙钟P90_ms": 27749,
          "社区版_墙钟P90_ms": 51978,
          "答案首字比_中位": 0.389,
          "答案首字比_P90": 0.49,
          "墙钟比_中位": 0.489,
          "墙钟比_P90": 0.534,
          "模型段比_中位": 0.67,
          "工具段比_中位": 0.147,
          "出字段比_中位": 4.199,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q2-thesis-review.都真做题"
        },
        "q3-factor-screen": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 8742.75,
          "社区版_首字中位_ms": 7057.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 8743.2,
          "社区版_答案首字中位_ms": 7057.0,
          "hca_答案首字P90_ms": 9799.3,
          "社区版_答案首字P90_ms": 8398,
          "hca_墙钟中位_ms": 12388.5,
          "社区版_墙钟中位_ms": 9676.0,
          "hca_墙钟P90_ms": 13833,
          "社区版_墙钟P90_ms": 10360,
          "答案首字比_中位": 1.239,
          "答案首字比_P90": 1.167,
          "墙钟比_中位": 1.28,
          "墙钟比_P90": 1.335,
          "模型段比_中位": 1.229,
          "工具段比_中位": 1.44,
          "出字段比_中位": 1.489,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q3-factor-screen.都真做题"
        },
        "q4-kronos-forecast": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 14370.7,
          "社区版_首字中位_ms": 4236.0,
          "hca_首字是路标的次数": 5,
          "hca_答案首字中位_ms": 18925.35,
          "社区版_答案首字中位_ms": 4236.0,
          "hca_答案首字P90_ms": 26007.2,
          "社区版_答案首字P90_ms": 4364,
          "hca_墙钟中位_ms": 21762.5,
          "社区版_墙钟中位_ms": 4469.0,
          "hca_墙钟P90_ms": 28166,
          "社区版_墙钟P90_ms": 4656,
          "答案首字比_中位": 4.468,
          "答案首字比_P90": 5.959,
          "墙钟比_中位": 4.87,
          "墙钟比_P90": 6.049,
          "模型段比_中位": 2.414,
          "工具段比_中位": null,
          "出字段比_中位": 9.54,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q4-kronos-forecast.都真做题"
        },
        "q5-intel-digest": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 13504.05,
          "社区版_首字中位_ms": 10465.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 13504.55,
          "社区版_答案首字中位_ms": 10465.0,
          "hca_答案首字P90_ms": 17581.9,
          "社区版_答案首字P90_ms": 10981,
          "hca_墙钟中位_ms": 16923.5,
          "社区版_墙钟中位_ms": 11526.0,
          "hca_墙钟P90_ms": 20171,
          "社区版_墙钟P90_ms": 12776,
          "答案首字比_中位": 1.29,
          "答案首字比_P90": 1.601,
          "墙钟比_中位": 1.468,
          "墙钟比_P90": 1.579,
          "模型段比_中位": 1.584,
          "工具段比_中位": 0.872,
          "出字段比_中位": 2.458,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q5-intel-digest.都真做题"
        },
        "q6-investor-panel": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 3974.3,
          "社区版_首字中位_ms": 47876.0,
          "hca_首字是路标的次数": 9,
          "hca_答案首字中位_ms": 35872.4,
          "社区版_答案首字中位_ms": 47876.0,
          "hca_答案首字P90_ms": 71289.6,
          "社区版_答案首字P90_ms": 56113,
          "hca_墙钟中位_ms": 45455.0,
          "社区版_墙钟中位_ms": 59706.0,
          "hca_墙钟P90_ms": 83630,
          "社区版_墙钟P90_ms": 73800,
          "答案首字比_中位": 0.749,
          "答案首字比_P90": 1.27,
          "墙钟比_中位": 0.761,
          "墙钟比_P90": 1.133,
          "模型段比_中位": 0.663,
          "工具段比_中位": 1.002,
          "出字段比_中位": 0.742,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q6-investor-panel.都真做题"
        },
        "q7-lhb": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 18099.5,
          "社区版_首字中位_ms": 37399.0,
          "hca_首字是路标的次数": 2,
          "hca_答案首字中位_ms": 22271.8,
          "社区版_答案首字中位_ms": 37399.0,
          "hca_答案首字P90_ms": 100852.1,
          "社区版_答案首字P90_ms": 45319,
          "hca_墙钟中位_ms": 24117.0,
          "社区版_墙钟中位_ms": 37797.5,
          "hca_墙钟P90_ms": 101851,
          "社区版_墙钟P90_ms": 45765,
          "答案首字比_中位": 0.596,
          "答案首字比_P90": 2.225,
          "墙钟比_中位": 0.638,
          "墙钟比_P90": 2.226,
          "模型段比_中位": 1.651,
          "工具段比_中位": 0.043,
          "出字段比_中位": 4.953,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q7-lhb.都真做题"
        },
        "q8-trap-detector": {
          "n_hca": 10,
          "n_社区版": 10,
          "hca_首字中位_ms": 4008.1,
          "社区版_首字中位_ms": 15963.5,
          "hca_首字是路标的次数": 8,
          "hca_答案首字中位_ms": 17635.2,
          "社区版_答案首字中位_ms": 15963.5,
          "hca_答案首字P90_ms": 22591.1,
          "社区版_答案首字P90_ms": 40220,
          "hca_墙钟中位_ms": 21753.0,
          "社区版_墙钟中位_ms": 17744.0,
          "hca_墙钟P90_ms": 27795,
          "社区版_墙钟P90_ms": 43035,
          "答案首字比_中位": 1.105,
          "答案首字比_P90": 0.562,
          "墙钟比_中位": 1.226,
          "墙钟比_P90": 0.646,
          "模型段比_中位": 0.959,
          "工具段比_中位": 2.097,
          "出字段比_中位": 1.944,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q8-trap-detector.都真做题"
        },
        "q9-multiturn": {
          "n_hca": 10,
          "n_社区版": 9,
          "hca_首字中位_ms": 5657.1,
          "社区版_首字中位_ms": 10781,
          "hca_首字是路标的次数": 10,
          "hca_答案首字中位_ms": 11991.85,
          "社区版_答案首字中位_ms": 10781,
          "hca_答案首字P90_ms": 15669.4,
          "社区版_答案首字P90_ms": 21326,
          "hca_墙钟中位_ms": 30748.0,
          "社区版_墙钟中位_ms": 49170,
          "hca_墙钟P90_ms": 43954,
          "社区版_墙钟P90_ms": 72310,
          "答案首字比_中位": 1.112,
          "答案首字比_P90": 0.735,
          "墙钟比_中位": 0.625,
          "墙钟比_P90": 0.608,
          "模型段比_中位": 0.862,
          "工具段比_中位": 0.346,
          "出字段比_中位": 2.817,
          "来源": "docs/eval/i4/summary.json → batches.i4-b.questions.q9-multiturn.都真做题"
        }
      },
      "逐题_A线_原样": {
        "q1-fundamental": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 8758.2,
          "社区版_首字中位_ms": 39045.0,
          "hca_首字是路标的次数": 4,
          "hca_答案首字中位_ms": 54661.35,
          "社区版_答案首字中位_ms": 39045.0,
          "hca_答案首字P90_ms": 107368.9,
          "社区版_答案首字P90_ms": 65117,
          "hca_墙钟中位_ms": 58548.5,
          "社区版_墙钟中位_ms": 40572.5,
          "hca_墙钟P90_ms": 110832,
          "社区版_墙钟P90_ms": 66393,
          "答案首字比_中位": 1.4,
          "答案首字比_P90": 1.649,
          "墙钟比_中位": 1.443,
          "墙钟比_P90": 1.669,
          "模型段比_中位": 1.415,
          "工具段比_中位": 1.256,
          "出字段比_中位": 2.581,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q1-fundamental.原样"
        },
        "q10-refusal": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 6508.55,
          "社区版_首字中位_ms": 15203.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 6508.75,
          "社区版_答案首字中位_ms": 15203.0,
          "hca_答案首字P90_ms": 8436.1,
          "社区版_答案首字P90_ms": 17773,
          "hca_墙钟中位_ms": 7368.5,
          "社区版_墙钟中位_ms": 15580.5,
          "hca_墙钟P90_ms": 9492,
          "社区版_墙钟P90_ms": 18354,
          "答案首字比_中位": 0.428,
          "答案首字比_P90": 0.475,
          "墙钟比_中位": 0.473,
          "墙钟比_P90": 0.517,
          "模型段比_中位": 0.43,
          "工具段比_中位": 0.0,
          "出字段比_中位": 2.001,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q10-refusal.原样"
        },
        "q2-thesis-review": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 6315.45,
          "社区版_首字中位_ms": 47598.0,
          "hca_首字是路标的次数": 4,
          "hca_答案首字中位_ms": 49930.75,
          "社区版_答案首字中位_ms": 47598.0,
          "hca_答案首字P90_ms": 64347.0,
          "社区版_答案首字P90_ms": 60351,
          "hca_墙钟中位_ms": 54551.5,
          "社区版_墙钟中位_ms": 49029.0,
          "hca_墙钟P90_ms": 69039,
          "社区版_墙钟P90_ms": 61848,
          "答案首字比_中位": 1.049,
          "答案首字比_P90": 1.066,
          "墙钟比_中位": 1.113,
          "墙钟比_P90": 1.116,
          "模型段比_中位": 0.986,
          "工具段比_中位": 1.042,
          "出字段比_中位": 3.243,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q2-thesis-review.原样"
        },
        "q3-factor-screen": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 9560.25,
          "社区版_首字中位_ms": 7614.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 9561.0,
          "社区版_答案首字中位_ms": 7614.0,
          "hca_答案首字P90_ms": 11026.3,
          "社区版_答案首字P90_ms": 10929,
          "hca_墙钟中位_ms": 13687.5,
          "社区版_墙钟中位_ms": 9868.0,
          "hca_墙钟P90_ms": 15181,
          "社区版_墙钟P90_ms": 13107,
          "答案首字比_中位": 1.256,
          "答案首字比_P90": 1.009,
          "墙钟比_中位": 1.387,
          "墙钟比_P90": 1.158,
          "模型段比_中位": 1.317,
          "工具段比_中位": 1.536,
          "出字段比_中位": 1.857,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q3-factor-screen.原样"
        },
        "q4-kronos-forecast": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 6202.2,
          "社区版_首字中位_ms": 4034.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 6202.6,
          "社区版_答案首字中位_ms": 4034.0,
          "hca_答案首字P90_ms": 15134.0,
          "社区版_答案首字P90_ms": 5502,
          "hca_墙钟中位_ms": 7034.0,
          "社区版_墙钟中位_ms": 4292.0,
          "hca_墙钟P90_ms": 15985,
          "社区版_墙钟P90_ms": 5822,
          "答案首字比_中位": 1.538,
          "答案首字比_P90": 2.751,
          "墙钟比_中位": 1.639,
          "墙钟比_P90": 2.746,
          "模型段比_中位": 1.537,
          "工具段比_中位": null,
          "出字段比_中位": 3.272,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q4-kronos-forecast.原样"
        },
        "q5-intel-digest": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 5651.55,
          "社区版_首字中位_ms": 10639.5,
          "hca_首字是路标的次数": 6,
          "hca_答案首字中位_ms": 17269.1,
          "社区版_答案首字中位_ms": 10639.5,
          "hca_答案首字P90_ms": 21910.3,
          "社区版_答案首字P90_ms": 11348,
          "hca_墙钟中位_ms": 20406.5,
          "社区版_墙钟中位_ms": 12057.5,
          "hca_墙钟P90_ms": 25633,
          "社区版_墙钟P90_ms": 12353,
          "答案首字比_中位": 1.623,
          "答案首字比_P90": 1.931,
          "墙钟比_中位": 1.692,
          "墙钟比_P90": 2.075,
          "模型段比_中位": 1.608,
          "工具段比_中位": 1.675,
          "出字段比_中位": 2.308,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q5-intel-digest.原样"
        },
        "q6-investor-panel": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 3897.95,
          "社区版_首字中位_ms": 56131.5,
          "hca_首字是路标的次数": 5,
          "hca_答案首字中位_ms": 56148.65,
          "社区版_答案首字中位_ms": 56131.5,
          "hca_答案首字P90_ms": 77032.4,
          "社区版_答案首字P90_ms": 61125,
          "hca_墙钟中位_ms": 67336.5,
          "社区版_墙钟中位_ms": 69153.0,
          "hca_墙钟P90_ms": 88346,
          "社区版_墙钟P90_ms": 76999,
          "答案首字比_中位": 1.0,
          "答案首字比_P90": 1.26,
          "墙钟比_中位": 0.974,
          "墙钟比_P90": 1.147,
          "模型段比_中位": 0.576,
          "工具段比_中位": 1.348,
          "出字段比_中位": 0.954,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q6-investor-panel.原样"
        },
        "q7-lhb": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 27657.55,
          "社区版_首字中位_ms": 38076.5,
          "hca_首字是路标的次数": 2,
          "hca_答案首字中位_ms": 51016.25,
          "社区版_答案首字中位_ms": 38076.5,
          "hca_答案首字P90_ms": 107800.8,
          "社区版_答案首字P90_ms": 45060,
          "hca_墙钟中位_ms": 52905.5,
          "社区版_墙钟中位_ms": 38369.5,
          "hca_墙钟P90_ms": 109492,
          "社区版_墙钟P90_ms": 45578,
          "答案首字比_中位": 1.34,
          "答案首字比_P90": 2.392,
          "墙钟比_中位": 1.379,
          "墙钟比_P90": 2.402,
          "模型段比_中位": 1.282,
          "工具段比_中位": 1.599,
          "出字段比_中位": 5.219,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q7-lhb.原样"
        },
        "q8-trap-detector": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 3949.5,
          "社区版_首字中位_ms": 19500.0,
          "hca_首字是路标的次数": 4,
          "hca_答案首字中位_ms": 23285.4,
          "社区版_答案首字中位_ms": 19500.0,
          "hca_答案首字P90_ms": 41138.3,
          "社区版_答案首字P90_ms": 48537,
          "hca_墙钟中位_ms": 27903.0,
          "社区版_墙钟中位_ms": 21072.5,
          "hca_墙钟P90_ms": 45690,
          "社区版_墙钟P90_ms": 50985,
          "答案首字比_中位": 1.194,
          "答案首字比_P90": 0.848,
          "墙钟比_中位": 1.324,
          "墙钟比_P90": 0.896,
          "模型段比_中位": 1.072,
          "工具段比_中位": 1.603,
          "出字段比_中位": 2.36,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q8-trap-detector.原样"
        },
        "q9-multiturn": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 4856.55,
          "社区版_首字中位_ms": 7616.0,
          "hca_首字是路标的次数": 6,
          "hca_答案首字中位_ms": 10883.65,
          "社区版_答案首字中位_ms": 7616.0,
          "hca_答案首字P90_ms": 14310.4,
          "社区版_答案首字P90_ms": 10332,
          "hca_墙钟中位_ms": 51289.0,
          "社区版_墙钟中位_ms": 42699.5,
          "hca_墙钟P90_ms": 123682,
          "社区版_墙钟P90_ms": 47692,
          "答案首字比_中位": 1.429,
          "答案首字比_P90": 1.385,
          "墙钟比_中位": 1.201,
          "墙钟比_P90": 2.593,
          "模型段比_中位": 1.844,
          "工具段比_中位": 0.838,
          "出字段比_中位": 4.06,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q9-multiturn.原样"
        }
      },
      "逐题_A线_都真做题": {
        "q1-fundamental": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 8758.2,
          "社区版_首字中位_ms": 39045.0,
          "hca_首字是路标的次数": 4,
          "hca_答案首字中位_ms": 54661.35,
          "社区版_答案首字中位_ms": 39045.0,
          "hca_答案首字P90_ms": 107368.9,
          "社区版_答案首字P90_ms": 65117,
          "hca_墙钟中位_ms": 58548.5,
          "社区版_墙钟中位_ms": 40572.5,
          "hca_墙钟P90_ms": 110832,
          "社区版_墙钟P90_ms": 66393,
          "答案首字比_中位": 1.4,
          "答案首字比_P90": 1.649,
          "墙钟比_中位": 1.443,
          "墙钟比_P90": 1.669,
          "模型段比_中位": 1.415,
          "工具段比_中位": 1.256,
          "出字段比_中位": 2.581,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q1-fundamental.都真做题"
        },
        "q10-refusal": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 6508.55,
          "社区版_首字中位_ms": 15203.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 6508.75,
          "社区版_答案首字中位_ms": 15203.0,
          "hca_答案首字P90_ms": 8436.1,
          "社区版_答案首字P90_ms": 17773,
          "hca_墙钟中位_ms": 7368.5,
          "社区版_墙钟中位_ms": 15580.5,
          "hca_墙钟P90_ms": 9492,
          "社区版_墙钟P90_ms": 18354,
          "答案首字比_中位": 0.428,
          "答案首字比_P90": 0.475,
          "墙钟比_中位": 0.473,
          "墙钟比_P90": 0.517,
          "模型段比_中位": 0.43,
          "工具段比_中位": 0.0,
          "出字段比_中位": 2.001,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q10-refusal.都真做题"
        },
        "q2-thesis-review": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 6315.45,
          "社区版_首字中位_ms": 47598.0,
          "hca_首字是路标的次数": 4,
          "hca_答案首字中位_ms": 49930.75,
          "社区版_答案首字中位_ms": 47598.0,
          "hca_答案首字P90_ms": 64347.0,
          "社区版_答案首字P90_ms": 60351,
          "hca_墙钟中位_ms": 54551.5,
          "社区版_墙钟中位_ms": 49029.0,
          "hca_墙钟P90_ms": 69039,
          "社区版_墙钟P90_ms": 61848,
          "答案首字比_中位": 1.049,
          "答案首字比_P90": 1.066,
          "墙钟比_中位": 1.113,
          "墙钟比_P90": 1.116,
          "模型段比_中位": 0.986,
          "工具段比_中位": 1.042,
          "出字段比_中位": 3.243,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q2-thesis-review.都真做题"
        },
        "q3-factor-screen": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 9560.25,
          "社区版_首字中位_ms": 7614.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 9561.0,
          "社区版_答案首字中位_ms": 7614.0,
          "hca_答案首字P90_ms": 11026.3,
          "社区版_答案首字P90_ms": 10929,
          "hca_墙钟中位_ms": 13687.5,
          "社区版_墙钟中位_ms": 9868.0,
          "hca_墙钟P90_ms": 15181,
          "社区版_墙钟P90_ms": 13107,
          "答案首字比_中位": 1.256,
          "答案首字比_P90": 1.009,
          "墙钟比_中位": 1.387,
          "墙钟比_P90": 1.158,
          "模型段比_中位": 1.317,
          "工具段比_中位": 1.536,
          "出字段比_中位": 1.857,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q3-factor-screen.都真做题"
        },
        "q4-kronos-forecast": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 6202.2,
          "社区版_首字中位_ms": 4034.0,
          "hca_首字是路标的次数": 0,
          "hca_答案首字中位_ms": 6202.6,
          "社区版_答案首字中位_ms": 4034.0,
          "hca_答案首字P90_ms": 15134.0,
          "社区版_答案首字P90_ms": 5502,
          "hca_墙钟中位_ms": 7034.0,
          "社区版_墙钟中位_ms": 4292.0,
          "hca_墙钟P90_ms": 15985,
          "社区版_墙钟P90_ms": 5822,
          "答案首字比_中位": 1.538,
          "答案首字比_P90": 2.751,
          "墙钟比_中位": 1.639,
          "墙钟比_P90": 2.746,
          "模型段比_中位": 1.537,
          "工具段比_中位": null,
          "出字段比_中位": 3.272,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q4-kronos-forecast.都真做题"
        },
        "q5-intel-digest": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 5651.55,
          "社区版_首字中位_ms": 10639.5,
          "hca_首字是路标的次数": 6,
          "hca_答案首字中位_ms": 17269.1,
          "社区版_答案首字中位_ms": 10639.5,
          "hca_答案首字P90_ms": 21910.3,
          "社区版_答案首字P90_ms": 11348,
          "hca_墙钟中位_ms": 20406.5,
          "社区版_墙钟中位_ms": 12057.5,
          "hca_墙钟P90_ms": 25633,
          "社区版_墙钟P90_ms": 12353,
          "答案首字比_中位": 1.623,
          "答案首字比_P90": 1.931,
          "墙钟比_中位": 1.692,
          "墙钟比_P90": 2.075,
          "模型段比_中位": 1.608,
          "工具段比_中位": 1.675,
          "出字段比_中位": 2.308,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q5-intel-digest.都真做题"
        },
        "q6-investor-panel": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 3897.95,
          "社区版_首字中位_ms": 56131.5,
          "hca_首字是路标的次数": 5,
          "hca_答案首字中位_ms": 56148.65,
          "社区版_答案首字中位_ms": 56131.5,
          "hca_答案首字P90_ms": 77032.4,
          "社区版_答案首字P90_ms": 61125,
          "hca_墙钟中位_ms": 67336.5,
          "社区版_墙钟中位_ms": 69153.0,
          "hca_墙钟P90_ms": 88346,
          "社区版_墙钟P90_ms": 76999,
          "答案首字比_中位": 1.0,
          "答案首字比_P90": 1.26,
          "墙钟比_中位": 0.974,
          "墙钟比_P90": 1.147,
          "模型段比_中位": 0.576,
          "工具段比_中位": 1.348,
          "出字段比_中位": 0.954,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q6-investor-panel.都真做题"
        },
        "q7-lhb": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 27657.55,
          "社区版_首字中位_ms": 38076.5,
          "hca_首字是路标的次数": 2,
          "hca_答案首字中位_ms": 51016.25,
          "社区版_答案首字中位_ms": 38076.5,
          "hca_答案首字P90_ms": 107800.8,
          "社区版_答案首字P90_ms": 45060,
          "hca_墙钟中位_ms": 52905.5,
          "社区版_墙钟中位_ms": 38369.5,
          "hca_墙钟P90_ms": 109492,
          "社区版_墙钟P90_ms": 45578,
          "答案首字比_中位": 1.34,
          "答案首字比_P90": 2.392,
          "墙钟比_中位": 1.379,
          "墙钟比_P90": 2.402,
          "模型段比_中位": 1.282,
          "工具段比_中位": 1.599,
          "出字段比_中位": 5.219,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q7-lhb.都真做题"
        },
        "q8-trap-detector": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 3949.5,
          "社区版_首字中位_ms": 19500.0,
          "hca_首字是路标的次数": 4,
          "hca_答案首字中位_ms": 23285.4,
          "社区版_答案首字中位_ms": 19500.0,
          "hca_答案首字P90_ms": 41138.3,
          "社区版_答案首字P90_ms": 48537,
          "hca_墙钟中位_ms": 27903.0,
          "社区版_墙钟中位_ms": 21072.5,
          "hca_墙钟P90_ms": 45690,
          "社区版_墙钟P90_ms": 50985,
          "答案首字比_中位": 1.194,
          "答案首字比_P90": 0.848,
          "墙钟比_中位": 1.324,
          "墙钟比_P90": 0.896,
          "模型段比_中位": 1.072,
          "工具段比_中位": 1.603,
          "出字段比_中位": 2.36,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q8-trap-detector.都真做题"
        },
        "q9-multiturn": {
          "n_hca": 6,
          "n_社区版": 6,
          "hca_首字中位_ms": 4856.55,
          "社区版_首字中位_ms": 7616.0,
          "hca_首字是路标的次数": 6,
          "hca_答案首字中位_ms": 10883.65,
          "社区版_答案首字中位_ms": 7616.0,
          "hca_答案首字P90_ms": 14310.4,
          "社区版_答案首字P90_ms": 10332,
          "hca_墙钟中位_ms": 51289.0,
          "社区版_墙钟中位_ms": 42699.5,
          "hca_墙钟P90_ms": 123682,
          "社区版_墙钟P90_ms": 47692,
          "答案首字比_中位": 1.429,
          "答案首字比_P90": 1.385,
          "墙钟比_中位": 1.201,
          "墙钟比_P90": 2.593,
          "模型段比_中位": 1.844,
          "工具段比_中位": 0.838,
          "出字段比_中位": 4.06,
          "来源": "docs/eval/i4/summary.json → batches.i4-a.questions.q9-multiturn.都真做题"
        }
      },
      "hook段": {
        "hooks_dir": "/workspace/.hooks",
        "repeat": 11,
        "hooks": {
          "guard": {
            "event": "PreToolUse",
            "median_ms": 16.4,
            "min_ms": 16.1,
            "max_ms": 16.7
          },
          "audit": {
            "event": "PostToolUse",
            "median_ms": 16.4,
            "min_ms": 16.2,
            "max_ms": 16.7
          },
          "context": {
            "event": "UserPromptSubmit",
            "median_ms": 16.4,
            "min_ms": 16.1,
            "max_ms": 16.6
          },
          "lang": {
            "event": "UserPromptSubmit",
            "median_ms": 16.5,
            "min_ms": 16.2,
            "max_ms": 16.6
          },
          "budget": {
            "event": "UserPromptSubmit",
            "median_ms": 16.4,
            "min_ms": 16.2,
            "max_ms": 17.3
          }
        },
        "per_tool_call_ms": 32.8,
        "per_prompt_ms": 16.5,
        "source": "docs/eval/i4/hook-bench.json",
        "说明": "hook 段在 SSE 上和模型调用 / MCP 往返混在一段里,分不开,所以是**单独直测**的(tools/eval/hook_bench.py 在 daemon 容器里按 hook 真实的调用方式跑 N 次取中位数)。一次运行要付的 hook = per_prompt_ms + 工具调用次数 × per_tool_call_ms —— 这是**推算**,不是从那次运行里量到的,报告里按推算标注。社区版没有 hook 机制,这一段不适用(不是 0,是没有这回事)。"
      },
      "明细文件": "docs/eval/i4/summary.json(含分段计时的每一次原值与 P90)",
      "注意": [
        "**首字延迟有两个**:`首字` 是第一块正文(本部署人设要求先发一行「路标」,所以它可能是进度提示不是答案);`答案首字` 是末轮第一块正文 —— 跨产品比较以**答案首字**为准。",
        "社区版的**收尾段量不到**(阻塞 POST,没有独立结束事件),写 null,上界是残差。",
        "P90 用最近秩法,不插值;n=6 时 P90 等于最大值。"
      ]
    }
  },
  "q2的两套口径": {
    "为什么": "社区版在 q2 上不稳定:论点原文只在工作区文件里,它多数时候不去读、直接回「你没存买入理由」就结束。所以这道题的达标与否取决于对照组抽到哪一类。",
    "原样口径_正式判定": {
      "hca_墙钟中位_s": 18.6,
      "社区版_墙钟中位_s": 9.4,
      "墙钟比": 1.98,
      "n": "12 / 11"
    },
    "都真做题口径": {
      "hca_墙钟中位_s": 18.6,
      "社区版_墙钟中位_s": 32.2,
      "墙钟比": 0.58,
      "n": "12(12/12 都读了) / 1(11 次里 1 次)"
    },
    "两档合并扩样本": {
      "hca_墙钟中位_s": 18.9,
      "社区版_墙钟中位_s": 32.2,
      "墙钟比": 0.59,
      "n": "24 / 5(23 次里 5 次)"
    },
    "来源": "python3 tools/eval/i2_taskdone.py docs/eval/i3 --batches opt5-fork-b-12,opt3-fork-b-重测-12"
  },
  "U-20_工具schema体积对首字延迟的影响": {
    "结论": "有因果,约 10.0 毫秒 / 千字节(R² 0.99,四档 × 每档 26 次);提示正文约 6.0 毫秒 / 千字节 —— 同一量级,所以吃时间的是「请求大小」本身。",
    "发行版→只留5个工具能省": "347 ms / 轮",
    "来源": "docs/eval/i3/schema-probe/报告.md(原始记录同目录)"
  },
  "不能放在一起比的": [
    "I1 的 q4 与 I3 / I2 / M2 的 q4(Kronos key 开没开不同)",
    "I2 报告里的 D 23.2 与 I3 的 22.2(隔天,且 I3 已用同配置重测证明差值来自漂移)",
    "A 线与 B 线(挂的 MCP 不同)",
    "I4 的绝对耗时与 I1 / I3 的绝对耗时(代码版本不同、时段不同;I4 里能比的只有同批次内两边的相对关系)",
    "I4 的「首字」与 I1 报告里那个「首字延迟」(I1 那个没有把人设的「路标」摘出去;I4 的「答案首字」才是同口径的那个)"
  ]
}