{
  "updated_at": "2026-11-03T11:35:05+08:00",
  "display_timezone": "Asia/Shanghai",
  "previous_experiment": {
    "status": "COMPLETE",
    "completed_at": "2026-11-02 20:58:58 +08:00",
    "training_rounds_completed": 21,
    "main_test_answers": 720,
    "auxiliary_test_answers": 180,
    "total_test_answers": 900,
    "gpu_stopped_at": "2026-11-02 21:01:28 +08:00",
    "results": [
      {
        "id": "N",
        "name": "原模型",
        "sampled": 27,
        "fixed": 15,
        "sampled_total": 60,
        "fixed_total": 30
      },
      {
        "id": "WT",
        "name": "Wide",
        "sampled": 30,
        "fixed": 14,
        "sampled_total": 60,
        "fixed_total": 30
      },
      {
        "id": "DT",
        "name": "Deep",
        "sampled": 30,
        "fixed": 14,
        "sampled_total": 60,
        "fixed_total": 30
      },
      {
        "id": "WDT",
        "name": "原 Wide＋Deep",
        "sampled": 27,
        "fixed": 15,
        "sampled_total": 60,
        "fixed_total": 30
      },
      {
        "id": "R",
        "name": "同题量普通训练",
        "sampled": 28,
        "fixed": 14,
        "sampled_total": 60,
        "fixed_total": 30
      },
      {
        "id": "WDT_OPT",
        "name": "改进 Wide＋Deep",
        "sampled": 30,
        "fixed": 14,
        "sampled_total": 60,
        "fixed_total": 30
      },
      {
        "id": "WDT_REPAIR",
        "name": "改进 Wide＋Deep＋纠错",
        "sampled": 34,
        "fixed": 15,
        "sampled_total": 60,
        "fixed_total": 30
      },
      {
        "id": "WDT_SCALE50",
        "name": "50 道种子版",
        "sampled": 29,
        "fixed": 13,
        "sampled_total": 60,
        "fixed_total": 30
      }
    ],
    "results_url": "final-test-summary.json",
    "analysis_url": "analysis-20261103.html"
  },
  "current_release": {
    "release": "wide_deep_fresh_revised_20261103",
    "prepared_at": "2026-11-02T17:49:07.312155+00:00",
    "status": "LOCAL_READY_WAITING_FOR_POD",
    "local_checks_passed": 179,
    "check_groups": [
      {
        "name": "主练与并行补答",
        "passed": 98,
        "scope": "实际模型更新、完整回答、共同说明和有效诊断保留"
      },
      {
        "name": "长题处理",
        "passed": 6,
        "scope": "实际分词器、旧失败答卷和四种显示方式"
      },
      {
        "name": "整套流程",
        "passed": 2,
        "scope": "三个训练组各三轮、四版本测试与完成后接续"
      },
      {
        "name": "固定材料学习比较",
        "passed": 5,
        "scope": "同一初始模型与同一批答卷的真实学习"
      },
      {
        "name": "老师与进度",
        "passed": 14,
        "scope": "临时故障、有限重试、回复恢复和调用记账"
      },
      {
        "name": "启动与退出",
        "passed": 8,
        "scope": "重复启动、安装、空间不足、中断与无进展退出"
      },
      {
        "name": "方法与纠错",
        "passed": 46,
        "scope": "真实验题、回访、有效改正与保存后接续"
      }
    ],
    "final_package_install": "PASS",
    "gpu_started": false,
    "formal_new_results": false,
    "archive_sha256": "ce151de771a88217290b10ae02aeaa09710f9c3b912960eae4f506c2c865db54",
    "code_sha256": "96e38f3a931cc961b3d6066d07fd2487a3769914cb1d3edf60d47c2ce0805262",
    "controlled_workflow_records_are_experiment_results": false
  },
  "next_experiment": {
    "arms": [
      "N",
      "R",
      "WDT_OPT",
      "WDT_REPAIR"
    ],
    "seed_questions": 20,
    "rounds": 3,
    "answers_per_question": 8,
    "question_cap_each_round": 60,
    "final_test_questions": 30,
    "final_test_answers": 360,
    "deferred_test_questions": 20,
    "automatic_deferred_tests": false,
    "new_main_updates_max": 9,
    "repair_updates_max": 3,
    "automatic_fixed_batch_comparisons": false
  },
  "time_and_cost_estimate": {
    "planning_hours": 13,
    "planning_range_hours": [
      8,
      31
    ],
    "historical_gpu_usd_per_hour": 1.79,
    "planning_gpu_usd": 23,
    "planning_gpu_range_usd": [
      15,
      55
    ],
    "teacher_and_storage_costs_separate": true,
    "new_gpu_rate_verified": false,
    "new_parallel_speedup_assumed": false
  },
  "changes": [
    {
      "title": "说明一致，长题也不丢信息",
      "problem": "材料规则没有说透时，答错可能来自理解方式不同；长题切换显示方式也可能漏说明。",
      "change": "学生、老师和普通训练组使用共同说明；长题仍保留完整动作、库存与反馈。",
      "verified": "本地用旧长题失败记录检查了四种显示方式，内容可完整还原。"
    },
    {
      "title": "老师漏分一条，不再丢掉整批诊断",
      "problem": "扩大组第二轮漏了一条分组，55 份有效诊断没有进入出题。",
      "change": "保留已经分好的有效诊断，遗漏项单独处理，让其余题继续进入课程。",
      "verified": "本地回放了原来的遗漏情况，其他有效诊断保留下来。"
    },
    {
      "title": "八份回答都参与下一步选题",
      "problem": "只看第一份回答，会漏掉其他回答已经暴露的失败。",
      "change": "第一份先给老师；八份补齐后，完整成败和同能力历史表现共同决定 Wide、Deep 的安排。",
      "verified": "本地检查覆盖首份成功、后续失败的情况，以及同能力跨题成绩的使用。"
    },
    {
      "title": "降难保留卡点，再接回原难题",
      "problem": "书架题缺造纸练习时直接送纸，或者已经会造纸却仍不会备齐数量，都需要更具体的过渡。",
      "change": "简化其他负担，保留当前不会的必要步骤；简单题成功后优先回原题，仍不会就补数量和制作次序的过渡。",
      "verified": "本地检查覆盖卡点保留、原题回访和过渡安排；可完成、能力必要、难度方向的验题继续保留。"
    },
    {
      "title": "全对但一路犯错，暂缓升难",
      "problem": "纠错版 34 份随机成功中，25 份仍出现被环境拒绝的动作；做成不等于过程稳定。",
      "change": "从已有答卷记录重复失败和恢复情况；八份全对但过程出错的题优先回访，暂缓升难。",
      "verified": "本地检查确认了回访优先级。这是课程安排的变化，八份仍全对时，主练本身仍没有成败差异。"
    },
    {
      "title": "纠错必须保留真实错误及后果",
      "problem": "错误发生前重写，与错误发生后现场补救，学到的能力不同。",
      "change": "现场补救保留至少一次实际失败和后果；回原题执行到成功，只学习改过、执行过且帮助原目标的动作。",
      "verified": "本地检查覆盖原题执行和动作筛选；反思文字、未改动作及无关产物不作为学习目标。"
    },
    {
      "title": "把等待压短，保存关键产出",
      "problem": "学生答完等待老师、重复准备和过细保存，会占用付费显卡时间。",
      "change": "老师工作时继续补齐旧题；作答最多并行八份，训练按显存余量安排。同时只驻留一个模型，只保存题目、完整答卷、关键结果和阶段模型。",
      "verified": "本地检查覆盖并行补答和完整流程。新题仍要等两组计划和匹配题量确定后再做；没有额外摸底、逐动作保存或反复扫描历史。"
    },
    {
      "title": "中断接续，不再重复学同一批",
      "problem": "模型已保存、汇总还没写完时重启，可能重复更新；老师正常工作也可能被误当作无进展。",
      "change": "直接接回已完成的模型；老师临时故障最多尝试三次，完成回复时记录进度。内部调用额度只记账。",
      "verified": "本地真实学习检查了保存、重载和继续学习的一致性；整套流程完成后重启，没有新增工作。"
    }
  ],
  "optional_comparison": {
    "implemented": true,
    "automatically_started": false,
    "uses_existing_answers": true,
    "new_teacher_calls": 0,
    "options": [
      "原来计算方式",
      "长短回答等权",
      "取消成功答卷中同库存重复被拒动作的正向学习项"
    ],
    "formal_results": null
  },
  "remaining_questions": [
    "全对但过程错误的题，主练仍无成败差异；回访不等于新增了学习反馈。",
    "现有主练仍按整题成败学习，长失败和正确前缀受到的影响需要单独比较。",
    "普通成功动作与纠错动作的等学习量对照仍未实现。",
    "新版真实显卡速度、首次实际学习和独立测试成绩等待正式实验。"
  ],
  "update_policy": "完成一轮、形成一批测试结果、确认原因或冻结修改后，核对记录并更新此页。",
  "background_auto_publication": false
}