Add 162 trajectories (phase 2, remaining runs; batch S10)

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-privileged--84xaxhg.json

@@ -0,0 +1,222 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-privileged--84xaxhg",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 990,
+   "sim_steps": 990,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 464,
+   "agent_resets": 26453,
+   "agent_sim_steps": 10412,
+   "agent_records": 0,
+   "agent_replays": 10,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 18.162,
+   "peak_processes": 32,
+   "peak_disk_gb": 0.987,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-privileged--84xaxhg/replay.mp4",
+  "duration_s": 39.64,
+  "fps": 25.0,
+  "frames": 991,
+  "sha256": "dc36673849e3ca9e38f112b0bb59f7ce43da1f0b639348b1b8f1170344509d3a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 19,
+      "text_zh": "观察：log step 19 所读模型将 ball1 绑定黄色材质、ball2 绑定粉色材质；本次视频 0.0 s 右侧近景可见黄色球在掌内靠拇指侧，粉色球靠其外侧，以下按此固定身份追踪，不按变色验收。"
+     },
+     {
+      "t": 4.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–9.5 s 每隔约0.5 s 的近景显示，MyoHand起初轻微屈指、两球沉降，随后黄色球基本停在拇指与掌面之间；小目标标记继续绕行，视频 4.0–6.5 s 可见标记移到粉色球一侧而黄色球未随行。"
+     },
+     {
+      "t": 39.6,
+      "log_step": null,
+      "text_zh": "推断：视频 10.0–39.0 s 逐秒近景及视频 38.6–39.6 s 末段显示黄色球始终留在同一掌内区域，无持续绕行，因此不存在连续200个控制步的完整目标跟踪窗口，终态也未满足；结论依据球与移动标记的相对运动，而非仅由总评分为零推得。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "托/舀",
+      "note_zh": "观察：视频 0.0–2.0 s MyoHand屈指、掌面和拇指侧承托黄色球1；视频 2.0–39.6 s 球仍由同一手的指掌承托，未见持续滚动换位。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 4.0,
+      "log_step": 19,
+      "text_zh": "观察：log step 19 的模型材质定义将粉色球标为 ball2；视频 0.0–9.5 s 的密集近景中，粉色球起初随手部沉降，随后停在指根附近。视频 4.0–7.5 s 的小目标标记从球旁经过、跨过其投影，粉色球未跟着标记在掌内绕转。"
+     },
+     {
+      "t": 39.6,
+      "log_step": null,
+      "text_zh": "推断：视频 10.0–39.0 s 的逐秒画面与视频 38.6–39.6 s 的密集末段中，粉色球一直位于黄色球的同一侧，缺少连续一整圈所需的换位，故球2持续跟踪节点从未达到且终态为假；球或标记的颜色变化不足以替代三维距离判定。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "托/舀",
+      "note_zh": "观察：视频 0.0–2.0 s 粉色球2随MyoHand指掌姿态变化落稳在指根附近，视频 2.0–39.6 s 持续由同一手的指掌承托，未见将其绕掌输送的手内调整。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 2.0,
+   "log_step": 189,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 2.0 s MyoHand已将两球托稳，随后视频 2.0–9.5 s 目标标记绕行而球1停在掌内；这一偏离持续到视频 39.6 s，球2也并行未能持续跟踪。log step 189 将最终交付改为恒定肌肉命令，log step 202 ��该990行文件的两次回放均为 success=false、instance_ok=true、deterministic=true；断点以并行失败的球1跟踪步骤为代表，未把此前调试方案的掉球算入本次轨迹。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 4.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 2.0–9.5 s，MyoHand保持近似固定托球姿势，两球中心基本不换位，小目标标记却持续绕行；视频 4.0–6.5 s 标记经过粉色球附近而黄色球仍停在拇指侧，之后视频 10.0–39.6 s 重复这一停球、走标记的状态。机器人没有跟上移动目标。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 189 的最终生成脚本默认用 np.tile(u0, (T, 1)) 重复同一肌肉命令，替换了未能稳定运行的脉冲控制；log step 202 重新生成并两次回放验证的是这一恒定命令文件，均未成功。视频 2.0–39.6 s 对应表现为托稳两球但不使其持续换位，说明交付的肌肉控制只能维持托持，未实现任务所需的掌内绕转；现有证据不足以细分为接触几何、逆解或反馈参数问题。"
+   }
+  ],
+  "gap_zh": "观察：本目录有完整视频与运行日志，但没有逐控制步状态日志；run.json 的 reward 未报告 target1_dist、target2_dist 或 solved_run，故三项原生指标均填 null。目标标记很小且部分被指掌和球遮挡，无法恢复各球精确三维误差及瞬间合格窗口；两球缺少整圈换位仍足以判定持续跟踪未完成。log step 203 所称约19步合格是agent自述，未作为节点到达时间；其“不可达”判断也未作为任务或环境根因。",
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  }
+ },
+ "summary_zh": "观察：本次视频 0.0–39.6 s 中MyoHand将双球托在掌内，起初轻微沉降后基本停位，两项持续跟踪节点均未完成；最终文件在 log step 202 两次一致回放中均失败。推断：断点取并行失败的球1持续跟踪（操作中），表现为双球未跟上移动目标，交付的恒定肌肉命令未形成持续掌内绕转；缺少逐步三维误差，不能细分控制根因或精确计量短暂合格窗口。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两节点均未到达、final=false 与视频一致（球全程停在掌内，无绕转）；断点、现象、根因有据，未见需改之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:02:32Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--84xaxhg/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--84xaxhg/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c536c79b1d916efc3e94db1e799b1464a66ef7959b9cdc242c1765884f4fb829",
+   "video_sha256": "dc36673849e3ca9e38f112b0bb59f7ce43da1f0b639348b1b8f1170344509d3a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-privileged--aab3yvm.json

@@ -0,0 +1,225 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-privileged--aab3yvm",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 460,
+   "sim_steps": 460,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 194,
+   "agent_resets": 13859,
+   "agent_sim_steps": 12583,
+   "agent_records": 44,
+   "agent_replays": 47,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 4.731,
+   "peak_processes": 14,
+   "peak_disk_gb": 0.266,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-privileged--aab3yvm/replay.mp4",
+  "duration_s": 18.44,
+  "fps": 25.0,
+  "frames": 461,
+  "sha256": "0076cc5c7027a247666864a1df832a4ca375a7ede879f8da29f7b5e94a0429a6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–7.75 s 的右侧近景中，MyoHand调整腕部及手指后，两球仍保持一球在画面左上、另一球在右下的相对布局，仅有小幅位移；小目标标记继续绕行，两球没有随之绕转换位。对object1的持续跟踪不能由短暂靠近或球体变色认定。"
+     },
+     {
+      "t": 18.4,
+      "log_step": null,
+      "text_zh": "推断：视频 8.00–18.40 s（含视频 17.40、17.60、17.80、18.00、18.20、18.40 s 末段近景）仍显示相同的双球布局，目标标记周期性移到球外侧；相比成功参考视频中双球随目标交换位置，本次object1没有覆盖完整目标绕行所需的连续200步跟踪窗口，因此本节点从未成立且终值为false。此判断基于运动不匹配，不是仅由总评分为0反推。"
+     },
+     {
+      "t": 18.4,
+      "log_step": 376,
+      "text_zh": "观察：log step 374恢复最终460步交付文件，log step 376对该文件的Session.replay返回steps=460、success=false、done=false、solved_run=0、hold_steps=200；log step 377官方评分器两次回放均success=false且终态哈希相同。共同计数只支持整项未完成，不单独替代逐球视频判断。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.00–18.40 s，MyoHand同一只手的手指及掌部持续接触并留住第1颗球，伴随腕指小幅运动在手内调整球的位置；未见计划松手，实际未形成随目标绕转换位。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–7.75 s 的右侧近景中，MyoHand调整腕部及手指后，两球仍保持一球在画面左上、另一球在右下的相对布局，仅有小幅位移；小目标标记继续绕行，两球没有随之绕转换位。对object2的持续跟踪不能由短暂靠近或球体变色认定。"
+     },
+     {
+      "t": 18.4,
+      "log_step": null,
+      "text_zh": "推断：视频 8.00–18.40 s（含视频 17.40、17.60、17.80、18.00、18.20、18.40 s 末段近景）仍显示相同的双球布局，目标标记周期性移到球外侧；相比成功参考视频中双球随目标交换位置，本次object2没有覆盖完整目标绕行所需的连续200步跟踪窗口，因此本节点从未成立且终值为false。此判断基于运动不匹配，不是仅由总评分为0反推。"
+     },
+     {
+      "t": 18.4,
+      "log_step": 376,
+      "text_zh": "观察：log step 374恢复最终460步交付文件，log step 376对该文件的Session.replay返回steps=460、success=false、done=false、solved_run=0、hold_steps=200；log step 377官方评分器两次回放均success=false且终态哈希相同。共同计数只支持整项未完成，不单独替代逐球视频判断。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.00–18.40 s，MyoHand同一只手的手指及掌部持续接触并留住第2颗球，伴随腕指小幅运动在手内调整球的位置；未见计划松手，实际未形成随目标绕转换位。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": 0
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 1.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.75–2.00 s，机器人已开始腕指操作并留住双球，但球主要停在手中原来的相对位置，未随继续绕行的目标换位；视频 2.00–18.40 s 未修复这种运动不匹配。两节点并行操作且均未完成，此处以球1持续跟踪作为共同跟踪失败的断点锚点，不表示球2先完成。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 1.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 1.00–7.75 s及视频 8.00–18.40 s，MyoHand上的双球基本维持并排布局，而小目标标记持续绕行并反复离开球心所在位置；机器人没有让双球跟上移动目标。视频末段双球仍在手上，没有可见掉落，失败表现为未跟上目标运动。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 321的生成脚本用周期函数设定腕部、拇指及其余手指关节目标，再以PD力矩和肌肉力矩映射求肌肉命令；log step 374、376–377确认最终选取并重复回放的是460步肌肉命令文件。结合视频 1.00–18.40 s中双球被留在手上却未随目标绕行，说明此次腕指协调控制未产生所需的球体跟踪运动；现有记录不足以定位到某个控制参数、接触几何或逆解子问题。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志；最终交付文件的回放记录（log step 376–377）未报告两颗球各自的距离历史，视频透视及局部遮挡也不能精确量出三维0.015米误差或阈值越界时刻。可由全段缺少绕转换位判定两节点均未形成连续200步窗口，但不能细分控制根因；native_metrics的两个距离值因此为null，solved_run=0取自log step 376的最终文件回放结果。"
+ },
+ "summary_zh": "本次460步评分轨迹中，MyoHand持续留住双球并进行小幅腕指调整，但双球基本维持并排布局，没有像成功参考视频那样随固定配对目标绕转换位，两个持续跟踪节点均未完成。断点记为球1持续跟踪的操作中，现象为未跟上移动目标，根因归入控制；log step 376–377确认最终交付回放失败且可确定复现。缺少逐球距离和保持历史，无法精确定位阈值越界时刻或进一步细分控制原因。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "两节点未成立、终值false与视频一致（双球全程停在手上、布局基本不变，无掉落），断点、动态时机现象和控制根因（停在父类）有证据支持，未发现需要改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:56:28Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--aab3yvm/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--aab3yvm/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "863deb7a904d1b7acab8896b634ff18067a6ebb6c40121c8836b631e73010c29",
+   "video_sha256": "0076cc5c7027a247666864a1df832a4ca375a7ede879f8da29f7b5e94a0429a6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-privileged--pqfxjpu.json

@@ -0,0 +1,237 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-privileged--pqfxjpu",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 155,
+   "sim_steps": 155,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 54,
+   "agent_resets": 276,
+   "agent_sim_steps": 4921,
+   "agent_records": 0,
+   "agent_replays": 3,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 1,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.0,
+   "peak_processes": 0.0,
+   "peak_disk_gb": 0.0,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-privileged--pqfxjpu/replay.mp4",
+  "duration_s": 6.24,
+  "fps": 25.0,
+  "frames": 156,
+  "sha256": "2172bd028f4f50589dbc4ca1cb14875291446f9035171b8a2b3483588cc30862"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.64,
+      "log_step": 18,
+      "text_zh": "观察：视频0.00–1.64 s，MyoHand屈伸手指、转动手腕，双球留在指掌接触区。log step 18的所交付轨迹生成记录在循环索引40报告球1误差0.0077 m、球2误差0.0111 m，共同连续合格计数36；这只是短时合格，尚未达到200步。"
+     },
+     {
+      "t": 2.04,
+      "log_step": 18,
+      "text_zh": "推断：log step 18的循环索引50是第51次控制，映射视频51/25=2.04 s；该处报告球1误差0.0157 m，之后抽样值持续超阈，至循环索引140为0.0768 m。视频2.04、2.44、3.24、4.04 s可见手指继续调整球；这些动作未使球1持续跟上移动目标。上述映射已用第155次控制掉球对应视频6.20 s末帧核对，不使用agent工作时间。"
+     },
+     {
+      "t": 6.2,
+      "log_step": 30,
+      "text_zh": "推断：评分器字段n_actions=sim_steps=155，log step 23保存的155行轨迹在log step 30两次回放均于第155个控制步失败、终态哈希一致；录像共156帧（含初态），末帧视频6.20 s可见掉球，与此吻合。每控制步对应一帧，155步不足节点要求的连续200步，因此本回合该节点从未成立，不能把短暂贴近目标记为完成。"
+     },
+     {
+      "t": 6.2,
+      "log_step": 18,
+      "text_zh": "观察：视频5.92–6.20 s，双球向指掌支承区外滑离，左侧视角可见它们与手明显分开。log step 18的末次循环索引154报告球1高度1.224 m、误差0.2609 m，低于掉球高度1.25 m；其末态哈希与log step 23、30的交付回放相同。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "视频0.00–5.92 s可见单只MyoHand的手指屈伸、配合掌面接触并调整掌中球1的位置，末段球滑离支承区。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.64,
+      "log_step": 18,
+      "text_zh": "观察：视频0.00–1.64 s，MyoHand在同一只手的指掌间同时调整两球。log step 18对应交付轨迹的循环索引40报告球2误差0.0111 m、共同连续合格计数36，尚未达到200步。"
+     },
+     {
+      "t": 2.04,
+      "log_step": 18,
+      "text_zh": "观察：log step 18在循环索引50报告球2误差0.0185 m；循环索引70、80又分别为0.0132、0.0128 m，但末次循环索引154为0.1226 m、高度1.343 m。视频2.04–5.92 s手指继续在掌中调整双球，视频6.00–6.20 s可见双球滑出指掌支承区；未将球的颜色作为球2自身距离达标的证据。"
+     },
+     {
+      "t": 6.2,
+      "log_step": 30,
+      "text_zh": "推断：评分器字段n_actions=sim_steps=155，log step 23保存的155行轨迹在log step 30两次回放均于第155个控制步失败、终态哈希一致；录像共156帧（含初态），末帧视频6.20 s可见掉球，与此吻合。每控制步对应一帧，155步不足节点要求的连续200步，因此本回合该节点从未成立，不能把短暂贴近目标记为完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "视频0.00–5.92 s可见同一只MyoHand以屈伸的手指和掌面同时接触、滚动调整球2，末段未能留住球。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 2.04,
+   "log_step": 18,
+   "where": "操作中",
+   "evidence_zh": "推断：MyoHand正在并行调整双球以跟踪移动目标。log step 18对应交付轨迹在循环索引40尚报告两球误差合格，循环索引50球1误差已为0.0157 m，此后抽样误差持续增大，末次球1高度降至1.224 m；log step 23保存该轨迹，log step 30两次回放均在155步失败且终态哈希一致。视频1.60–2.40 s可见手指持续调整但没有完成绕转，视频5.92–6.20 s可见掉球，故将球1持续跟踪操作中的未修复失跟定位为断点，视频2.04 s为粗略导航锚点。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 2.04,
+    "log_step": 18,
+    "evidence_zh": "推断：移动目标继续绕行时，MyoHand未能让球1同步跟上。log step 18记录循环索引40球1误差0.0077 m，循环索引50为0.0157 m，循环索引80为0.0306 m、100为0.0409 m；对应视频1.64、2.04、3.24、4.04 s，手指仍在调整球，球集中于指掌一侧，没有随目标完成持续绕转。此处是运动目标的持续失跟。"
+   },
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 6.0,
+    "log_step": 18,
+    "evidence_zh": "观察：视频5.20–5.92 s双球仍处指掌附近，视频6.00–6.20 s它们滑离原支承区并与手分开，MyoHand没有重新留住球。log step 18末次记录球1高度1.224 m低于1.25 m，log step 23、30对所交付文件的回放均报告掉球失败；这发生在继续掌内操作时。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 16–18的脚本以两球目标误差为代价搜索手部关节设定，再用肌肉控制层执行；log step 18显示交付轨迹在短时合格后误差持续扩大，log step 23、30确认原执行与交付回放终态一致。视频2.04–6.20 s中手部持续动作却未维持双球绕行，最终失去支承，支持肌肉与手指协调控制不足；现有证据不能把它进一步归为特定接触布局、反馈律或参数错误。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志，log step 18仅每隔若干控制步输出误差，无法确定两颗球首次越阈的精确先后或重建所有瞬时接触；断点以球1持续失跟并最终掉落为依据。小目标标记受手指及球遮挡，二维画面不能独立测定0.015 m三维阈值，球2变色也不能证明自身距离合格。评分器reward未单列target1_dist、target2_dist、solved_run，原生指标按要求保留null，交付执行日志的数值仅写入证据；根因停留在控制大类。"
+ },
+ "summary_zh": "MyoHand确实通过指掌接触同时调整双球，但交付只有155个控制步，两个要求连续200步的跟踪节点均未达成，终态为0/2。球1在视频约2.04 s已失跟且未完成修复，视频6.00–6.20 s双球滑离支承区，球1掉落触发失败；交付回放一致，证据支持控制执行不足。缺少逐控制步状态及评分器原生距离字段，无法精确区分两球首次越阈的先后与更细控制根因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "两节点均未达成（0/2）、时钟映射（控制步/25）与末帧掉球视频6.20 s吻合，断点在球1失跟、现象与根因均有log step 18/23/30及视频支持，未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:00:44Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--pqfxjpu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--pqfxjpu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "55fb6b1c6c05182df464977a31cb7d10d9d2da4942fe4a31693bebe57bbae8eb",
+   "video_sha256": "2172bd028f4f50589dbc4ca1cb14875291446f9035171b8a2b3483588cc30862",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-privileged--rrazfuz.json

@@ -0,0 +1,218 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-privileged--rrazfuz",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 400,
+   "sim_steps": 400,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 47,
+   "agent_resets": 9576,
+   "agent_sim_steps": 7074,
+   "agent_records": 9280,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.646,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.837,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 1,
+   "session_log_from_monitor": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-privileged--rrazfuz/replay.mp4",
+  "duration_s": 16.04,
+  "fps": 25.0,
+  "frames": 401,
+  "sha256": "77ff87080f9802e303bfe6c99c7cef4d8781812b7e0b79dc658f8f930483d197"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 4.0,
+      "log_step": 206,
+      "text_zh": "观察：视频 0.0–16.0 s 的两视角及右侧近景加密帧显示，MyoHand屈指承托两球并摆动腕部，但两球在掌内大体保持原来的相对位置；视频 2.0–4.0 s、10.0–12.0 s 可见小目标标记绕至球侧而球未随之绕转换位。log step 205 保存ES9检查点生成的400步动作；log step 206 对该次生成记录的球1目标距离，在控制步100、150、200、250、300、350、400分别为0.0454、0.0442、0.0181、0.0200、0.0457、0.0446、0.0179米，全部高于0.015米。"
+     },
+     {
+      "t": 16.0,
+      "log_step": 206,
+      "text_zh": "推断：交付记录（log step 205–206）与评分器n_actions=sim_steps=400、deterministic=1一致；视频含初始帧及400个动作帧，25 fps，故控制步100/25对应视频 4.0 s、控制步400/25对应视频 16.0 s，两处均已核对球仍在掌中且与小目标标记相离。上述每隔50控制步的越界采样排除了全程任何连续200步合格窗口，因此球1持续跟踪从未到达、终态为false；不由总失败分数或球体变色单独反推。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.0–0.6 s、4.8–6.4 s、12.8–16.0 s，单只MyoHand的屈曲手指配合掌面接触并约束掌中object1，随腕部摆动做小幅滚动、换位；没有把球松开后重新抓取，实际调整未形成所需持续绕转。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 4.0,
+      "log_step": 206,
+      "text_zh": "观察：视频 0.0–16.0 s 的两视角及右侧近景加密帧显示，MyoHand屈指承托两球并摆动腕部，但两球在掌内大体保持原来的相对位置；视频 2.0–4.0 s、10.0–12.0 s 可见小目标标记绕至球侧而球未随之绕转换位。log step 205 保存ES9检查点生成的400步动作；log step 206 对该次生成记录的球2目标距离，在控制步100、150、200、250、300、350、400分别为0.0470、0.0468、0.0180、0.0207、0.0468、0.0455、0.0166米，全部高于0.015米。"
+     },
+     {
+      "t": 16.0,
+      "log_step": 206,
+      "text_zh": "推断：交付记录（log step 205–206）与评分器n_actions=sim_steps=400、deterministic=1一致；视频含初始帧及400个动作帧，25 fps，故控制步100/25对应视频 4.0 s、控制步400/25对应视频 16.0 s，两处均已核对球仍在掌中且与小目标标记相离。上述每隔50控制步的越界采样排除了全程任何连续200步合格窗口，因此球2持续跟踪从未到达、终态为false；不由总失败分数或球体变色单独反推。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.0–0.6 s、4.8–6.4 s、12.8–16.0 s，单只MyoHand的屈曲手指配合掌面接触并约束掌中object2，随腕部摆动做小幅滚动、换位；没有把球松开后重新抓取，实际调整未形成所需持续绕转。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 2.0,
+   "log_step": 206,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.4–4.0 s，MyoHand继续接触承托两球并摆动手腕，小目标标记继续绕行，两球却仍并排停留在原来的指掌区域；视频 10.0–12.0 s 再次出现同样偏离。log step 206 记录最终交付轨迹的控制步100、150两球误差均超过0.04米，最长共同合格仅33控制步。推断：两项并行跟踪均在操作中失败且之后未完成200步窗口，以球1节点作断点导航，球2同时未达，不表示两球有先后顺序。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 2.0,
+    "log_step": 206,
+    "evidence_zh": "观察：视频 1.4–6.4 s 及9.6–14.4 s，MyoHand保持两球在指掌间，两球主要随腕部整体移动，小目标标记则绕到球的侧面、下方再绕回，球没有持续跟上移动标记；视频 15.0–16.0 s 仍是相同掌内布局。log step 206 对最终保存动作记录最长共同合格33控制步、最终solved_run=0。推断：可见缺陷是对周期移动目标的跟随不足，属动态时机。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 205,
+    "evidence_zh": "观察：log step 158 的评估器以两球目标误差、速度、手关节、肌肉激活和目标相位计算肌肉动作；log step 165、173、185、205沿用该评估器并最终保存ES9生成的400步动作。log step 206 在这次生成中记录两球误差周期性增至约0.045米，最长共同合格33控制步；视频 2.0–6.4 s、10.0–14.4 s 对应地显示手腕在动而两球未完成掌内绕转。推断：该肌肉控制策略能保住球，却未产生追随目标所需的连续掌内运动，归为控制；现有证据不能进一步区分接触设计、轨迹与反馈环节。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志或接触力、肌肉响应记录；目标标记较小且部分被球与手指遮挡，不能从二维画面独立复核每步三维距离或精确定位首次越界。节点未达由视频的缺失绕转与最终交付日志的密集越界采样共同确定，精确失败时刻仅用约视频 2.0 s 导航；根因仅判到控制大类。评分器未报告DAG三个原生指标的数值，native_metrics的value均为null。"
+ },
+ "summary_zh": "交付的400控制步轨迹中，MyoHand持续接触承托并小幅调整两球，但两球主要随腕部整体摆动，未随各自移动目标持续绕转，两个跟踪节点均未达到连续200步要求，终态0/2。断点记球1持续跟踪（两节点并行失败），现象为动态时机，根因为控制策略未产生足够掌内运动；log step 206记录最长共同合格仅33控制步，评分器回放失败且未触发掉球终止。缺少逐步状态与接触力记录，不能细分控制根因；成功参考视频中明显的掌内绕转换位仅用于外观比较。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个跟踪节点未达、终态0/2与评分器success=0、log step 206的7个采样点（全部>0.015 m，最长共同合格33步）和视频帧一致；断点、现象、根因有证据支持，未发现需要修改之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:58:26Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--rrazfuz/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--rrazfuz/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "5e1ecc58b95953f1241508ab832ff801b92889448b13a1b144c0b50d16bb82bb",
+   "video_sha256": "77ff87080f9802e303bfe6c99c7cef4d8781812b7e0b79dc658f8f930483d197",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-privileged--x2hpb7x.json

@@ -0,0 +1,228 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-privileged--x2hpb7x",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 62,
+   "agent_resets": 65,
+   "agent_sim_steps": 7166,
+   "agent_records": 0,
+   "agent_replays": 7,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 2.074,
+   "peak_processes": 15,
+   "peak_disk_gb": 0.291,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-privileged--x2hpb7x/replay.mp4",
+  "duration_s": 40.04,
+  "fps": 25.0,
+  "frames": 1001,
+  "sha256": "c022a85bc94066a0767cad23943c94c17cab6b53fbc744b2758d3a1797f1714a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–3.0 s，MyoHand收拢手指，两球在指掌间移动；球1（黄色球）随后停留在掌中一侧。起始位置或短暂靠近目标不构成连续200步跟踪。"
+     },
+     {
+      "t": 8.0,
+      "log_step": null,
+      "text_zh": "观察：视频 3.0–8.0 s 的近景及视频 9.0–40.0 s 每秒帧显示，球1仅小幅移动，未随移动标记绕掌持续换位；末秒视频 39.04–40.0 s 仍如此。成功参照视频 0.0–8.0 s 则能看见双球沿指掌持续绕转，该运动线索在本次视频中缺失。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 88,
+      "text_zh": "推断：本次视频可见球1长期近乎停在掌中，结合任务要求跟随绕行目标连续200步，判定该节点未完成。log step 87 保存最终轨迹，log step 88 对该文件两次回放均返回 success=false、steps=1000，原因均为目标轨迹耗尽，与视频一致；未把会话其他候选的数值当作该球的达成记录。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.2–3.0 s，MyoHand单手屈伸手指，配合掌面使球1（黄色球）在手内滚动换位；视频 4.0–40.0 s 主要围托球并作小幅调整，未形成持续绕转。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–3.0 s，MyoHand收拢手指，两球在指掌间移动；球2（橙色球）随后停留在掌中一侧。起始位置或短暂靠近目标不构成连续200步跟踪。"
+     },
+     {
+      "t": 8.0,
+      "log_step": null,
+      "text_zh": "观察：视频 3.0–8.0 s 的近景及视频 9.0–40.0 s 每秒帧显示，球2仅小幅移动，未随移动标记绕掌持续换位；末秒视频 39.04–40.0 s 仍如此。成功参照视频 0.0–8.0 s 则能看见双球沿指掌持续绕转，该运动线索在本次视频中缺失。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 88,
+      "text_zh": "推断：本次视频可见球2长期近乎停在掌中，结合任务要求跟随绕行目标连续200步，判定该节点未完成。log step 87 保存最终轨迹，log step 88 对该文件两次回放均返回 success=false、steps=1000，原因均为目标轨迹耗尽，与视频一致；未把会话其他候选的数值当作该球的达成记录。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.2–3.0 s，MyoHand单手屈伸手指，配合掌面使球2（橙色球）在手内滚动换位；视频 4.0–40.0 s 主要围托球并作小幅调整，未形成持续绕转。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 3.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 2.0–4.0 s，MyoHand已收拢并接触双球，但球的位置逐渐停在掌中两侧；之后直到视频 40.0 s 都未出现持续绕行。推断：双球两个并行跟踪步骤共同受阻，以球1持续跟踪作为断点定位，球2也未完成；视频 3.0 s 为停滞的粗略导航锚点，不是精确距离越界时刻。log step 88 对最终文件的两次回放均在1000步目标轨迹耗尽时判失败。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 3.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.0–3.0 s 双球随手指动作短暂移位；视频 3.0–8.0 s 及视频 9.0–40.0 s 双球基本保持各自的掌中位置，而小目标标记继续变化位置，球未跟随目标持续绕转。推断：这是对移动目标没有跟上的表现；末秒视频 39.04–40.0 s 仍有手掌承托双球，失败不表现为掉球。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 87,
+    "source_zh": "日志",
+    "evidence_zh": "观察：log step 78 的搜索脚本使用固定姿态、衰减瞬态和周期肌肉命令生成候选动作；log step 83 输出 best5 候选，log step 84–87 将该参数写入生成脚本，修复语法后成功生成并保存最终轨迹。log step 88 对最终文件两次确定性回放均失败；视频 3.0–40.0 s 显示其肌肉命令主要形成静态围托，未产生双球所需的连续手内绕行动作。推断：交付的肌肉控制序列能够留住球，却未实现目标轨迹跟踪；证据不足以进一步区分接触设计、轨迹求解或具体参数问题，停在控制层。"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志，评分器reward也未提供target1_dist、target2_dist和solved_run数值，三个原生指标均填null；不能从二维投影精确恢复三维0.015米误差或某次瞬间越界时刻，也不以球体变色代替验收。两个节点的false依据是全程可见的持续绕行缺失，配合最终文件回放失败，而非仅依据零分；断点两个并行步骤同步受阻，选球1定位不代表已证实它先于球2失败。"
+ },
+ "summary_zh": "MyoHand在评分视频中收拢手指并留住双球，但两球随后基本停在掌中，两个持续跟踪节点均未完成，终态为0/2。断点定位球1持续跟踪的操作中，球2同步未完成；现象是未跟上移动目标，交付的周期肌肉控制序列未能产生持续手内绕转，最终文件两次回放均因1000步目标轨迹耗尽失败（log step 88）。没有逐步三维距离与连续计数记录，不能精确定位阈值越界或进一步细分控制根因。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "两节点未完成、断点、现象（动态时机，没跟上）和控制根因与视频及log step 88的两次回放失败一致；未发现需要修改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:56:40Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--x2hpb7x/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--x2hpb7x/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7a057f0a16ba487103ef01f42e33d30de9d4828055b92b46bbdf9c06645cbcc9",
+   "video_sha256": "c022a85bc94066a0767cad23943c94c17cab6b53fbc744b2758d3a1797f1714a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-privileged--y8ssqwd.json

@@ -0,0 +1,219 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-privileged--y8ssqwd",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 82,
+   "sim_steps": 82,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 32,
+   "agent_resets": 25738,
+   "agent_sim_steps": 1375,
+   "agent_records": 0,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.833,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.244,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-privileged--y8ssqwd/replay.mp4",
+  "duration_s": 3.32,
+  "fps": 25.0,
+  "frames": 83,
+  "sha256": "b5bcad7b5c6aa879cb36d9af3806a2593b223137da7b5b66994ff5a27eade25a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.8,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–3.04 s，MyoHand的手指弯曲，黄色球1与另一球在指掌内移动；视频3.12–3.28 s，两球迅速分离并离开手指围成的支承区域。已检查本次全部83帧，包含最后一秒的逐帧画面。"
+     },
+     {
+      "t": 3.28,
+      "log_step": 63,
+      "text_zh": "推断：评分器reward.n_actions与sim_steps均为82；log step 63中agent脚本明确生成并保存形状为(82,39)的失败轨迹，与本视频的短回合及末段脱手一致。球1所需的连续200控制步窗口不可能在仅82控制步的交付回合内形成，故全程无达标区间、终态为false；此判断不以总分0或球体变色代替判据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.40–3.04 s，MyoHand单手的弯曲手指围住掌内的黄色球1，随手指屈曲使其在指掌间滚动、换位；视频3.12–3.28 s最终脱手。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.8,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–3.04 s，MyoHand的手指弯曲，粉褐色球2与另一球在指掌内移动；视频3.12–3.28 s，两球迅速分离并离开手指围成的支承区域。已检查本次全部83帧，包含最后一秒的逐帧画面。"
+     },
+     {
+      "t": 3.28,
+      "log_step": 63,
+      "text_zh": "推断：评分器reward.n_actions与sim_steps均为82；log step 63中agent脚本明确生成并保存形状为(82,39)的失败轨迹，与本视频的短回合及末段脱手一致。球2所需的连续200控制步窗口不可能在仅82控制步的交付回合内形成，故全程无达标区间、终态为false；此判断不以总分0或球体变色代替判据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.40–3.04 s，MyoHand单手的弯曲手指围住掌内的粉褐色球2，随手指屈曲使其在指掌间滚动、换位；视频3.12–3.28 s最终脱手。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 3.12,
+   "log_step": 63,
+   "where": "操作中",
+   "evidence_zh": "观察：视频3.04 s时两球仍在弯曲手指附近，视频3.12–3.28 s黄色球1从指掌左侧迅速飞离，粉褐色球2也向另一侧脱离，结束前没有接回；两项跟踪并行，此处以球1明显脱手定位失败步骤。log step 63显示agent生成并保存的就是82步失败轨迹；评分器报告task_failure=1、sim_steps=82、deterministic=1。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 3.12,
+    "log_step": null,
+    "evidence_zh": "观察：视频0.80–3.04 s，MyoHand弯曲手指将两球留在指掌间并使其移动；视频3.12–3.28 s，左侧视角可见黄色球1迅速远离指掌，粉褐色球2也与手指拉开间隙，右侧近景可见球2下坠，均未重新承接。已有指掌接触后丢失承托，且不是到达目标后的计划松手。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 3.12,
+    "log_step": 63,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 44及58的agent脚本以周期肌肉激励叠加两球位置误差反馈来驱动双球，目标与任务的跟踪要求一致；log step 61选择的参数仍只有64步共同连续合格，log step 63用该参数生成并保存的轨迹明确在82步失败。结合视频3.12–3.28 s手指仍弯曲而两球突然脱离支承，说明该肌肉控制未能持续维持球的稳定手内运动；现有证据只能归到控制，不能确定是接触布局、反馈增益还是其他执行细节。"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志；视频中的目标标记小且受球和手指遮挡，不能精确恢复两球三维误差、各自短暂合格窗口或掉球瞬间的世界高度。不过评分器的82步回合长度足以排除任何200步节点达成。评分器未报告target1_dist、target2_dist、solved_run，三项native_metrics均留null；日志中的64是agent脚本计算的共同最长窗口，不当作评分器原生指标。缺少接触力及控制对照试验，根因不细分。"
+ },
+ "summary_zh": "MyoHand已同时对两球进行手内调整，但交付回合仅82控制步，两项要求连续200步合格的跟踪节点均未达成，终态0/2。视频约3.12 s，球1迅速飞离指掌，球2也脱手，断点为球1持续跟踪的操作中，现象是没保持住，根因归为控制。缺少逐步状态和接触力，无法精确复核短暂跟踪误差或细分控制失稳原因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与视频、日志一致：回合仅82控制步，两个200步节点均不可能达成，断点与现象（没保持住）有帧证据，根因控制已保守表述。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:58:27Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--y8ssqwd/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-privileged--y8ssqwd/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8a2526dfb7b1f22bb1288ad741b225e54531d8f149ae5c39cfe91e5c6d5c2f10",
+   "video_sha256": "b5bcad7b5c6aa879cb36d9af3806a2593b223137da7b5b66994ff5a27eade25a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-standard--4uq3ivr.json

@@ -0,0 +1,261 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-standard--4uq3ivr",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 212,
+   "sim_steps": 212,
+   "physics_steps": 2120,
+   "steps": 229,
+   "calls": 4,
+   "episodes": 2,
+   "resets": 1,
+   "reset_budget": 10,
+   "sim_steps_total": 229,
+   "elapsed_s": 2708.0,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.4,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.371,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-standard--4uq3ivr/replay.mp4",
+  "duration_s": 8.52,
+  "fps": 25.0,
+  "frames": 213,
+  "sha256": "59808fa95fef8b592e0a93655c6bd6c69fefc8db9d168522b093293e7fbdd1db"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.4,
+      "log_step": 37,
+      "text_zh": "观察：视频 0.96–3.36 s 的近景中，MyoHand转动腕掌并屈伸手指，两球仍在指掌间，但没有持续绕掌交换位置；log step 37 中 r14.log 的规划读数为控制步40球1误差0.0148 m、控制步60球1误差0.0248 m且run=0。这些是会话中的规划读数，不作为交付回合的实时状态；交付序列的球1失跟另由 log step 41 对 a14.npy 的诊断支持。"
+     },
+     {
+      "t": 5.6,
+      "log_step": 41,
+      "text_zh": "推断：视频 2.4–7.6 s 的球与手指相对运动，结合 log step 41 对最终交付 a14.npy 的诊断（索引39–69中球1绕掌角约82–91度，而目标角继续从115度到169度），支持球1未跟上移动目标。成功参考视频 2.4–6.8 s 则可见双球连续换位；该参考仅用于识别绕转表现。"
+     },
+     {
+      "t": 8.48,
+      "log_step": 46,
+      "text_zh": "推断：log step 46 明确将 a14.npy 重置后提交为第2回合，服务终止于212控制步；本视频含初态及212次推进，最后下落事件在视频 7.84–8.48 s，与该回合一致。一控制步对应一帧，即视频0.04 s，200步窗口跨7.96个视频秒；上述视频中段的跟踪偏离使整个短回合无法积满球1的200步窗口，故从未达到节点。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.4–7.6 s，MyoHand单手的屈曲手指配合掌面持续接触掌内双球，通过转腕及手指开合尝试滚动、换位球1；没有换手或使用工具。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 5.6,
+      "log_step": 40,
+      "text_zh": "观察：视频 4.48–5.92 s 的近景中，一球长时间停在远端手指附近，另一球留在掌侧，没有像成功参考视频 4.4–6.4 s 那样继续绕掌换位。log step 40 是最终交付 a14.npy 的本地诊断：索引134–139中球2的归一化椭圆半径为2.46、2.36，而目标轨道半径为1；这些数值不是服务的实时球位姿。"
+     },
+     {
+      "t": 7.92,
+      "log_step": 46,
+      "text_zh": "观察：视频 7.76 s 时浅褐色球仍贴近伸出的手指，视频 7.84–8.08 s 它与手指间的间距迅速增大，视频 8.16–8.48 s 继续远离并下落，黄色球仍在掌指间；log step 46 的第2回合服务返回 failure: a ball dropped or left its target，control_steps=212。"
+     },
+     {
+      "t": 8.48,
+      "log_step": 46,
+      "text_zh": "推断：视频中段没有持续绕转，末段又失去一球，结合 log step 40 中交付文件的球2轨道偏离诊断，支持球2未形成连续200步合格窗口。log step 46 确认同一 a14.npy 的本地掉球和服务失败均在控制步212，已用视频 7.84–8.48 s 的掉球核对回合；212控制步仅跨8.48个视频秒，不能把短暂在掌内或变色当作持续200步达标。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频 0.4–7.76 s，MyoHand单手手指配合掌面接触并滚动、承托双球，尝试在手内换位球2；视频约7.84 s后远端浅褐色球脱离指掌支承。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 2.0,
+   "log_step": 41,
+   "where": "操作中",
+   "evidence_zh": "推断：视频约1.6–2.4 s起，机器人继续转腕和屈指，但掌内两球没有随目标持续绕转，之后始终未修复为完整跟踪窗口；log step 41 对交付序列 a14.npy 的本地诊断显示，索引4–69中球1绕掌角约76–95度，而目标角从52度推进到169度，其中索引39–69球1仅约82–91度、目标角从115度推进到169度，支持球1未跟上移动目标。因此将最早未修复步骤定位为球1持续跟踪，视频约2.0 s为粗略锚点。log step 46 确认交付该文件，视频 7.84–8.48 s 后续掉球与服务212步失败一致；掉球不是另一个断点。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 2.4,
+    "log_step": 41,
+    "evidence_zh": "推断：视频 1.6–3.36 s 中MyoHand虽持续运动，但两球没有继续完成绕掌换位；视频 4.48–5.92 s 仍主要分居掌侧与远端手指。结合 log step 41 对已交付 a14.npy 的本地诊断，球1在索引39–69保持约82–91度的绕掌角，而其移动目标从115度转到169度，说明需要跟踪的运动目标已被落下；这是没跟上移动目标。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 16、34–35 的脚本明确以两球各自的移动目标误差搜索39个肌肉命令，并非瞄准了别的任务目标；log step 37 是 r14.log 的会话规划读数；log step 40–41 对最终交付 a14.npy 的本地诊断显示球1绕转停滞、球2随后偏出轨道。log step 46 把该序列提交给服务后同样在212控制步失败，视频 1.6–7.6 s 可见腕掌大幅转动却未实现持续双球换位，视频 7.84–8.48 s 又掉球。据此归到肌肉控制未实现所需滚动跟踪；现有材料不足以细分接触设计、轨迹求解或反馈控制。"
+   }
+  ],
+  "gap_zh": "没有逐控制步的实时球心、目标三维位置、接触对和各球连续合格计数；评分器也未报告 target1_dist、target2_dist 或 solved_run，因此原生指标均填null。小目标标记受球和手指遮挡，不能从变色精确确认0.015 m阈值；节点未达与球1先失跟的判断结合视频中的绕转缺失及最终交付文件的本地诊断作推断，不能把本地诊断等同于实时状态日志。球的固定编号与画面颜色对应也主要依赖该诊断的运动一致性，精确越界帧及控制子类无法独立核实。"
+ },
+ "summary_zh": "第2回合中，MyoHand通过转腕及手指开合在掌内操作双球，但两颗球均未完成连续200控制步跟踪，最终节点为0/2。球1先未跟上移动目标，之后视频约7.84–8.48 s一球脱离指掌支承并下落，服务在212控制步判失败，根因归为控制。没有实时三维球位姿或连续合格计数，精确距离、越界时刻和控制子类依赖交付文件的本地诊断辅助推断。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 7,
+    "summary_zh": "节点0/2、断点球1、动态时机/控制整体有据可循（回合在212控制步掉球，视频7.84–8.48 s一致）。仅有一处证据归属不准：把r14.log的读数说成最终交付a14.npy的诊断。",
+    "n_issues": 1,
+    "n_major": 0
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点0/2、断点球1、动态时机/控制与视频及log一致；第1轮的证据归属问题已修正，未发现新的内容性错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:11:21Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--4uq3ivr/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--4uq3ivr/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--4uq3ivr/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--4uq3ivr/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "62f3e465a092bcb4a53f550a98009c0642032024498da1633acc3b2ca544183f",
+   "video_sha256": "59808fa95fef8b592e0a93655c6bd6c69fefc8db9d168522b093293e7fbdd1db",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-standard--fvtkunk.json

@@ -0,0 +1,240 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-standard--fvtkunk",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 319,
+   "sim_steps": 319,
+   "physics_steps": 3190,
+   "steps": 2104,
+   "calls": 2120,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 2104,
+   "elapsed_s": 3617.0,
+   "video_rendered": 1,
+   "task_failure": 0,
+   "timeout": 1,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.52,
+   "peak_processes": 9,
+   "peak_disk_gb": 1.251,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-standard--fvtkunk/replay.mp4",
+  "duration_s": 12.8,
+  "fps": 25.0,
+  "frames": 320,
+  "sha256": "e2383767340a7e07808528a878e59aa533e3954ff7ab00a904c614c502717213"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 91,
+      "text_zh": "观察：log step 91启动最后一次reset后的run_wave.py；log step 92的服务元数据确认episode=11、resets_left=0。视频0.0 s两球在掌中，随后MyoHand屈指、摆腕；本节点的200步持续条件在初始帧尚未累计完成。"
+     },
+     {
+      "t": 4.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0–12.76 s的双视角及每约0.2 s近景显示，两球一直相邻留在指掌间，主要随手腕往复摆动，彼此相对方位只小幅改变；视频2.0、4.0、6.4、8.0、10.4 s未见球绕到另一侧的持续换位。参考成功视频0.0–8.28 s则可见两球沿掌内轨道交换方位。"
+     },
+     {
+      "t": 12.76,
+      "log_step": 93,
+      "text_zh": "推断：结合视频全程缺少完整绕转、视频11.8–12.76 s仍停留在相近掌内区域，以及任务中每个固定目标绕掌运行的要求，球1未形成连续200步的固定配对跟踪窗口，故intervals为空、final=false；不是仅由总成功分数为0推断。log step 93的agent自述“两球在振荡而未完成绕行”与画面一致，但未用其视觉估计充当精确距离。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.2–12.76 s，MyoHand单手屈曲的手指与掌面持续包托球1所在的双球组，随腕部摆动在手内小幅拨动、调整；球未脱离手掌，未完成持续绕转。不据遮挡画面细分每根手指的接触。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 91,
+      "text_zh": "观察：log step 91启动最后一次reset后的run_wave.py；log step 92的服务元数据确认episode=11、resets_left=0。视频0.0 s两球在掌中，随后MyoHand屈指、摆腕；本节点的200步持续条件在初始帧尚未累计完成。"
+     },
+     {
+      "t": 4.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0–12.76 s的双视角及每约0.2 s近景显示，两球一直相邻留在指掌间，主要随手腕往复摆动，彼此相对方位只小幅改变；视频2.0、4.0、6.4、8.0、10.4 s未见球绕到另一侧的持续换位。参考成功视频0.0–8.28 s则可见两球沿掌内轨道交换方位。"
+     },
+     {
+      "t": 12.76,
+      "log_step": 93,
+      "text_zh": "推断：结合视频全程缺少完整绕转、视频11.8–12.76 s仍停留在相近掌内区域，以及任务中每个固定目标绕掌运行的要求，球2未形成连续200步的固定配对跟踪窗口，故intervals为空、final=false；不是仅由总成功分数为0推断。log step 93的agent自述“两球在振荡而未完成绕行”与画面一致，但未用其视觉估计充当精确距离。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.2–12.76 s，MyoHand单手屈曲的手指与掌面持续包托球2所在的双球组，随腕部摆动在手内小幅拨动、调整；球未脱离手掌，未完成持续绕转。不据遮挡画面细分每根手指的接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 2.0,
+   "log_step": 91,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约2.0 s手已包住两球并摆腕，但双球仍保持相邻的近似方位；随后至视频12.76 s仅反复小幅摆动，未恢复为参考成功视频中的掌内绕转。两节点并行同时受阻，以球1持续跟踪作为断点；没有证据显示球2先于球1发生独立失败。log step 91为该评分回合的启动记录，前一回合在log step 90的掉球不属于本次视频。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 2.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频约2.0–12.76 s，MyoHand手腕和手指持续活动，两球却只在掌中往复小幅移动，未随需要绕掌运动的目标持续换位；视频11.8–12.76 s仍如此。推断：这是对移动目标未跟上的表现；未见本回合掉球，不将调试回合掉球标为没保持住。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 91,
+    "source_zh": "日志",
+    "evidence_zh": "观察：log step 86的run_wave.py用control_steps生成固定正弦相位和手腕、手指关节目标，再以本体感知计算肌肉命令；BallTracker的球位置只在每20步的分支内检测、打印及保存，不进入下一步动作计算。log step 89、91仅修改固定摆幅和屈指姿态，并在log step 91启动最终评分回合；log step 93的agent自述仍为振荡。推断：该关节周期动作没有根据实际球与移动目标的偏差修正掌内运动，视频全程只有振荡而未形成绕转，说明所实现控制未能产生所需双球运动。现有证据不足以进一步区分接触几何、轨迹或底层肌肉跟踪问题，因此停在控制。"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志，也没有评分器报告的target1_dist、target2_dist或solved_run；小目标标记与部分指球接触被遮挡，不能从二维画面或变色复原三维0.015米误差、短暂合格时段及精确首次越界时刻。两球均未形成可见的绕转窗口，故可判两个持续节点未完成；断点视频约2.0 s仅为振荡停滞的导航锚点，无法判两节点失败的先后，也不能将画面颜色可靠地绑定object编号。",
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  }
+ },
+ "summary_zh": "评分采用log step 91最后重置后的第11回合：MyoHand持续包托并小幅调整双球，但视频0.0–12.76 s均未形成跟随移动目标的完整绕转，两个持续跟踪节点均未达到，最终0/2。断点记球1持续跟踪的操作中，球2并行受阻；固定周期关节控制只产生振荡，球位置检测未用于动作纠偏，评分器最终记录timeout=1、live_success=0。缺少逐步三维距离与合格计数，无法量化瞬时误差或精确定位第一次跟踪越界。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "未发现需要修改的错误：两节点final=false与视频一致（两球全程留在掌中小幅摆动），断点、现象与根因证据充分且标注了推断，缺口说明合理。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:59:17Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--fvtkunk/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--fvtkunk/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "1f863a7f5a84873642684abaf23247e749fd17c71e0be23c2ecb19ab48fc044e",
+   "video_sha256": "e2383767340a7e07808528a878e59aa533e3954ff7ab00a904c614c502717213",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-standard--ghmmw85.json

@@ -0,0 +1,214 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-standard--ghmmw85",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 9,
+   "sim_steps": 9,
+   "physics_steps": 90,
+   "steps": 145,
+   "calls": 134,
+   "episodes": 7,
+   "resets": 6,
+   "reset_budget": 10,
+   "sim_steps_total": 145,
+   "elapsed_s": 3489.8,
+   "video_rendered": 1,
+   "task_failure": 0,
+   "timeout": 1,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.995,
+   "peak_processes": 12,
+   "peak_disk_gb": 0.383,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-standard--ghmmw85/replay.mp4",
+  "duration_s": 0.4,
+  "fps": 25.0,
+  "frames": 10,
+  "sha256": "2b6cd4c4ed618920c0e572a7a98b3a19bf1246ff09847accc4571c9d261b2924"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–0.36 s的全部10帧中，MyoHand由较伸展姿态逐渐屈指成杯状，两颗球始终位于指掌附近，球1没有出现持续绕行一整圈的过程。"
+     },
+     {
+      "t": 0.36,
+      "log_step": null,
+      "text_zh": "推断：评分器reward仅记录最后回合n_actions=9、sim_steps=9，视频恰为初始画面加9步的10帧；即使球1每步都满足距离阈值，也不可能累计节点所需的200个连续控制步，故全程未达到。视频时间按25 fps读取，未将agent工作时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.00–0.36 s，MyoHand单只手的手指在球1周围屈曲，与掌面配合承托、挪动该球；两球仍在指掌间，没有可见的主动释放。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–0.36 s的全部10帧中，MyoHand由较伸展姿态逐渐屈指成杯状，两颗球始终位于指掌附近，球2没有出现持续绕行一整圈的过程。"
+     },
+     {
+      "t": 0.36,
+      "log_step": null,
+      "text_zh": "推断：评分器reward仅记录最后回合n_actions=9、sim_steps=9，视频恰为初始画面加9步的10帧；即使球2每步都满足距离阈值，也不可能累计节点所需的200个连续控制步，故全程未达到。视频时间按25 fps读取，未将agent工作时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.00–0.36 s，MyoHand单只手的手指在球2周围屈曲，与掌面配合承托、挪动该球；两球仍在指掌间，没有可见的主动释放。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": null,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频0.00–0.36 s，MyoHand已在两球周围屈指并进行掌内操作；评分器记录最后回合只有9个控制步、success=0、timeout=1、task_failure=0。推断：球1持续跟踪尚未完成且此后没有修复，断点定位该操作；球2的并行跟踪同样未完成。录像没有呈现可单独定位的不可逆物理失败时刻，因此时间置空。"
+  },
+  "phenomena": [],
+  "root_causes": [],
+  "gap_zh": "观察：inputs/含本次录像、完整会话日志和成功参考录像，但��有states.jsonl；评分器没有报告target1_dist、target2_dist或solved_run，原生指标均置null。本次视频0.00–0.36 s仅展示9个控制步，末帧仍见两球处于指掌附近；低分辨率及遮挡不能确定精确三维跟踪误差，也未见明确掉球或其他可归类的失败现象，故现象层留空。log step 64、66含先前回合的服务终止信息，log step 67为agent对会话的总结，均不能证明最后9步回合发生同样故障；缺少最后回合的逐步估计与执行误差，无法区分感知、接触设计和反馈控制等根因，故根因层留空。",
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  }
+ },
+ "summary_zh": "评分录像中MyoHand已对掌中两球进行屈指与手内调整，但最后回合只有9个控制步，两个要求连续200步跟踪的节点均未达到，终态为0/2。断点定位球1持续跟踪的操作中，球2并行操作也未完成；视频0.36 s末帧仍见两球处于指掌附近，不能把先前测试回合的掉球套用于本次评分轨迹。缺少最后回合的三维球位、目标误差和连续合格计数，具体失败现象与根因留空。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与证据一致：评分回合只有9个控制步（n_actions=9，视频10帧0.4 s），两个要求连续200步的节点不可能达到，final=0；断点、空现象/根因层均带gap_zh，且未把先前回合的掉球套用到评分回合。未发现需要改动的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:12:20Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--ghmmw85/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--ghmmw85/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8aa3281885d290ea255e22159d52cb8c75174e63cbe0dc34596b991773ff874d",
+   "video_sha256": "2b6cd4c4ed618920c0e572a7a98b3a19bf1246ff09847accc4571c9d261b2924",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-standard--jisq83e.json

@@ -0,0 +1,257 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-standard--jisq83e",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 271,
+   "sim_steps": 271,
+   "physics_steps": 2710,
+   "steps": 341,
+   "calls": 15,
+   "episodes": 3,
+   "resets": 2,
+   "reset_budget": 10,
+   "sim_steps_total": 341,
+   "elapsed_s": 1937.8,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 2.526,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.663,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-standard--jisq83e/replay.mp4",
+  "duration_s": 10.88,
+  "fps": 25.0,
+  "frames": 272,
+  "sha256": "2ef4693c093a60ac620f3439a80e6c096a8d9eeec04c6ae1a27b93060b921ddf"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 6.4,
+      "log_step": null,
+      "text_zh": "观察：视频0.0–10.84 s的双视角中，单手屈指托住黄色球1和粉色球2，但两球长时间保持同侧相邻位置；视频6.0–7.2 s右侧近景可见黄色目标标记经过粉色球附近，黄色球1仍在其右侧，未随对应标记换位。成功参考视频3.0–5.0 s则可见两球随目标交换掌中位置。"
+     },
+     {
+      "t": 10.84,
+      "log_step": 318,
+      "text_zh": "观察：视频10.52–10.84 s，黄色球1脱离指掌，左侧全景可见球与手之间的间距迅速增大；log step 318中，机器人服务在最后第3回合control_steps=271报告done=True、reason=failure: a ball dropped or left its target。"
+     },
+     {
+      "t": 6.4,
+      "log_step": 318,
+      "text_zh": "推断：log step 318的271个控制步对应录像初态加271帧（272帧、25 fps），末步271/25=视频10.84 s与可见脱手终止核对一致；200个连续步跨199/25=7.96个视频秒。视频6.4 s球1明显偏离配对目标，任何可能的完整200步窗口都包含此段偏离，因此没有完成保持节点；这里未用变色或零分代替位置判断。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.4–10.4 s，MyoHand单手的屈曲手指配合掌面持续接触黄色球1，随腕姿变化使其在指掌间滚动；视频10.52–10.84 s球1脱手。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 6.4,
+      "log_step": null,
+      "text_zh": "观察：视频0.4–7.2 s，粉色球2持续位于黄色球1左侧的指掌处，未随移动目标绕过另一颗球；视频6.4 s右侧近景中，两球及小目标标记的相对位置已错开，黄色小标记位于粉色球附近，另一配对位置在黄色球一侧。成功参考视频3.0–5.0 s呈现两球交换位置，本次没有该绕转过程。"
+     },
+     {
+      "t": 10.84,
+      "log_step": 318,
+      "text_zh": "推断：视频6.0–7.2 s的配对目标与球2错位排除了连续200步跟踪；该偏离段位于本录像所有可能的7.96视频秒保持窗口内（log step 318的271控制步与272帧录像末端脱手相符，每控制步一帧、25 fps）。视频10.84 s粉色球2仍在手边，但留在掌中不足以满足移动目标跟踪；未把本地测试数值当作本回合逐步测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.4–10.84 s，MyoHand单手的弯曲手指和掌面包围、承托粉色球2，随手腕转动在手内小幅滚动换位；没有双手协作或工具接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 3.2,
+   "log_step": 318,
+   "where": "操作中",
+   "evidence_zh": "推断：视频2.4–4.0 s机器人已在屈指、转腕操作双球，但两球相对指掌的位置变化很小，未继续沿移动目标绕行；视频6.0–7.2 s可见配对目标与球明显错位，此后也未形成完整跟踪窗口。两个跟踪步骤并行，取黄色球1的持续跟踪作为断点，视频3.2 s仅为较早停滞段的导航锚点，并非精确越过0.015 m的时刻；视频10.52–10.84 s球1最终脱手，log step 318的第3回合在271控制步终止。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 3.2,
+    "log_step": null,
+    "evidence_zh": "观察：视频2.4–7.2 s机器人虽持续屈指和转腕，两球仍大体维持原来的左右关系，移动目标已绕向另一侧；视频6.4 s黄色小目标已到粉色球附近，黄色球1没有跟上。目标在移动，双球未持续同步绕行。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 282,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 282交付脚本按固定200步周期的基频及二次谐波预生成39维肌肉动作，执行时每50步提交，仅检查服务done/reason，没有按球与目标位置修正动作；log step 318确认第3回合执行该脚本及其谐波参数。该肌肉控制轨迹未使双球持续绕转，并在视频10.52–10.84 s失去对球1的保持；log step 315的独立本地测试也只得到最长48步共同合格、270步存活，作为控制能力的辅助证据，不能作为本次视频的精确计数。现有材料不足以进一步区分接触布局、轨迹或具体肌肉参数问题。"
+   }
+  ],
+  "gap_zh": "观察：inputs/中有本次录像、运行日志和成功参考录像，但没有逐控制步状态日志；评分器reward未报告target1_dist、target2_dist或solved_run，故三个原生指标均为null。推断：视频足以排除完整跟踪窗口并确认黄色球1脱手，但遮挡和二维投影不能给出精确三维误差、最早越界帧及两个并行节点谁先越界；断点时间是可见停滞段的粗略锚点。log step 315等本地仿真结果仅用于根因辅助，不与评分回合合并。"
+ },
+ "summary_zh": "观察：最后第3回合中，MyoHand单手接触并调整双球，但两球没有持续随各自移动目标绕行，两个200步跟踪节点均未完成，终态0/2。推断：断点取球1持续跟踪，固定谐波肌肉控制未维持同步绕转；视频10.52–10.84 s黄色球1脱手，log step 318的机器人服务在271控制步报告失败。观察：缺少本回合逐步三维距离与合格计数，无法确定精确越界时刻及更细控制根因。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点值（两个跟踪节点未达成、final=0）与断点选择可接受；但把评分器终止事件（球脱手）单列为第二个现象不合规则，应删除，改为在证据中提及。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点值（两个跟踪节点未达成、final=0）、断点（球1持续跟踪，操作中）、现象与根因均可接受；第1轮意见已处理，时钟换算（271控制步=272帧）与末帧脱手一致。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:10:51Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--jisq83e/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--jisq83e/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--jisq83e/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--jisq83e/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "3c2aaf16c9c305bb88b54f25af00f513817afb35eb9a3f13a0221658fcc4eea0",
+   "video_sha256": "2ef4693c093a60ac620f3439a80e6c096a8d9eeec04c6ae1a27b93060b921ddf",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-standard--pntstup.json

@@ -0,0 +1,229 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-standard--pntstup",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 60,
+   "sim_steps": 60,
+   "physics_steps": 600,
+   "steps": 96,
+   "calls": 5,
+   "episodes": 2,
+   "resets": 1,
+   "reset_budget": 10,
+   "sim_steps_total": 96,
+   "elapsed_s": 1450.5,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 4.283,
+   "peak_processes": 20,
+   "peak_disk_gb": 1.322,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-standard--pntstup/replay.mp4",
+  "duration_s": 2.44,
+  "fps": 25.0,
+  "frames": 61,
+  "sha256": "ff223d8c921ea2ef5260ebe0f3acbc1d4cc2da2a3d6a9ff91b61dcdd6becd3e7"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 106,
+      "text_zh": "推断：log step 106 的服务返回确认评分对象为重置后的 episode 2，共执行60个控制步即失败；60步不足节点要求的连续200步，因此该节点全程未达到，不能把掌中短时靠近目标算作完成。视频0.00–2.40 s为该回合起止帧，已核对末段粉色球离手；61帧对应初始帧加60步，每步占视频0.04 s（60/25=视频2.40 s），不是每步0.025 s的仿真时钟。"
+     },
+     {
+      "t": 2.4,
+      "log_step": null,
+      "text_zh": "观察：视频0.00、0.40、0.80、1.20、1.60、2.00及2.40 s，黄色球随MyoHand指掌姿态改变而在手中移动，末帧仍在指掌围成的区域；这只能说明有操球，不能证明连续200步的三维目标跟踪。log step 38 的模型片段将ball1对应MatBallY。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.00–2.40 s，MyoHand单手的手指弯曲并配合掌面改变黄色球在手中的位置；球一直留在指掌区域，未见主动释放。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 106,
+      "text_zh": "推断：log step 106 的服务返回确认评分对象为重置后的 episode 2，共执行60个控制步即失败；60步不足节点要求的连续200步，因此该节点全程未达到，不能把掌中短时靠近目标算作完成。视频0.00–2.40 s为该回合起止帧，已核对末段粉色球离手；61帧对应初始帧加60步，每步占视频0.04 s（60/25=视频2.40 s），不是每步0.025 s的仿真时钟。"
+     },
+     {
+      "t": 2.16,
+      "log_step": 38,
+      "text_zh": "观察：log step 38 的模型片段将ball2对应MatBallP；视频0.00–2.08 s粉色球由指掌承托，视频2.12–2.24 s逐渐脱离指掌区域，视频2.40 s左侧全景中已远离手部，未被接回。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "手内调整",
+      "note_zh": "观察：视频0.00–2.08 s，MyoHand单手弯曲手指并配合掌面移动粉色球；视频约2.12–2.24 s球从指掌区域逸出，此后失去承托。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball2_target_tracked",
+   "t": 2.16,
+   "log_step": 106,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约2.12–2.24 s，MyoHand正在掌中操动双球时，粉色球离开指掌区域；视频2.40 s左侧全景显示它已远离手，黄色球仍在掌中。log step 106 的服务返回episode 2在60个控制步结束，reason为“failure: a ball dropped or left its target”。推断：两球跟踪为并行操作，此处以发生不可恢复失球的球2持续跟踪节点作为断点。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 2.16,
+    "log_step": 106,
+    "evidence_zh": "观察：视频1.76–2.08 s粉色球仍靠在MyoHand指掌间，视频2.12–2.24 s与指掌的距离迅速增大，视频2.40 s左侧全景可见其已脱离手部；机器人没有接回。log step 106服务同时报告该评分回合以掉球或离开目标失败结束。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "观察：log step 104中agent脚本生成并保存最终肌肉命令序列，本地rollout输出surv=60、nsucc=0、maxrun=0；这些是会话测试结果。log step 106中部署脚本重置后向服务开环播放同一序列，服务在60个控制步报失败；视频约2.12–2.40 s实际显示粉色球逸出指掌。推断：所生成的肌肉控制未能维持操球所需的承托与运动稳定性；本地测试与实时评分均暴露失球，支持控制层归因。没有足够证据区分接触设计、反馈律或参数问题，保留在控制总类。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "无逐控制步状态日志，服务未报告两球各自的目标距离、世界高度、共同连续合格计数或接触力；native_metrics三项均无评分器数值，故填null。二维画面不能精确定位0.015米越界或1.25米高度阈值的时刻，但评分回合只有60步，足以确定两个200步节点均未达到。根因仅能确定至控制总类，不能验证具体控制子类。"
+ },
+ "summary_zh": "评分对象为重置后的第2回合，MyoHand确实在掌中操动双球，但仅执行60个控制步，两个要求连续200步的跟踪节点均未达到，最终0/2。视频约2.12–2.24 s粉色球脱离指掌且未被接回，断点为球2持续跟踪，现象为没保持住；最终序列的本地测试和服务执行均在60步失效，支持控制层归因。缺少逐步距离、高度及接触数据，无法细分控制根因；成功参考视频仅用于比较视角与操球外观。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "评分回合为重置后的第2回合（log step 108状态：episode 2，control_steps 60，failure）。两节点未达到、断点球2、现象没保持住（粉球约2.12–2.24 s脱手）与视频一致；根因停在控制总类有日志支持（本地孪生与服务均在60步失败），未发现需改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:59:14Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--pntstup/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--pntstup/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f66116c4cef7f06858652953a067d5cbb9c73d8d0154026238d8f961b18018de",
+   "video_sha256": "ff223d8c921ea2ef5260ebe0f3acbc1d4cc2da2a3d6a9ff91b61dcdd6becd3e7",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-standard--q9ytca9.json

@@ -0,0 +1,262 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-standard--q9ytca9",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 70,
+   "sim_steps": 70,
+   "physics_steps": 700,
+   "steps": 778,
+   "calls": 98,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 778,
+   "elapsed_s": 1752.7,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.477,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.857,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-standard--q9ytca9/replay.mp4",
+  "duration_s": 2.84,
+  "fps": 25.0,
+  "frames": 71,
+  "sha256": "3e5d90c182fa94210ffe5e7c1e73f31794db3c4f19a9400a4993a7fd125af624"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 147,
+      "text_zh": "观察：log step 147 的服务回复显示最后一次重置进入 episode 11、resets_left=0；log step 149 执行控制脚本，评分视频从双球在指掌间的初态开始。此前回合的调试不计入本节点。"
+     },
+     {
+      "t": 2.8,
+      "log_step": 149,
+      "text_zh": "观察：视频 0.00–2.56 s，MyoHand 单手指掌小幅运动，两球仍在指掌附近；视频 2.60–2.80 s 粉色球逐渐离开支承，末帧黄色球仍在手边。评分器 reward.sim_steps=70、n_actions=70，live_matches_replay=1；视频共71帧。"
+     },
+     {
+      "t": 2.8,
+      "log_step": null,
+      "text_zh": "推断：本评分回合仅执行70个控制步，即使本球每一步都满足距离阈值，也不足节点要求的连续200步，所以整个视频内均未达到持续跟踪条件，终值为false；并非仅由success=0推定。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "托/舀",
+      "note_zh": "观察：视频 0.00–2.40 s，MyoHand 单只手的指掌承托本节点对应球，随腕部及手指的小幅运动使球在手边滚动；没有清楚的合手夹紧动作。两球均受到承托，颜色与object编号未在输入中明确对应。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 147,
+      "text_zh": "观察：log step 147 的服务回复显示最后一次重置进入 episode 11、resets_left=0；log step 149 执行控制脚本，评分视频从双球在指掌间的初态开始。此前回合的调试不计入本节点。"
+     },
+     {
+      "t": 2.8,
+      "log_step": 149,
+      "text_zh": "观察：视频 0.00–2.56 s，MyoHand 单手指掌小幅运动，两球仍在指掌附近；视频 2.60–2.80 s 粉色球逐渐离开支承，末帧黄色球仍在手边。评分器 reward.sim_steps=70、n_actions=70，live_matches_replay=1；视频共71帧。"
+     },
+     {
+      "t": 2.8,
+      "log_step": null,
+      "text_zh": "推断：本评分回合仅执行70个控制步，即使本球每一步都满足距离阈值，也不足节点要求的连续200步，所以整个视频内均未达到持续跟踪条件，终值为false；并非仅由success=0推定。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "托/舀",
+      "note_zh": "观察：视频 0.00–2.40 s，MyoHand 单只手的指掌承托本节点对应球，随腕部及手指的小幅运动使球在手边滚动；没有清楚的合手夹紧动作。两球均受到承托，颜色与object编号未在输入中明确对应。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 2.6,
+   "log_step": 149,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 2.40–2.56 s 粉色球仍贴近指掌边缘，视频 2.60–2.80 s 它与指掌的间隙迅速扩大并离开右侧画面，机器人未接回；黄色球仍在手边。评分器 reward.sim_steps=70、n_actions=70，少于两个节点各自要求的连续200步。推断：两个并行跟踪节点在同一段双球操作中均未完成，未见可区分的失败先后，按DAG所列第一个节点ball1_target_tracked记录断点，球2节点同样未完成；此记录不表示粉色球就是object1，颜色与object编号仍无明确对应，也无法确认两球各自首次误差越界的时刻。"
+  },
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 2.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.00–2.40 s 粉色球受MyoHand指掌承托；视频 2.60–2.80 s 粉色球脱离指掌并迅速远离，机器人没有恢复接触。此为原有承托丢失，未见主动将球放到目标上的释放。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 2.6,
+    "log_step": 146,
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 146 的控制脚本读取相机与本体感知，以每次5个控制步的肌肉脉冲驱动手部；log step 149 记录实际执行14次submit。视频 0.00–2.56 s 指掌虽有运动，却没有维持住双球的绕转与支承，视频 2.60–2.80 s 粉色球滑脱，支持肌肉控制未能维持操作的判断；没有充分证据细分到接触设计、反馈控制或具体参数。"
+   }
+  ],
+  "gap_zh": "观察：inputs/提供本次和成功参照视频、完整会话日志，但没有状态日志；评分器未报告target1_dist、target2_dist或solved_run。log step 146 的脚本中间有截断，log step 149–150未收录最终控制器的打印输出；log step 151仅有agent自述“粉色球低于1.25米”，不能当作精确高度测量。推断：两节点可由70步不足200步确定未完成，但无法复核每步三维误差、精确掉落高度及控制子类；颜色与object编号无明确对应，不能把粉色球脱离支承归到某一个物体编号。断点按两个并行未完成节点中的首个节点记录，不表示已识别掉落球编号或失败先后。"
+ },
+ "summary_zh": "本次评分对象是最后的episode 11：机器人单手承托并尝试操纵双球，但仅运行70个控制步，两个要求连续200步的持续跟踪节点均未完成，终态0/2。视频约2.60–2.80 s粉色球脱离指掌且未接回，现象为没保持住，证据支持一般控制问题；两个节点并行未完成，断点按DAG首个节点ball1_target_tracked记录。没有逐步误差与高度日志，且颜色与object编号未明确对应，不能确定掉落球编号、两球首次越界先后或控制子类。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点终值false、时钟（70控制步≈71帧）、粉球约2.6 s脱离指掌均与视频一致；唯一问题是断点ref为unknown，而两个并行节点同时失败，应取其一。",
+    "n_issues": 1,
+    "n_major": 0
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两节点终值false（70控制步<200步）、粉球约2.6 s脱离指掌、断点ref取首个并行节点、根因控制均与视频和评分字段一致，未发现需改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:12:25Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--q9ytca9/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--q9ytca9/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--q9ytca9/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--q9ytca9/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "5fc5fa4ab6bf32ab075fe28a7712aeb6bd7a017e65a01565a03b70bb11f1faee",
+   "video_sha256": "3e5d90c182fa94210ffe5e7c1e73f31794db3c4f19a9400a4993a7fd125af624",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/baoding/trajectories/myosuite-baoding-i00-standard--qftdcfj.json

@@ -0,0 +1,230 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-baoding-i00-standard--qftdcfj",
+ "taskKey": "myosuite/baoding",
+ "dag_sha256": "d601c830f12d5f8aa686b3f3d6b6e58345f56dd2da1103766529b1f1c3ddafd8",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "physics_steps": 10000,
+   "steps": 1036,
+   "calls": 6,
+   "episodes": 2,
+   "resets": 1,
+   "reset_budget": 10,
+   "sim_steps_total": 1036,
+   "elapsed_s": 2381.3,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 2.721,
+   "peak_processes": 15,
+   "peak_disk_gb": 0.501,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-baoding-i00-standard--qftdcfj/replay.mp4",
+  "duration_s": 40.04,
+  "fps": 25.0,
+  "frames": 1001,
+  "sha256": "34cf003498971fba7fa3eaa90c500baa22ec9beb7d83de78149ad93fc988c906"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "ball1_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 4.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0、0.4、1.2、2.4、4.0、6.4、8.0 s的两侧视角显示，MyoHand先收拢手指并使双球小幅移位，随后两球分别停在指掌间的固定区域；小目标标记继续绕行，球没有随之持续换位。球1对应的持续跟踪条件未见达成，球体变色不作为达标依据。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 69,
+      "text_zh": "推断：视频9.6–38.4 s每约0.8 s的近景以及视频39.04–40.00 s的密集末帧仍显示双球维持原有相对位置，没有任何完整目标周期的绕行；因此球1未连续200个控制步跟随其固定配对目标，最终为false。log step 69的机器人服务确认评分第2回合执行至control_steps=1000，因目标轨迹耗尽结束；这支持未成功，但两球各自未完成的判断来自上述画面，非仅由零分反推。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频0.4–4.0 s及视频39.04–40.00 s，单只MyoHand弯曲的手指配合掌面包夹、留住球1；起始有小幅位移，随后主要保持该握持位置，未见持续滚动换位。"
+     }
+    ]
+   },
+   "ball2_target_tracked": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 4.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0、0.4、1.2、2.4、4.0、6.4、8.0 s的两侧视角显示，MyoHand先收拢手指并使双球小幅移位，随后两球分别停在指掌间的固定区域；小目标标记继续绕行，球没有随之持续换位。球2对应的持续跟踪条件未见达成，球体变色不作为达标依据。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 69,
+      "text_zh": "推断：视频9.6–38.4 s每约0.8 s的近景以及视频39.04–40.00 s的密集末帧仍显示双球维持原有相对位置，没有任何完整目标周期的绕行；因此球2未连续200个控制步跟随其固定配对目标，最终为false。log step 69的机器人服务确认评分第2回合执行至control_steps=1000，因目标轨迹耗尽结束；这支持未成功，但两球各自未完成的判断来自上述画面，非仅由零分反推。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频0.4–4.0 s及视频39.04–40.00 s，单只MyoHand弯曲的手指配合掌面包夹、留住球2；起始有小幅位移，随后主要保持该握持位置，未见持续滚动换位。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target1_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target2_dist",
+   "unit": "m",
+   "threshold": 0.015,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "solved_run",
+   "unit": "control_steps",
+   "threshold": 200,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "ball1_target_tracked",
+   "t": 2.4,
+   "log_step": 69,
+   "where": "操作中",
+   "evidence_zh": "观察：视频0.4–2.4 s手指收拢后双球停在指掌间，视频2.4–9.2 s目标标记移动而球未完成相应绕行，后续直至视频40.00 s也未修复。两个并行跟踪步骤遇到同一问题，以球1持续跟踪作为断点；log step 69表明这正是服务重置后的评分第2回合，结束原因为1000个控制步的目标轨迹耗尽，未发生视频中可见的掉球。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 2.4,
+    "log_step": null,
+    "evidence_zh": "观察：视频0.4–9.2 s的近景中，MyoHand使双球稳定在两个近乎固定的位置，小目标标记从球旁及球面投影处经过；视频9.6–38.4 s仍反复出现相同情况。球没有跟上运动中的目标，不是失去持球接触；视频39.04–40.00 s两球仍留在手中。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "log_step": 69,
+    "evidence_zh": "推断：log step 38、43、58中的脚本把静态握持肌肉命令叠加周期性傅里叶项，按预先生成的命令序列执行；log step 69实际向机器人服务提交该B2策略的1000行命令，首段400个控制步的手部关节轨迹与离线测试比较最大差为0.00000。视频0.4–9.2 s及视频9.6–38.4 s显示策略只产生小幅手内运动，未形成双球轮流绕转，说明所找到的肌肉控制序列不足以实现滚动跟踪。log step 53、68、85的离线诊断及变体测试也显示握持附近的控制未能得到持续绕行，但这些测试不作为评分回合的节点值或掉球现象。",
+    "note_zh": "控制序列未能产生所需绕转有直接证据；没有接触力及逐步球状态，无法进一步区分抓取接触布局、肌肉激活幅相和反馈控制的贡献，保留上位控制标签。"
+   }
+  ],
+  "gap_zh": "本输入没有状态日志；两侧视频可判明未持续绕转，但小目标部分被手指和球遮挡，不能测得各球逐控制步的三维距离、独立连续合格计数或精确越界时刻。log step 69的final dists和best solved streak=24来自agent脚本的离线模拟，不是机器人服务或评分器报告的距离与计数；不据此填写原生指标，也不据此设定节点区间。两球并行出现同一失败，无法确定谁先越过数值阈值，断点时间仅锚定早期可见的未跟随。",
+  "unmet": {
+   "ball1_target_tracked": false,
+   "ball2_target_tracked": false
+  }
+ },
+ "summary_zh": "评分第2回合中，MyoHand握住双球并执行满1000个控制步，但双球主要停在掌内固定位置，没有持续跟随绕行目标，两个节点均未完成。断点取并行的球1持续跟踪，现象为未跟上移动目标，日志与画面支持肌肉控制序列未能形成绕转，服务最终因目标轨迹耗尽结束。没有逐步球状态和评分器距离值，精确误差、连续合格时长及控制子类无法核实。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两球节点均未达成、断点与动态时机现象、控制根因与帧和服务读数一致，无需修改。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:02:36Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--qftdcfj/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-baoding-i00-standard--qftdcfj/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c83e3e951e8d7ba48f87e6a231f4b88e3301c45163e8b515d24fd892c2b0200c",
+   "video_sha256": "34cf003498971fba7fa3eaa90c500baa22ec9beb7d83de78149ad93fc988c906",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-privileged--ahkgq8f.json

@@ -0,0 +1,240 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-privileged--ahkgq8f",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 471,
+   "sim_steps": 471,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 532,
+   "agent_resets": 575,
+   "agent_sim_steps": 338500,
+   "agent_records": 463,
+   "agent_replays": 1,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 6.302,
+   "peak_processes": 39,
+   "peak_disk_gb": 0.395,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-privileged--ahkgq8f/replay.mp4",
+  "duration_s": 18.88,
+  "fps": 25.0,
+  "frames": 472,
+  "sha256": "881dbc4fcab865b390e5376a7bc1f95ac6cb545288763daa048c4983d834c6e7"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [
+     [
+      12.5,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 8.0,
+      "log_step": null,
+      "text_zh": "观察：视频 8.0–12.6 s 的俯视画面中，肌肉右手的拇指与其余手指从相对两侧夹住红边灰色长方体；视频 13.2–15.8 s 物体离开起始柱并随右手抬升，形成实际抓握。"
+     },
+     {
+      "t": 12.5,
+      "log_step": 106,
+      "text_zh": "推断：log step 106 交付并回放的文件有471个动作，本视频为472帧、25 fps，含初态帧，故100个控制步对应约4 s视频；结合视频约 8.0 s 开始的夹持接触及视频 13.2–15.8 s 的持物抬升，右臂累计接触100步约在视频 12.5 s 达标（估计误差约0.5 s，无逐步接触计数）。该节点按DAG是历史事件，视频 16.4 s 滑脱不撤销已达记录。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 8.0–16.2 s，肌肉右手的拇指与其余手指夹住红边长方体，随后将其抬离起始柱；视频 16.4 s 物体滑出，白色假肢手未参与。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 16.4,
+      "log_step": null,
+      "text_zh": "观察：本次视频 0–18.84 s 只有肌肉右手围绕起始柱抓取和抬升；视频 16.2–16.8 s 物体从右手滑出，未由白色假肢手承接，视频 17.0–18.84 s 也未见补接。成功参照视频 16.0–19.0 s 可见白手伸到物体处承接，本次没有出现这一动作或物体随白手移动的结果。"
+     },
+     {
+      "t": 18.84,
+      "log_step": 106,
+      "text_zh": "观察：log step 106 的交付生成器只执行肌肉手抓取抬升脚本，其实际回放返回 success=False、done=True、steps=471；该结果与视频末段掉落一致。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 18.84,
+      "log_step": null,
+      "text_zh": "观察：视频 0–18.84 s 物体始终在起始柱及肌肉右手附近，视频 16.4–18.84 s 脱手后向柱侧下方掉落，没有被左手运到另一根目标柱顶，也没有松手留在目标柱上的阶段。成功参照视频 24.0–31.12 s 可见物体留在另一柱顶、假肢手离开的结果，本次未出现。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_held_by_myo",
+   "t": 16.3,
+   "log_step": 106,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 15.8–16.2 s 肌肉右手仍夹着离柱的长方体，视频 16.4 s 手与物体分开，视频 16.8–18.84 s 物体继续翻转下落，未被重新抓住或由白手接住。log step 106 对交付文件的回放在471个控制步结束，返回 failure: object dropped or the time limit passed。推断：机器人在右手抓取抬升阶段未保持住物体，断点落在该步；累计接触历史达标不等于抓稳或已开始交接。"
+  },
+  "unmet": {
+   "object_held_by_myo": true,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 16.3,
+    "log_step": null,
+    "evidence_zh": "观察：视频 13.2–15.8 s 长方体被肌肉右手夹持并抬离起始柱；视频 16.2–16.4 s 它从弯曲的手指间滑出，随后翻转下落，到视频 18.84 s 仍未被承接。此前已有实际持物接触，失败发生在保持抓握阶段。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "视频",
+    "evidence_zh": "推断：视频 13.2–16.4 s 显示肌肉右手能到达物体、闭合并抬起它，但物体在指间转动并滑出，表明执行中的抓握控制未能保持承载接触。日志中的交付生成器（log step 106）确实使用抓取抬升脚本，agent 也在 log step 107 自述手指和拇指跟踪欠佳、夹持不稳；这一自述与所见滑脱相符，但不足以区分接触布局、肌肉力分配或跟踪反馈的具体责任，根因停在控制。"
+   }
+  ],
+  "gap_zh": "无状态日志；两视角中侧视部分被起始柱遮挡，右臂100步累计接触的首次达标时刻由俯视接触和持物过程估计，不能精确恢复每个控制步的接触标签。评分器未报告DAG所列五项原生指标，均填null；没有本次交付轨迹的指尖力与目标/实际关节跟踪序列，故无法确定控制根因的子类。"
+ },
+ "summary_zh": "肌肉右手夹住并抬起物体，按DAG的累计接触历史规则估计在视频约 12.5 s 达标，最终保留1/3节点；假肢手交接和目标柱放置均未达到。右手在视频约 16.3 s 未保持住物体，物体滑脱下落，交付回放在471个控制步结束并判失败。缺少逐步接触、力和关节跟踪日志，达标时刻为估计，根因仅归入控制。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与视频一致：右手自视频约 8 s 夹住长方体、13–16 s 抬起、约 16.3 s 滑脱；节点1估计达标时刻（约 12.5 s）在约1 s误差内合理，交接与放置未达成，断点、现象（没保持住）和根因（控制，停在大类）均有证据支持。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:40:27Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--ahkgq8f/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--ahkgq8f/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "96fa2810a9e2d24b6a4673c7cc1f6fca4d7382541625cc4776c2743d75a67a63",
+   "video_sha256": "881dbc4fcab865b390e5376a7bc1f95ac6cb545288763daa048c4983d834c6e7",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-privileged--t8bv7go.json

@@ -0,0 +1,284 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-privileged--t8bv7go",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 29,
+   "agent_resets": 79,
+   "agent_sim_steps": 19832,
+   "agent_records": 9,
+   "agent_replays": 1,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.577,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.25,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-privileged--t8bv7go/replay.mp4",
+  "duration_s": 40.04,
+  "fps": 25.0,
+  "frames": 1001,
+  "sha256": "0df7f7acf7e117d57bcfff8294e3d7a39828886e695836d1ffba13e84d7edc1f"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [
+     [
+      6.5,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 6.5,
+      "log_step": 52,
+      "text_zh": "观察：视频1.25–7.5 s，肌肉右手及腕部贴近并接触红边长方体，物体在起始柱面上倾倒、滑动；视频8.5–16 s，右手移走而物体留在原柱。log step 52 是生成并保存最终交付动作的运行，脚本直接统计其接触历史，输出 myo 122、pros 0、goal 0；log step 53 对该文件回放得到 deterministic=true、1000步失败，与视频相符。"
+     },
+     {
+      "t": 6.5,
+      "log_step": 52,
+      "text_zh": "推断：按本DAG的实际验收，肌肉臂任一部位累计接触122步已超过100步，因此该历史事件达成并保持为真；不把它解释为成功夹起物体。视频共1001帧、25 fps，对应交付1000控制步及初帧，每控制步约0.04视频秒；log step 52 的控制步200物体仍在柱上，与视频8.0 s核对相符。达标时刻估在视频6.5 s附近，误差约1 s；日志只有总计，不能精确重建第100个接触步。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "观察：视频1.25–8.0 s，肌肉右手/腕部与长方体侧面接触，把物体沿起始柱顶推移、压倒；未见手指从相对两侧稳定夹住物体，故不标抓握。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 19,
+      "log_step": 52,
+      "text_zh": "观察：视频8.5–16 s，红边长方体留在起始柱，肌肉右手空手移向两臂之间；视频18–26 s，白色假肢手伸到空交接位并合拢，物体没有进入白手。log step 52 对实际保存运行的接触历史统计为 pros 0，问题为PROSTH_SHORT。"
+     },
+     {
+      "t": 40,
+      "log_step": null,
+      "text_zh": "观察：视频27–40.0 s，长方体始终留在起始柱，白手移向另一侧后没有带回物体。成功参考视频18–20 s可清楚看到长方体从肌肉手转到白手；本次同类视角没有该物体转移，故交接未达成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "其他",
+      "note_zh": "观察：视频19–26 s，白色假肢手在空交接位张合，未接触仍在起始柱的长方体。候选“空抓尝试”：没有实际包围或夹住物体，不能写抓握，也没有两手共同接触物体；现有方式无对应值。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 30,
+      "log_step": 52,
+      "text_zh": "观察：视频27–36 s，白色假肢手空手从交接区移向目标柱一侧；长方体仍可在起始柱上辨认，没有随白手移动或落上另一个目标柱。log step 52 的交付运行接触历史统计 goal 0。"
+     },
+     {
+      "t": 40,
+      "log_step": null,
+      "text_zh": "观察：视频39.0、39.25、39.5、39.75、40.0 s，长方体持续留在肌肉臂旁的起始柱面，未出现在目标柱顶；目标放置关系从未建立。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "其他",
+      "note_zh": "观察：视频27–36 s，白��假肢手执行向目标柱一侧移动的空手放置动作，长方体留在起始柱。候选“空手放置尝试”：手中没有物体，无实际抓握或释放接触，现有接触方式没有对应值。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_handed_to_prosthesis",
+   "t": 8.5,
+   "log_step": 52,
+   "where": "操作中",
+   "evidence_zh": "观察：视频7.5–9.5 s，肌肉右手开始离开而长方体落平并留在起始柱；视频12–26 s，两臂继续执行交接位置的动作，白手在空处合拢，未承接物体，此后也未修复。log step 52 生成并保存的1000步轨迹中，肌肉臂历史接触122步达标，假肢接触0步，最终PROSTH_SHORT；log step 53确定性回放同一文件仍失败。推断：按评分器口径第一历史事件已达成，首个始终未完成的操作步骤是交给左手；视频8.5 s作为物体被留在起始柱的导航锚点。"
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 8.5,
+    "log_step": 52,
+    "evidence_zh": "观察：视频1.25–8.0 s，肌肉右手/腕部贴靠红边长方体并使其在起始柱顶倾倒、滑动；视频8.5–14.0 s，右手抬起并移向交接区，物体仍留在起始柱，未见随右手离柱搬运。log step 52 是生成并保存交付轨迹的运行，控制步300–900的obj位置约为[-0.299,-0.294,1.103]，而palm位置继续变化，累计myo接触122步、pros接触0步。推断：右手虽有推靠接触，却未建立将物体带到交接位所需的夹持或承载接触，导致交接失败；视频19–26 s白手空合拢仅作后续旁证，不作为独立的接收抓取失误。现有证据未显示物体曾被右手持起后再滑落，故保留“没建立接触”，不改为“没保持住”。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "source_zh": "视频 / 日志",
+    "evidence_zh": "推断：所交付承托方案没有形成可随手搬运的承载接触。视频1.25–8.5 s，肌肉手/腕贴靠长方体并将其沿柱面推移，随后手上移、移出而物体留柱；视频12–26 s因此执行空手交接。log step 34 的控制器明确采用自由下垂后把手当托盘的方案，log step 43 将其改为固定初始手姿缓慢上提，log step 51将该策略内联到最终生成脚本，log step 52显示物体在控制步300后停在约[-0.299,-0.294,1.103]，而手继续移向交接区。该接触布局未托住或夹住物体；没有证据把失败进一步归到某个力控增益或逆解误差。成功参考展示任务可完成，但未验证本次冻结初态下的具体替代接触方案。"
+   }
+  ],
+  "gap_zh": "未提供状态日志或逐控制步接触历史；肌肉臂100步达标时间只能由早段画面和最终累计122步估计，视频6.5 s锚点可能相差约1 s。接触标签覆盖整臂，画面不足以逐项区分腕部与各指接触。原生评分reward字段未报告DAG列出的五项指标，native_metrics均留null；log step 52的myo/pros/goal是最终生成脚本的直接诊断，用于节点证据，不冒充reward指标。未验证本冻结初态的替代夹持方案。",
+  "unmet": {
+   "object_held_by_myo": true,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  }
+ },
+ "summary_zh": "本次交付完成肌肉臂累计接触要求（log step 52：122步），按实际评分口径计1/3节点；视频中物体只在起始柱上被推移，没有被带到交接位。断在交给左手：承托/夹持布局未能携物，白手在空处合拢，假肢接触及目标柱接触均为0；视频40.0 s物体仍留起始柱。缺少逐步接触历史，首节点达标时刻只能估计，不能把接触计数达标等同于稳定持起。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 5,
+    "summary_zh": "节点值与断点基本与视频及log step 52/53一致（物体始终留在起始柱，白手空合拢，假肢接触0）。唯一问题：现象“没建立接触”落在白手上，而根因与证据说明失败源于右手没有把物体带到交接位，现象与根因对象不一致。",
+    "n_issues": 1,
+    "n_major": 0
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点值、断点和现象与视频及log step 52/53一致：物体全程留在起始柱，肌肉手只推移，假肢接触0；第一节点按评分器接触计数(122)判达成，已在证据中标为推断。未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:10:23Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--t8bv7go/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--t8bv7go/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--t8bv7go/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--t8bv7go/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7de4f03b070a6e2e56daba31148187658debb8f1ecf0f244bc4a6edf49529d91",
+   "video_sha256": "0df7f7acf7e117d57bcfff8294e3d7a39828886e695836d1ffba13e84d7edc1f",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-privileged--vfsmu3q.json

@@ -0,0 +1,266 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-privileged--vfsmu3q",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 960,
+   "sim_steps": 960,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 149,
+   "agent_resets": 202,
+   "agent_sim_steps": 19515,
+   "agent_records": 7,
+   "agent_replays": 3,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 1,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 11.513,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.246,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 2,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "over_oom": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-privileged--vfsmu3q/replay.mp4",
+  "duration_s": 38.44,
+  "fps": 25.0,
+  "frames": 961,
+  "sha256": "09d683fc1b9ca10d1052541779952bdc0d863c174f69f0ffe8a90bb587c141d4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [
+     [
+      5.5,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 1.4,
+      "log_step": null,
+      "text_zh": "观察：视频约1.4 s起，肌肉右手掌指贴住红边长方体上缘，视频2.8–9.6 s手指弯曲并持续压推物体，物体沿起始柱顶移动，未被稳定抬离柱面。"
+     },
+     {
+      "t": 5.5,
+      "log_step": 8,
+      "text_zh": "推断：按DAG及log step 8的评分代码，此节点接受MYO累计100个控制步接触，不另要求抬升。最终交付有960步、录像有961帧且为25 fps，每步约0.04视频秒；log step 102、103、108给出的最终阶段长度对应假肢手约视频34.8 s开始张手，已与视频35.0–36.0 s画面核对。结合视频1.4–9.6 s持续贴靠及随手推移，估计约视频5.5 s累计达标，首次达标时刻约有1 s不确定性，历史事件此后保留。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "肌肉右手掌指从上缘和侧面压住红边长方体，并沿起始柱顶推移；未见对向夹紧后承重抬起。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 14.4,
+      "log_step": null,
+      "text_zh": "观察：视频9.6–16.3 s，白色假肢左手靠近起始柱、弯曲手指，但手与红边长方体未会合；右侧画面中物体一直留在起始柱，左侧画面可见白手在物体另一侧较远处闭合。"
+     },
+     {
+      "t": 22.56,
+      "log_step": 108,
+      "text_zh": "观察：视频18.2–30.0 s白色假肢手离开时手中没有红边物体，起始柱原位仍有该物体，视频38.4 s亦如此。最终交付的log step 108显示960步实时与回放终态哈希相同且solved=False；该日志只用来确认交付身份，不把其他会话试验当作本次交接。"
+     },
+     {
+      "t": 16.3,
+      "log_step": null,
+      "text_zh": "推断：白手未建立承接接触，也未使物体随其移动，因而此节点未达到；不是已经接到后滑落。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "压/擦",
+      "note_zh": "交接尝试时肌肉右手仍贴压红边长方体，假肢左手同时靠近并空合手指；仅右手实际接触物体，未形成双手抓握或左手承接。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 30.0,
+      "log_step": 108,
+      "text_zh": "观察：视频22.6–30.0 s假肢左手空手向另一侧移动，红边长方体仍留在起始柱；最终交付log step 108的搬运后诊断也显示物体位置约为[-0.3359,-0.2051,1.134]，假肢掌部已经向另一侧移动。"
+     },
+     {
+      "t": 38.4,
+      "log_step": null,
+      "text_zh": "观察：视频31.0–38.4 s，肌肉右手逐渐张开，红边长方体留在同一起始柱顶；未见物体被送到与起始柱分开的目标柱。末秒按约0.12 s间隔检查至视频38.4 s，未出现交接或目标柱放置。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "其他",
+      "note_zh": "候选：空手搬放。白色假肢左手在未接到红边长方体时执行朝目标方向移动及张手撤离���动作，没有对物体的搬放接触；现有抓握、托/舀等值均不适合这次实际空手动作。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_handed_to_prosthesis",
+   "t": 14.4,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约14.4 s，假肢左手在起始柱另一侧合拢，红边物体仍在肌肉右手附近的起始柱顶，没有进入白手；视频18.2–30.0 s白手空手离开，直至视频38.4 s交接仍未修复。右手节点按累计接触规则已达到，最先未修复的是左手接收。"
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 14.4,
+    "log_step": null,
+    "evidence_zh": "观察：视频13.4–16.3 s，白色假肢手弯指位置与红边物体分离；视频18.2–22.6 s假肢手移动后物体仍在起始柱。所需的左手承接接触从未建立，未见一次持有后丢失。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "source_zh": "视频 / 日志",
+    "evidence_zh": "推断：视频2.8–14.4 s肌肉右手以掌指压推物体，未形成承重夹持或把物体送到白色假肢手闭合处；视频14.4–16.3 s两手的作用位置分开，承载与交接接触布局未建立。log step 109中agent自述多种握法均把物体推开、无法抬起，是会话层面的辅助证据；未把其可达性判断或早期控制错误当成本次已证实原因。"
+   }
+  ],
+  "gap_zh": "无状态日志，也没有最终交付的逐步接触集合或累计接触计数；右手100步达标及约视频5.5 s的时刻由画面接触持续性估计，不能精确恢复。log step 109先称右手持续接触超过100帧、后又称两臂条件均未满足，不能据此替代画面判定。白手闭合处部分被肌肉臂遮挡，但起始柱物体始终未被带走、后续白手为空的画面足以判交接失败；无法进一步量化接触力或区分具体逆解与跟踪误差。评分器未报告DAG五项原生指标的数值，均保留null。",
+  "unmet": {
+   "object_held_by_myo": true,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  }
+ },
+ "summary_zh": "本次交付中肌肉右手持续压推起始柱上的物体，按DAG的累计接触验收估计达到右手节点（约视频5.5 s），但未把物体稳定抬起。白色假肢手靠近、空合后空手移开，交接失败且未修复，物体到视频38.4 s仍在起始柱，最终1/3；可见的接触与承载布局不足支持控制层面的抓取和接触设计问题。无状态日志及逐步接触计数，右手首次达标时间存在约1 s不确定性，原生指标未报告。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "右手节点按累计接触规则估计达成（约视频5.5 s）、交接未达成、目标柱未达成，断点在交接（操作中）、现象没建立接触、根因抓取和接触设计，均与视频和log一致；未发现改变读者结论的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:46:46Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--vfsmu3q/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--vfsmu3q/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "e513e7d0aba79491cc50fc620d525d1e78e8c7c87ab6010b3921badc5397469d",
+   "video_sha256": "09d683fc1b9ca10d1052541779952bdc0d863c174f69f0ffe8a90bb587c141d4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-privileged--xxpz8gf.json

@@ -0,0 +1,255 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-privileged--xxpz8gf",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 441,
+   "agent_resets": 381,
+   "agent_sim_steps": 99321,
+   "agent_records": 0,
+   "agent_replays": 7,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 1,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 11.888,
+   "peak_processes": 11,
+   "peak_disk_gb": 1.013,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 2,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "over_oom": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-privileged--xxpz8gf/replay.mp4",
+  "duration_s": 40.04,
+  "fps": 25.0,
+  "frames": 1001,
+  "sha256": "88c70b129aa4784bf3a49dd1a524a738bc9e1e236aaf67d4f12cd8b982f838c6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 11.8,
+      "log_step": null,
+      "text_zh": "观察：视频 0–40.0 s 的两视角中，肌肉右手反复偏离起始柱；视频 11.6–12.2 s 指尖从盒侧经过，未形成对向包围，随后空手离开。视频 24.4–25.6 s、32.4–35.8 s 再次经过时，盒仍由起始柱承托，没有随手搬走。"
+     },
+     {
+      "t": 35.8,
+      "log_step": 273,
+      "text_zh": "推断：逐段检查近物画面后，右手仅有短暂擦碰机会，未达到累计100控制步的接触要求，故本历史事件不记达成；并非由零分或未抬高单独判定。交付文件在 log step 273 的两次回放均为1000步、确定性失败；1001个视频帧对应初态及1000动作，25 fps下100控制步约占视频4 s，已用视频4.4 s附近的合指动作核对 log step 103 的110步合指日程。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "推断：视频11.6–12.2 s肌肉右手指尖沿红边盒侧短暂擦推；没有看到拇指与其他手指对向夹住盒，不能标为抓握。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 18,
+      "log_step": null,
+      "text_zh": "观察：视频 14–21 s 白色假肢手在另一侧空手运动，右手也为空手；视频 24–26 s 和32–35.8 s 盒再次进入画面时仍在原起始柱上，未见白手包围或承接盒并带其移动。"
+     },
+     {
+      "t": 40,
+      "log_step": null,
+      "text_zh": "推断：结合视频 0–40.0 s 两视角的前后位置，未发生物体向假肢臂的交接，也没有足量假肢接触的可见过程；末秒白手仍空手，故该事件未达成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频14–21 s可见白色假肢手执行空手动作，但没有可标注的盒接触；不把空手合指或右手在别处擦盒记为双臂抓握。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 35.8,
+      "log_step": null,
+      "text_zh": "观察：视频 32.4–35.8 s 盒仍在起始柱上，分开的目标柱未承接盒；视频 39.0–40.0 s 右手和白色假肢手均未持盒。全片没有盒从起始柱被送往目标柱的过程，故放置条件始终不成立。"
+     },
+     {
+      "t": 40,
+      "log_step": 273,
+      "text_zh": "观察：log step 273 对最终交付文件两次回放的success均为false，instance_ok和deterministic均为true；这与视频未完成放置一致，输出未报告goal_dist、goal_touch或末尾目标柱独占接触计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频后段白色假肢手为空手姿态，没有看到它持盒放置；无法为脚本的空手放置动作赋予物体抓握接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_held_by_myo",
+   "t": 1.5,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0–1.5 s 肌肉右手从盒上方移向远离起始柱的位置，手中没有盒；视频 9.5–13 s、24–26 s 和32–35.8 s虽再次经过起始柱，仍未完成足量接触及持物搬运。这是最先未修复的步骤，后面的白手空手动作没有修复它。"
+  },
+  "unmet": {
+   "object_held_by_myo": false,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 1.5,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.4–1.5 s 肌肉右手迅速远离仍在起始柱上的盒；视频 2–8 s、14–23 s 多数时间手在空处运动。视频 11.6–12.2 s 即使扫到盒侧也未形成包围，之后空手离开；未看到盒随右手到达交接位置。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 156给出的控制器用掌心雅可比生成关节目标，再经肌肉激活映射执行；log step 169、171的独立跟踪测试中，掌心目标为[-0.522,-0.226,1.292]，实际终点却约为[0.03,0.20,1.73]，误差从约0.115 m增长至约0.83 m。log step 211随后用同一gen.py阶段脚本生成最终1000步文件，log step 273验证该文件确定性失败；评分视频0–8 s及14–23 s的空手大幅偏离与此控制能力缺陷一致。测试属于会话根因证据，不用作节点进度或接触计数；不足以进一步区分反馈控制与接触设计。"
+   }
+  ],
+  "gap_zh": "未提供状态日志，评分器也未输出逐步接触历史和各原生指标，右手累计100控制步是否达标依据近物片段作视觉推断，不能恢复精确计数或接触身体标签；擦推接触也是结合前后画面的推断。末帧盒不在视野内，但视频35.8 s仍在起始柱、随后双手均空手，没有到达目标柱的过程。agent在log step 280自述“扫下起始柱并掉落”，与评分视频中物体反复仍在柱上的观察不一致；不采纳该掉落叙述，log step 273的统一失败原因也不能单独证明掉落。"
+ },
+ "summary_zh": "评分轨迹未完成右手足量接触、交给左手或放上目标柱，最终进度为0/3；肌肉右手反复偏离盒，仅短暂擦过，白色假肢手始终未承接盒。断点在右手持起步骤，现象为位姿不准，会话测试与评分视频共同支持控制问题；无状态接触历史，右手计数及擦推接触为视觉推断，agent关于掉落的最终自述不作为评分轨迹事实。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "三个节点均未达成、断点为右手持起、现象位姿不准、根因控制，与视频0–40 s两视角所见（手多数时间在空处、盒在起始柱附近未被携走）一致；未发现需要改动的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:42:22Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--xxpz8gf/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--xxpz8gf/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "cdf416aa6147381788ac639e8c8b87be64444253baf21749fbc2c1b29eae4a22",
+   "video_sha256": "88c70b129aa4784bf3a49dd1a524a738bc9e1e236aaf67d4f12cd8b982f838c6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-privileged--zrrewr9.json

@@ -0,0 +1,277 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-privileged--zrrewr9",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 350,
+   "sim_steps": 350,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 254,
+   "agent_resets": 6,
+   "agent_sim_steps": 22156,
+   "agent_records": 3,
+   "agent_replays": 4,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.806,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.275,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-privileged--zrrewr9/replay.mp4",
+  "duration_s": 14.04,
+  "fps": 25.0,
+  "frames": 351,
+  "sha256": "ccfdfc762fca1e96e30452e8e4f230ed38fbbd886db77c0e627d7227a4210ab2"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [
+     [
+      5.7,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 2.4,
+      "log_step": null,
+      "text_zh": "观察：已检查视频 1.2–4.4 s 的密集画面；视频约 1.7 s 起，肌肉右手指端接触红边灰色盒并使盒体倾斜，视频 2.4–4.4 s 盒体继续位于手指下方，随后保持接触并随手移动。该节点按实际评分的累计接触判定，不另要求先明显抬高。"
+     },
+     {
+      "t": 5.7,
+      "log_step": 411,
+      "text_zh": "推断：log step 411 对交付文件逐动作重放，统计肌肉臂接触共238个控制步，超过100步；log step 407、418 对同一350步文件回放的终态哈希一致且 deterministic=true。视频约 1.7 s 起已见右手接触并压斜盒体，之后画面支持持续接触，因此以此作为累计起点估计：100个控制步约为4.0视频秒，首次达标约在视频 5.7 s，估计不确定性约±0.7视频秒。350个动作对应351帧、25 fps，即每控制步约0.04视频秒；log step 405 的交付脚本从控制步130开始向滑移目标插值，对应视频约 5.2 s，已与视频 5.0–5.6 s 盒体转动并向柱缘移动核对。未换算agent工作时间；总计数不证明每步均连续接触。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "肌肉右手先压倒盒体，随后拇指与其余手指夹住红边灰色盒，沿起始柱面向外移动；没有把柱面上的移动误作额外抬升验收。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 12.8,
+      "log_step": 411,
+      "text_zh": "观察：视频 12.0–14.0 s，两手在起始柱外侧会合，盒体下缘接触白色假肢手伸出的指尖；白手没有合拢包住盒体，也没有携盒离开交接处。log step 411 对交付文件逐动作重放，假肢臂累计接触只有33个控制步，未达到100步；这是已由log step 407、418确认确定性的交付文件，非较早试验的计数。"
+     },
+     {
+      "t": 14.0,
+      "log_step": 410,
+      "text_zh": "观察：视频 13.28–14.0 s 盒仍斜悬在起始柱旁和白手指尖处。log step 410 重放同一交付文件，终态为350步、obj_pos=[-0.185,-0.266,1.05]、touching_body=[0,1,0,0,0]、success=false、done=false，哈希与交付回放一致；末帧存在假肢接触不等于累计100步交接达标。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "视频末段肌肉右手夹持盒体送向白色假肢手，两手在盒体处会合；此处抓握指肌肉右手，白手未形成包围夹持。"
+     },
+     {
+      "who": "双臂协作",
+      "how": "托/舀",
+      "note_zh": "白色假肢手保持手指伸直，在右手送出的盒体下缘以指尖承接；可见接触限于指尖附近，未托入掌内。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 14.0,
+      "log_step": 410,
+      "text_zh": "观察：已检查视频全程以及视频 13.0–14.0 s 的密集末段，盒体从起始柱向白手移动后始终停留在交接侧，未被白手带到另一目标柱顶���也未出现放柱后双手离开的状态。log step 410 的交付回放终态仍处于交接坐标，success=false；成功参考视频约24–31 s中另柱附近的放置过程仅用于辨认目标位置，不作为本次完成的证据。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_handed_to_prosthesis",
+   "t": 13.3,
+   "log_step": 410,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 12.8–14.0 s，肌肉右手把盒送出起始柱，白手伸指在盒底承接，但盒体只落在伸直的指端附近，未进入掌内，未随白手搬走。log step 411 对交付350步文件的回放给出假肢臂33步接触，未达该步100步验收；log step 418 两次确定性回放均success=false，且没有后续修复。"
+  },
+  "unmet": {
+   "object_held_by_myo": true,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 13.3,
+    "log_step": 410,
+    "evidence_zh": "观察：视频 12.8–14.0 s，白色假肢手掌在盒体侧下方，伸直的指尖只碰到盒体下缘，盒体在柱缘外向下倾斜，始终没有送入白手掌内的承接区域。log step 410 的交付文件终态仍仅有假肢臂接触；这里是交接承接位置偏差，末帧没有显示物体已经掉到地面。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 405 的交付生成脚本在接收阶段持续使用固定q_prosth_bridge，全部350个控制步均令假肢手指命令p_fingers=-1，未在盒到达后加入合指或继续托入掌内的动作；视频 12.8–14.0 s 也显示白手仅用伸直指端承接、盒体位于掌外。该接触布局未把物体送入足够的承载区域，交付回放的假肢接触仅33步（log step 411），因此归因于接触设计；没有在本次交付中验证替代接收布局。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或交付轨迹的逐步接触计数；依据视频约 1.7 s 起的可见接触及后续夹持、每控制步约0.04视频秒，将右臂累计100步的首次达标估计为视频约 5.7 s，估计不确定性约±0.7视频秒。日志可确认最终238/33步，但不能精确恢复每次接触的起止或排除短暂中断。评分器reward未报告DAG五项原生指标，native_metrics均保留null。视频在白手指尖承接中结束，不能把会话其他测试的后续掉落当成本次评分事件。"
+ },
+ "summary_zh": "本次交付中，肌肉右手夹持并将盒体移出起始柱，按累计接触验收完成右手节点，最终进度1/3；白手仅在指尖处承接，交付回放接触33步，交接和目标柱放置均未达标。断点在交给左手：盒体未送入掌内，固定接收姿态与持续张开的手指缺少进一步承接布局。没有逐帧接触日志，按视频约 1.7 s 起接触并累计100控制步推断，右手达标约在视频 5.7 s（估计不确定性约±0.7视频秒），其他会话试验中的掉落未计入本次评分轨迹。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点2/3、断点（交给左手·操作中）、位姿不准与接触设计的标注与视频和 log 一致；主要问题是右手节点达标时刻 7.5 s 的推算起点错误，视频显示右手自约 1.7 s 起已持续接触盒体，100 步累计应在约 5.5–6 s。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "第一轮已修正右手节点时刻；节点 1/3、断点（交给左手·操作中）、位姿不准与接触设计均与视频和 log 一致，未发现新错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:11:43Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--zrrewr9/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--zrrewr9/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--zrrewr9/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-privileged--zrrewr9/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "12c572cb4c56beac43446f9705b9ea260684f222232f45185ce6c4487509c3f1",
+   "video_sha256": "ccfdfc762fca1e96e30452e8e4f230ed38fbbd886db77c0e627d7227a4210ab2",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-standard--9hfsnqg.json

@@ -0,0 +1,273 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-standard--9hfsnqg",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 919,
+   "sim_steps": 919,
+   "physics_steps": 4595,
+   "steps": 1909,
+   "calls": 35,
+   "episodes": 4,
+   "resets": 3,
+   "reset_budget": 10,
+   "sim_steps_total": 1909,
+   "elapsed_s": 2700.9,
+   "video_rendered": 1,
+   "task_failure": 0,
+   "timeout": 1,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 4.075,
+   "peak_processes": 34,
+   "peak_disk_gb": 0.402,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-standard--9hfsnqg/replay.mp4",
+  "duration_s": 36.8,
+  "fps": 25.0,
+  "frames": 920,
+  "sha256": "637bc841d93aa9f147b055e3ab5c8f1fb01c098d5b9ceb91ed0beab4517d2add"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": null,
+  "peak": null,
+  "final": null,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "unknown",
+    "intervals": null,
+    "final": null,
+    "evidence": [
+     {
+      "t": 3.6,
+      "log_step": null,
+      "text_zh": "观察：视频 3.0、3.6、5.0、7.0 s 的俯视画面显示肌肉右手位于红边长方体一侧，拇指贴近其上缘，其余手指在柱顶外侧；另一视图的接触区域被起始柱遮挡，不能分辨拇指与物体是否持续实接触。"
+     },
+     {
+      "t": 14.0,
+      "log_step": null,
+      "text_zh": "观察：视频 8.0–14.0 s 右手合指并移开，视频 14.0、18.0、22.0、24.0 s 红边长方体仍在起始柱上，右手中没有物体。物理持起没有发生；但本节点按肌肉臂累计接触至少100个控制步验收，并不要求抬升，故不能仅凭空抓把接触计数判为不达标。"
+     },
+     {
+      "t": null,
+      "log_step": 79,
+      "text_zh": "观察：log step 79 将动作序列发送到重置后的服务，服务状态为 episode=4、control_steps=919；log step 81 仍为同一局同一控制步数。log step 78 的 myo=586、pros=184 来自本地仿真，不是该受评局的接触计数，不能据此判本节点达标。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "肌肉右手在视频 3.0–12.0 s 于物体旁合指；拇指是否实接触、如何施力不能由柱体遮挡后的两视图可靠确认，未见包住物体并持起，不标为实际抓握。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 24.0,
+      "log_step": null,
+      "text_zh": "观察：视频 14.0、18.0、22.0、24.0 s 物体留在起始柱，肌肉右手已空手移向两臂之间；视频 26.0–32.0 s 白色假肢手掌朝上来到右手下方，两手会合时均未携带红边长方体。"
+     },
+     {
+      "t": 36.76,
+      "log_step": null,
+      "text_zh": "推断：视频 33.0–36.76 s 白手空手离开，未见物体进入其手掌或随其移动；结合视频 24.0 s 仍在起始柱的物体及此前右手空手搬运，假肢臂没有获得该物体的100步接触，节点未达成。成功参考录像在参考视频 18.0–21.0 s 可见红边长方体在两手间转移，本局没有该可见变化。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "视频 26.0–32.0 s 肌肉右手与白色假肢手执行空手会合；没有物体进入两手之间，未建立可标为抓握或托/舀的物体接触。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 24.0,
+      "log_step": null,
+      "text_zh": "观察：视频 24.0 s 红边长方体仍在起始柱；视频 26.0–33.0 s 白色假肢手执行承接、移动及撤手动作时，掌中没有物体。"
+     },
+     {
+      "t": 36.76,
+      "log_step": 79,
+      "text_zh": "推断：已逐段检查视频 0–36.76 s，含视频 35.8、36.0、36.2、36.4、36.6、36.76 s 的末段画面，没有物体被持离起始柱或送至目标柱的过程；所以未形成目标柱放置关系。评分器 success=0 与此一致，但不是仅凭总分判该节点为假。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "视频 32.0–36.76 s 白色假肢手空手移开，没有可分类的手与物体的搬运或释放接触；后续空手动作不能记为实际抓握放置。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_held_by_myo",
+   "t": 12.5,
+   "log_step": 79,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 8.0–12.5 s 肌肉右手合指后开始离开起始柱，视频 13.44–14.0 s 手与红边长方体明显分离，物体仍在柱面；视频 18.0–24.0 s 也没有补抓，之后两手执行空手交接。最早未修复的是右手持起的实际操作；该节点的历史接触阈值本身仍因计数缺失保留未知。"
+  },
+  "unmet": {
+   "object_held_by_myo": null,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 12.5,
+    "log_step": null,
+    "evidence_zh": "观察：视频 8.0–14.0 s 右手在物体旁屈指后空手移开，红边长方体没有随手离开起始柱；视频 18.0–24.0 s 物体仍在柱面。没有建立能带起物体的对向夹持；这不排除拇指曾有轻微擦碰，也不是已持起后滑落。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 53 的服务控制脚本将本地计算的80执行器命令按100行一批发送，循环中没有使用摄像头或本体反馈修改抓取，只在全部命令后取图；log step 79 实际运行该脚本形成 episode 4。视频 12.5–24.0 s 的空手右臂仍继续搬运，视频 26.0–32.0 s 白手继续空接，说明受评控制序列没有在抓取未建立时纠偏。log step 82 中 agent 也自述本地成功抓取没有在服务复现；具体是接触几何还是执行跟踪差异，现有证据不能再细分。"
+   }
+  ],
+  "gap_zh": "输入没有状态日志，评分器没有报告两臂累计接触计数及各项原生指标。视频 3.0–12.0 s 的右手指与物体上缘在俯视图中紧邻，另一视图被柱体遮挡；无法可靠区分持续接触与间隙，故右手100步历史接触事件为未知，而不是用未抬升代替评分规则判假。左手交接与放置可由物体始终留在原处、两手空手执行后续动作判未达成。log step 78 的接触计数和掉落是本地测试，不用于本局进度或现象；log step 82 猜测物体质量、摩擦差异未获独立验证，未作为环境根因。"
+ },
+ "summary_zh": "受评的是 log step 79 运行的最后一局 episode 4：右手在起始柱旁合指后空手移走，物体没有被持起，随后假肢手执行空手承接与撤离，交接和目标柱放置均未达成。断点为右手持起操作，现象是未建立有效夹持；脚本按预计算命令继续执行而没有根据抓取失败纠偏，根因标为控制。没有受评局的接触计数，且关键近接触面被柱体遮挡，故按实际100步接触验收的右手事件保留未知，确认完成节点为0/3。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "逐帧核对后，右手在视频 2–12 s 于物体旁合指、14 s 起空手离开、物体留在起始柱，与文件一致；两个下游节点为假，评分器 success=0 且无状态日志，对第一个节点标 unknown 有依据；断点、现象（没建立接触）与根因（停在“控制”）有证据支持，未发现需要改动的问题。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:41:32Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--9hfsnqg/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--9hfsnqg/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "6d41a14e3c278ff5326ff58a2507d187a92d1598cbd69da52878fe515e45fcd2",
+   "video_sha256": "637bc841d93aa9f147b055e3ab5c8f1fb01c098d5b9ceb91ed0beab4517d2add",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-standard--d2hna7y.json

@@ -0,0 +1,282 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-standard--d2hna7y",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 450,
+   "sim_steps": 450,
+   "physics_steps": 2250,
+   "steps": 460,
+   "calls": 19,
+   "episodes": 2,
+   "resets": 1,
+   "reset_budget": 10,
+   "sim_steps_total": 460,
+   "elapsed_s": 2601.0,
+   "video_rendered": 1,
+   "task_failure": 0,
+   "timeout": 1,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 2.754,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.389,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-standard--d2hna7y/replay.mp4",
+  "duration_s": 18.04,
+  "fps": 25.0,
+  "frames": 451,
+  "sha256": "8922835744863e63ef6805b2c29bef1c8197f7423a5be4ef51c1b8bb6bc9860a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": null,
+  "peak": null,
+  "final": null,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [
+     [
+      5.5,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": null,
+      "text_zh": "观察：视频0.8–2.2 s，肌肉右手弯曲手指接触红边盒体，盒体随手指运动沿起始柱面转动、滑移；视频4.8–18.0 s，盒体停在起始柱缘，仍贴靠肌肉臂腕部，未完成脱柱搬运。"
+     },
+     {
+      "t": 5.5,
+      "log_step": 344,
+      "text_zh": "推断：按DAG的实际验收，肌肉臂任一部位累计接触100个控制步即可，不另要求抬离柱面。视频约0.8 s开始接触，随后手指至腕部的接触延续，约视频5.5 s已累计足量接触（达标时间为视觉估计，误差约0.5 s）；log step 344的评分局提交450控制步，与451帧、25 fps、视频18.04 s对应，100步约占视频4 s，已核对视频18.0 s的末态。未用agent工作时间换算。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约0.8–2.2 s，右侧肌肉手弯指包夹红边盒体；盒体仍由起始柱支承，未见稳定抬离。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频约4.8–18.0 s，盒体由起始柱缘支承并贴靠右侧肌肉臂腕部；此处为腕部抵压，不能据此称为持续抓握。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "unknown",
+    "intervals": null,
+    "final": null,
+    "evidence": [
+     {
+      "t": 6.4,
+      "log_step": 344,
+      "text_zh": "观察：视频3.6–6.4 s，白色假肢手伸向右手下方；视频6.4–18.0 s，两手停在起始柱边，盒体仍贴靠肌肉臂腕部和起始柱缘，没有随白手离开。白手指尖与盒体之间的局部区域被肌肉手腕及柱体遮住，不能逐帧核实微小接触。"
+     },
+     {
+      "t": 18.0,
+      "log_step": 267,
+      "text_zh": "观察：log step 267的415步肌肉接触、128步假肢接触来自agent自行建立的本地MuJoCo测试；log step 273的假肢累计达标也来自本地测试，并非评分局服务读取。log step 345与346分别记录远端及本地末态，肘屈曲读数分别为1.2775与1.3959 rad，不能视为同一轨迹。"
+     },
+     {
+      "t": 18.0,
+      "log_step": 358,
+      "text_zh": "推断：视频明确未展示物体由白手接走，但DAG实际只验收假肢臂累计接触100控制步，缺少搬运不等于该计数未达标；局部接触被遮挡且没有评分局接触历史，因此保留未知，不采用agent在log step 358引用的本地128步作为节点值。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频6.4–18.0 s的交接尝试中，盒体仍抵在肌肉右臂腕部和起始柱缘，右侧作用端没有携物离开。"
+     },
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频5.8–7.8 s及末段，白色假肢手伸展指尖靠近物体所在的右腕下方，接触界面被右腕及柱体遮挡，无法确认白手是否有轻触���其方式；未见白手包夹并携走盒体。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 18.0,
+      "log_step": 357,
+      "text_zh": "观察：视频0.0–18.0 s中，盒体从起始柱面移到同一柱的边缘后停住；视频17.0、17.2、17.4、17.6、17.8和18.0 s均未出现白手将盒体送上另一根目标柱并离开的动作。log step 357的评分局服务状态仍为control_steps=450、done=False；评分器最终success=0。"
+     },
+     {
+      "t": 6.4,
+      "log_step": 354,
+      "text_zh": "观察：log step 344提交的动作及log step 354保存的脚本只让假肢手插值到交接位并维持，没有目标柱运输或释放阶段；视频中也未发生向另一根目标柱的携物运动。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_handed_to_prosthesis",
+   "t": 6.4,
+   "log_step": 344,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约6.4 s，白色假肢手到达肌肉右手下方后未包住或接走盒体；盒体继续留在起始柱缘和右腕处，直至视频18.0 s。log step 344是评分局唯一450步动作提交，随后log step 355、357均显示仍停在control_steps=450。推断：最早未修复的物理操作停在交接承接处；假肢累计接触是否已达100步另保留未知，不能将本地测试达标当作交接成功。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 6.4,
+    "log_step": 344,
+    "evidence_zh": "观察：视频5.8–7.8 s及视频10.0–18.0 s，白色假肢手保持张开，伸展手指位于肌肉右手的下方，红边盒体则停在右腕上侧的起始柱缘，没有进入白手可包夹的位置，也没有随白手移动。推断：接收手与盒体的相对位置未形成有效承接；不据此断言假肢从未发生过轻触。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "观察：log step 344在整段评分动作中将假肢手10个手指执行器目标设为-1.0并维持固定交接臂姿；log step 354的对应脚本明确将此设定注释为张开、伸展手指。视频约6.4–18.0 s显示白手伸指停在右手下方，物体仍依靠起始柱缘与右腕，没有被白手承载。推断：实际接收姿态与指形未建立能接走物体的承载或对向夹持布局；现有证据未验证本次相同状态下的替代动作。"
+   }
+  ],
+  "gap_zh": "没有状态日志，也没有评分局逐步接触标签或两臂累计接触计数；白手与盒体的微小接触界面被右腕及起始柱遮挡，因此假肢接触100步节点无法定值。log step 193、267、273的接触计数是本地测试，log step 345与346证实本地和远端关节状态存在差异，不能转用其计数或达标时刻；肌肉臂节点达标时间为视觉估计。评分器未报告DAG所列五项原生指标，均记null。",
+  "unmet": {
+   "object_held_by_myo": true,
+   "object_handed_to_prosthesis": null,
+   "object_on_goal_pillar": false
+  }
+ },
+ "summary_zh": "评分局中，肌肉右手接触并将盒体沿起始柱面移至柱缘，按累计接触的实际验收估计完成右臂节点；白手靠近后没有携物承接，目标柱放置未完成。断点在交接操作：接收手保持张开并停在右手下方，未形成有效承载或夹持布局。假肢累计接触是否达到100步因界面遮挡及缺少评分局接触历史而未知，agent报告的415/128步来自本地测试，不能用于评分局计数。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "抽查视频0.5–18 s各帧：盒体始终在起始柱上，右手肌肉臂接触，白手伸到右手下方未携物，未上目标柱；节点值、断点（交接）、位姿不准与抓取和接触设计根因与证据一致，假肢节点保留未知合理。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:45:43Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--d2hna7y/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--d2hna7y/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0ef4f239ba737e2c7bb6d42e9785d169ee873f4508ff7fa7c1579fbe4d170951",
+   "video_sha256": "8922835744863e63ef6805b2c29bef1c8197f7423a5be4ef51c1b8bb6bc9860a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-standard--fyvhoer.json

@@ -0,0 +1,259 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-standard--fyvhoer",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "physics_steps": 5000,
+   "steps": 1078,
+   "calls": 5,
+   "episodes": 3,
+   "resets": 2,
+   "reset_budget": 10,
+   "sim_steps_total": 1078,
+   "elapsed_s": 3170.0,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 2.306,
+   "peak_processes": 12,
+   "peak_disk_gb": 0.601,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-standard--fyvhoer/replay.mp4",
+  "duration_s": 40.04,
+  "fps": 25.0,
+  "frames": 1001,
+  "sha256": "c042cf5f258790fa973b7eb0c38d184a1621a8acbfd1f2be86bd2bd856f744d4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.5,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 2.5,
+      "log_step": null,
+      "text_zh": "观察：视频约0.5–5.5 s的逐0.25 s双画面中，右侧红色肌肉手的弯曲手指及掌部贴压红边长方体，物体倾倒后仍由起始柱支承；视频5.75–6.5 s右手撤走，物体留在原柱，未见持物抬升。"
+     },
+     {
+      "t": 4.5,
+      "log_step": 65,
+      "text_zh": "推断：log step 65提交的是最后第3局1000个控制步，视频1001帧、25 fps，故每控制步约对应视频0.04 s；脚本在控制步140切换肌肉命令，与视频约5.6 s右手开始撤离相符。可见贴压从视频约0.5 s持续至5.5 s，累计100步约在视频4.5 s达到，时间误差约0.5 s；按本DAG实际累计接触验收记达成并锁存，而不额外要求真正抬起。此判断不把本地优化输出的116步当成本局计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频约0.5–5.5 s，右侧肌肉手的手指及掌部压碰起始柱上的红边长方体；物体受压倾倒但没有被合手夹住并带离柱面。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 6.0,
+      "log_step": 65,
+      "text_zh": "观察：视频5.75–6.5 s，肌肉右手空手撤离，红边长方体仍在起始柱；视频0–40.0 s的全片粗查及视频39.04–40.0 s末段细查，白色假肢手始终远离起始物体，没有双手会合、承接或携物移动。log step 65最终提交代码将假肢17个执行器在全部1000步均设为初始关节位置的保持命令，只切换右臂肌肉命令。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 65,
+      "text_zh": "推断：右手离开前物体未被取走，此后两手无交接动作，且最终提交脚本无假肢接收阶段，因此假肢累计接触条件不成立；并非仅由总成功分数0推断。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 6.5,
+      "log_step": 65,
+      "text_zh": "观察：视频6.0–6.5 s，红边物体仍留在起始柱，肌肉手已空手离开；后续至视频40.0 s，两手没有携物到另一根柱顶再松开的动作。成功参考视频约17–25 s可见两手靠拢及向另一柱运送的动作，本局没有相应过程；参考只用于辨认可见线索。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 65,
+      "text_zh": "推断：本局未发生离开起始柱后的搬运，假肢手全程执行固定保持命令，故物体放上独立目标柱及柱独占接触尾窗均未达成；视频约7 s后原柱离开手部相机视野，不声称末帧直接显示物体位置。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_handed_to_prosthesis",
+   "t": 5.7,
+   "log_step": 65,
+   "where": "首步前",
+   "evidence_zh": "观察：视频约5.7 s肌肉右手撤走而物体留在起始柱，白色假肢手没有迎向或接触物体，此后至视频40.0 s也未开始交接。log step 65最终第3局仅提交右手触碰后撤离及假肢定姿保持命令；此前log step 64的第1局在78步结束后已被重置，不属于本视频。断点为未启动且未补做的交接步骤，视频5.7 s仅为右手撤离的导航锚点。"
+  },
+  "unmet": {
+   "object_held_by_myo": true,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 5.7,
+    "log_step": 65,
+    "evidence_zh": "观察：视频约0.5–5.5 s只有肌肉右手在起始柱处触碰物体，视频5.75–6.5 s空手撤离；白手在视频全程及39.04–40.0 s末段都未伸向物体，未启动交接。log step 65的最终提交中，假肢所有执行器始终保持初始关节目标，没有接收动作。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：会话log step 44、50的本地肌肉抓持搬运测试反复输出post=0，log step 62–63的完整本地控制器测试仍输出PROSTH_SHORT(0)、NO_GOAL，物体没有被送到假肢接收处；这些是控制开发的会话证据，不是本视频的状态计数。其后log step 65明确换成只触碰物体、假肢定姿保持的最后一局，与视频5.75–6.5 s空手撤离相符，表明未形成可执行的交接控制而退回触碰基线。现有证据不足以进一步区分抓取几何、逆解或跟踪问题；不将用时、先前局的掉落或agent关于工作空间不可达的自述当作本局物理现象。"
+   }
+  ],
+  "gap_zh": "无状态日志，评分器字段未报告本局两臂精确接触计数、goal_dist、goal_touch或末尾柱独占步数；右臂100步达标及视频约4.5 s时刻由持续可见接触和每控制步对应一视频帧推断，可能相差约0.5 s。视频约7 s后起始柱离开两路手部相机视野，末帧不能直接核实物体位置，但此前物体留柱、全程无交接动作及最终固定假肢命令足以判定后两节点未达。日志中的本地副本测试与被重置的第1局仅作控制开发背景，不用于给本局计数；控制根因无法细分，agent自述假肢工作空间限制也未独立证实。"
+ },
+ "summary_zh": "受评的是log step 65最后重置后的第3局：右手在视频约0.5–5.5 s贴压起始柱上的物体，按累计接触100步的实际验收，第一节点推断于视频约4.5 s达成，虽未真正持起，最终进度为1/3。视频约5.7 s右手空手撤离，假肢手未启动交接，后续放上目标柱也未达；会话控制测试未形成可用交接控制，最后退回了只触碰的基线。无本局精确接触计数，且视频约7 s后原柱不在视野内，因此第一节点时刻与控制细分原因仍有不确定性。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "标签与视频、log一致：右手视频约0.5–5.5 s持续贴压起始柱上的物体（约125控制步，足以支持100步累计的推断），白手全程未动，后两节点未达，断点首步前+没尝试+控制（子类不明）均有证据支持，未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:41:56Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--fyvhoer/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--fyvhoer/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "d15636894ab7542080ae32f6bcb2ba5e47834a5ffe5bde78a232f5176e65b520",
+   "video_sha256": "c042cf5f258790fa973b7eb0c38d184a1621a8acbfd1f2be86bd2bd856f744d4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-standard--mnb62cw.json

@@ -0,0 +1,254 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-standard--mnb62cw",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 739,
+   "sim_steps": 739,
+   "physics_steps": 3695,
+   "steps": 6182,
+   "calls": 2321,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 6182,
+   "elapsed_s": 5362.8,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-standard--mnb62cw/replay.mp4",
+  "duration_s": 29.6,
+  "fps": 25.0,
+  "frames": 740,
+  "sha256": "05941e6d344104932f5ddd781b41841fe3d1c0eb856f11b7244bdcf61e2ad826"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [
+     [
+      9.8,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 5.8,
+      "log_step": null,
+      "text_zh": "观察：视频约5.8–11.6 s，肌肉右手拇指持续抵住直立物体的上缘，物体仍由起始圆柱承托；视频11.8–13.6 s，其他手指进一步收拢到物体上缘两侧。"
+     },
+     {
+      "t": 9.8,
+      "log_step": 231,
+      "text_zh": "推断：本节点依DAG实际验收口径只需MYO累计接触100个控制步，不另要求抬升。该局739个控制步对应740帧、25 fps，视频时间约为control_steps/25；log step 231的334步对应视频13.36 s，与视频13.4 s的收指结束、开始抬手核对一致。视频约5.8 s起的持续触碰到视频约9.8 s已覆盖约100步，故以视频9.8 s为近似达标锚点（约±0.5 s），事件随后保留；这不表示已抓稳或带离起始柱。"
+     },
+     {
+      "t": 14.2,
+      "log_step": 236,
+      "text_zh": "观察：视频13.4–15.4 s，肌肉右手抬起时物体从上缘夹持处转落，平躺回起始柱；视频18.4–23.2 s的再次收指、抬手也未带走物体。log step 236中agent自述“The lift failed again, and the box is now flat on the pillar”，与该局视频一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "肌肉右手先以拇指抵压物体上缘，随后拇指与其余手指在物体上缘两侧收拢作边缘夹持并尝试抬升；夹持未能稳定承重，物体回落到起始柱。末段右手再次收指并拨动物体，使其离开柱缘。仅右手与物体发生此接触，白色假肢手未共同夹持。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 17.2,
+      "log_step": 235,
+      "text_zh": "观察：视频16.0–24.0 s的连续抽帧中，白色假肢手向身体附近的接收区抬起，但与起始柱上的红边物体始终分离；肌肉右手仍在起始柱处反复收指重抓。log step 235仅令假肢手移向接收姿态，没有已完成交接的服务读数。"
+     },
+     {
+      "t": 29.56,
+      "log_step": 242,
+      "text_zh": "观察：视频24.8–29.56 s，白手仍远离物体；物体先留在起始柱，再从柱缘掉落，未随白手移动。没有成功参考视频约17–20 s中两手在物体处会合并由白手持物的可见过程；本局未开始实际的物体交接，假肢臂累计100步物体接触的条件不成立。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 15.4,
+      "log_step": null,
+      "text_zh": "观察：视频0–27.6 s，物体一直在肌肉右手旁的起始圆柱上，被触碰、夹住上缘并翻倒；没有被运往另一目标柱。"
+     },
+     {
+      "t": 29.56,
+      "log_step": 242,
+      "text_zh": "观察：视频27.8–29.56 s，物体越过起始柱边缘并向下掉落，末帧肌肉手为空，白手没有持物，也没有物体留在目标柱上的画面。log step 242的服务状态为episode 11、control_steps 739、done True、reason为“failure: object dropped or the time limit passed”；评分记录task_failure=1、timeout=0，与视频掉落一致。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_held_by_myo",
+   "t": 14.2,
+   "log_step": 232,
+   "where": "操作中",
+   "evidence_zh": "观察：视频13.4–15.4 s，肌肉右手正进行收指后的抬升检验，物体却从上缘夹持处转落并平躺回起始柱；log step 230的脚本在收指后命令上移，log step 232仍在等待这次抬升结果。视频18.4–23.2 s及24.8–29.56 s的重抓均未修复承重夹持，最后物体跌出柱缘。断点按实际正在操作的右手拿取步骤记录；该节点较弱的累计触碰验收已达标，并不等于物理抓取已成功，也不能据此把断点移到尚未开始的交接或目标柱放置。"
+  },
+  "unmet": {
+   "object_held_by_myo": true,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 14.2,
+    "log_step": 236,
+    "evidence_zh": "观察：视频11.8–13.6 s，肌肉右手拇指与其余手指从两侧夹住物体上缘；视频13.8–15.4 s抬手时，物体相对手指转动、滑脱，仍落在起始柱面上，未随手移动到接收区。视频27.8–29.56 s的最后一次抓取后物体又跌出柱缘，白手未承接；log step 236中agent自述抬升失败，与前一段画面一致。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "t": 13.4,
+    "log_step": 230,
+    "source_zh": "视频",
+    "evidence_zh": "推断：视频11.8–14.4 s显示肌肉右手主要捏住长方体上缘的小片区域，没有形成能把整个物体随手带走的承载夹持；手上移时物体在夹持处转动并脱离。log step 230的脚本在收指后直接安排抬升，log step 236、241中agent分别自述抬升失败和只勾到边缘，与视频18.4–23.2 s的边缘重抓、空手抬起相符。证据支持接触布局不能稳定承重；未在本局验证一套可行替代握法，不能进一步断定具体肌肉参数或接触力数值。"
+   }
+  ],
+  "gap_zh": "未提供状态日志或逐控制步接触标签，评分字段也未报告五项原生指标；右臂100步累计接触的达标时刻由可见持续触碰及逐控制步渲染映射估计，无法精确恢复接触计数、力或三维距离。右手接触达标与实际拿取失败并存，源于DAG已说明的验收口径差异；没有证据把本次物理滑脱归因于任务设计。"
+ },
+ "summary_zh": "第11局的肌肉右手持续触碰物体，按DAG累计接触100个控制步的实际口径，右手节点约在视频9.8 s达标并保留，最终进度为1/3；这不代表物体已被稳定持起。实际操作在视频约14.2 s的右手抬升中出现边缘夹持滑脱，之后重抓未修复，物体在视频27.8–29.56 s跌出起始柱，未交给白手或放上目标柱，根因为抓取和接触设计。无状态日志和接触计数，达标时刻及接触布局的解释含视觉推断，原生指标均未报告。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点值、断点、现象与根因与视频一致：右手约5 s起触碰物体、13.4 s后抬升时滑脱、末段物体跌出柱缘，白手始终未接触物体。节点达标时刻属DAG口径下的合理推断，已标为推断。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:43:28Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--mnb62cw/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--mnb62cw/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "04ce3f1cd0b995c421b1e1439b3d0bb510fb992852a2893dbb693ba133a34bb9",
+   "video_sha256": "05941e6d344104932f5ddd781b41841fe3d1c0eb856f11b7244bdcf61e2ad826",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-standard--x4egqsj.json

@@ -0,0 +1,260 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-standard--x4egqsj",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "physics_steps": 5000,
+   "steps": 3400,
+   "calls": 79,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 3400,
+   "elapsed_s": 1609.1,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.706,
+   "peak_processes": 14,
+   "peak_disk_gb": 1.548,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-standard--x4egqsj/replay.mp4",
+  "duration_s": 40.04,
+  "fps": 25.0,
+  "frames": 1001,
+  "sha256": "12b02f935fc7bc59b6cee67f5d363b9657fdd9ee80280e5d0a2f57c756e4972b"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.0,
+      "log_step": 182,
+      "text_zh": "观察：视频 0.0–2.0 s 的双画面中，红色肌肉右手张开并从红边灰盒旁向下离开，盒子仍立在起始柱顶，没有随手抬升或被包夹。log step 182 的最后一次重置后先提交200步恒定肌肉命令；该步后处理报错，但已提交的运动属于第11局。"
+     },
+     {
+      "t": 33.2,
+      "log_step": 192,
+      "text_zh": "观察：视频 30.4–38.4 s 的接近段中，右手在盒子下方或侧方摆动；视频 33.2、34.8、35.6、36.8、38.4 s 可见盒子仍在原起始柱上，右手未形成包夹，未见持续的物体接触。视频 39.0–40.0 s 双手仍空着。"
+     },
+     {
+      "t": 30.4,
+      "log_step": 192,
+      "text_zh": "推断：本局1000个控制步对应1001帧、25 fps，视频每步约0.04 s；log step 192 从本局control_steps=760开始接近，对应视频 30.4 s，已用视频 30.4–30.8 s 手臂从保持转为摆动核对。全局未见可累计达100步的右臂接触，故按DAG的历史接触验收判未达，而非仅因没有抬高物体判未达。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "其他",
+      "note_zh": "候选名“未建立接触的接近尝试”：视频 30.4–38.4 s 肌肉右手手指向起始柱上的红边灰盒接近，但未见夹住或推动盒子；不能标为实际抓握、推/扫或压/擦。短暂擦碰无法从画面排除，未虚构接触计数。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 35.6,
+      "log_step": null,
+      "text_zh": "观察：视频 25.0–30.4 s 以及视频 33.2、35.6、36.8、38.4 s，盒子仍在起始柱，白色假肢手与盒子分离；视频 39.0–40.0 s 白手仍空着。全片未见物体随白手运动，亦未见足量的白臂接触。成功参考录像的视频 20.0 s 则能看见白手包住并承接盒子，此处仅用于辨认可见线索。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 192,
+      "text_zh": "推断：本局末段脚本（log step 192）仍在对右手做接近，未进入交接控制；结合盒子始终留在原柱的可见片段与白手空手的后续画面，假肢臂累计100步接触节点未达。没有把被重置的早期试验算入本局。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 38.4,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0、1.4、22.0、25.0–30.4、33.2、35.6、36.8、38.4 s 盒子保持在同一起始柱上，未见被搬往另一根目标柱；视频 39.0–40.0 s 两手仍空着，没有放置或松手动作。"
+     },
+     {
+      "t": 40.0,
+      "log_step": 197,
+      "text_zh": "推断：结合全片未出现搬运、可见片段中盒子仍在起始柱，以及 log step 197 服务在本局1000个控制步给出failure，物体放上目标柱节点未达；该判断不是仅从success=0倒推。末帧盒子在画外，不据此声称直接观察了末帧柱面关系。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_held_by_myo",
+   "t": 33.2,
+   "log_step": 192,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 30.4–38.4 s 机器人肌肉右手尝试接近起始柱上的盒子，却在其下方或侧方摆动；视频 33.2 s 盒子仍在柱上，未被夹住，后续也未修复为持物。log step 192 正在给右手求逆解并分块提交接近命令；log step 197 本局1000步后服务结束并判失败。断在右手持起这一步，后两步尚未开始。"
+  },
+  "unmet": {
+   "object_held_by_myo": false,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 33.2,
+    "log_step": 192,
+    "evidence_zh": "观察：视频 30.4–38.4 s 机器人右手向盒子附近摆动却没有包夹、承载盒子；视频 33.2、35.6、38.4 s 盒子仍在起始柱，视频 39.0–40.0 s 右手仍空着。所需的持物接触未建立，未见先持有再滑落的过程。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 188,
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：右臂的命令执行存在明显偏差，使手未能到达并保持在物体接触位姿。log step 188 的脚本要求保持当前关节姿态，提交120步后打印的关节绝对误差中包含0.979和1.527 rad；log step 192 明确以盒子附近指尖位置为目标求逆解，结果指尖仍偏离目标，视频 30.4–38.4 s 亦见空手反复摆动。证据支持控制层问题，但未验证具体接触方案或哪组参数能成功，故不细分为参数调试、反馈控制或抓取设计。"
+   }
+  ],
+  "gap_zh": "没有状态日志，也没有本局物体接触标签、两臂历史接触累计数、goal_dist、goal_touch和末尾独占接触计数；native_metrics均未报告。随手移动的相机在部分时段及视频 39.0–40.0 s 将盒子遮挡或移出画面，最终未达依据此前反复可见的起始柱关系、空手动作和服务终局结果综合判断。微小擦碰不能精确排除，但无达标的持续接触或搬运迹象；控制失配的具体机制未被查明，log step 182 的后处理索引错误发生在动作提交后且本局继续运行，不能把它当作外部中断。"
+ },
+ "summary_zh": "第11局为受评局：物体始终未被从起始柱带走，右手持起、交给左手和放上目标柱三个节点均未达。视频 30.4–38.4 s 的右手接近未建立持物接触，日志中的姿态保持误差支持控制层问题，最终本局1000个控制步后判失败。缺少接触历史和状态日志，短暂擦碰及控制失配的具体机制不能精确确定。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "三个节点均未达、断点在右手持起、现象没建立接触、根因停在控制，与视频及服务终局读数一致；未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:43:12Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--x4egqsj/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--x4egqsj/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c2da55792d751ffaab6a2916a1b04f456140ef5eec07876a12f550f7673955f8",
+   "video_sha256": "12b02f935fc7bc59b6cee67f5d363b9657fdd9ee80280e5d0a2f57c756e4972b",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/bimanual/trajectories/myosuite-bimanual-i00-standard--ydvnvvj.json

@@ -0,0 +1,244 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-bimanual-i00-standard--ydvnvvj",
+ "taskKey": "myosuite/bimanual",
+ "dag_sha256": "0af83fb75572a8dee8e6c0b8abc848886eb871d14f1068bf9e0ea0f234e3bb96",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 1,
+   "sim_steps": 1,
+   "physics_steps": 5,
+   "steps": 1,
+   "calls": 2,
+   "episodes": 1,
+   "resets": 0,
+   "reset_budget": 10,
+   "sim_steps_total": 1,
+   "elapsed_s": 3640.0,
+   "video_rendered": 1,
+   "task_failure": 0,
+   "timeout": 1,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.366,
+   "peak_processes": 7,
+   "peak_disk_gb": 0.474,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-bimanual-i00-standard--ydvnvvj/replay.mp4",
+  "duration_s": 0.08,
+  "fps": 25.0,
+  "frames": 2,
+  "sha256": "5114182e3c768f781d467ac2efc0f338c13a41aba77051463293636e57482924"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 3,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "object_held_by_myo": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：已查看本次视频全部两帧（视频 0.00 s、0.04 s）；右侧视图中红边长方体仍立在起始圆柱上，带红色肌肉束的手张开停在其上方，手指与物体之间可见间隙，没有靠近抓取或持起动作。"
+     },
+     {
+      "t": null,
+      "log_step": 57,
+      "text_zh": "观察：log step 57 中 agent 通过实时机器人服务 r.submit 仅提交一行80维全零执行器指令，dt=0.01；服务回报 control_steps=1。run.json 的评分字段也记录 n_actions=1、sim_steps=1、resets=0。"
+     },
+     {
+      "t": null,
+      "log_step": 57,
+      "text_zh": "推断：受评局总共只有一个控制步，不可能达到本节点要求的肌肉臂累计接触至少100个控制步；此结论同时与视频中未抓取的画面一致，并非仅根据失败评分反推。"
+     }
+    ]
+   },
+   "object_handed_to_prosthesis": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.04,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s、0.04 s 的两个视图均没有两手在长方体处会合、白色假肢手承接或物体随假肢手移动；物体始终留在起始圆柱上。"
+     },
+     {
+      "t": null,
+      "log_step": 57,
+      "text_zh": "推断：log step 57 的实时服务仅执行一个控制步，且评分字段 sim_steps=1、resets=0，因此受评局也不可能达到假肢臂累计接触至少100个控制步。log step 159 的承接记录来自 agent 另建的本地仿真测试，不能算作受评交接。"
+     }
+    ]
+   },
+   "object_on_goal_pillar": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.04,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s、0.04 s 中红边长方体一直在起始圆柱上，没有离开原位或被送至另一个目标柱，两个视图均未出现放置动作。"
+     },
+     {
+      "t": null,
+      "log_step": 159,
+      "text_zh": "观察：log step 159 的 solved=True、goal_dist=0.0696 与 goal_touch=11 是本地测试脚本的输出；该脚本在 log step 156 中以 gym.make 新建环境并调用 env.step，未调用实时服务 r.submit。这些数值不属于本次受评局，故不用于节点达成或原生指标。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "goal_dist",
+   "unit": "m",
+   "threshold": 0.17,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "myo_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "prosthesis_contact_count",
+   "unit": "control_steps",
+   "threshold": 100,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_touch",
+   "unit": "control_steps",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "goal_only_tail_count",
+   "unit": "control_steps",
+   "threshold": 8,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "object_held_by_myo",
+   "t": 0.0,
+   "log_step": 57,
+   "where": "首步前",
+   "evidence_zh": "观察：本次视频全部两帧（视频 0.00 s、0.04 s）中物体仍在起始柱上，肌肉手未抓取物体。log step 57 是唯一实时动作提交，只执行一个全零指令控制步；评分记录 n_actions=1、sim_steps=1，受评局未启动右手持起物体这一步。"
+  },
+  "unmet": {
+   "object_held_by_myo": false,
+   "object_handed_to_prosthesis": false,
+   "object_on_goal_pillar": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.0,
+    "log_step": 57,
+    "evidence_zh": "观察：视频 0.00 s、0.04 s 中肌肉手保持张开并与长方体分离，长方体仍在起始柱上；log step 57 仅由实时服务执行一行全零诊断指令。后续 log step 156–160 的动作在 agent 本地 gym 环境测试和修改，受评录像没有抓取尝试。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "观察：inputs/ 没有状态日志；本次录像只有视频 0.00 s、0.04 s 两帧，评分字段未报告各项原生指标的实测值，故这些指标均留空。log step 159 虽显示本地测试通过接触历史检查，log step 160 仍仅修改该本地脚本；完整记录没有说明为何未将后续控制序列用于实时服务，也没有最终完成声明。推断：可以确定受评抓取未启动，但不足以把原因细分为规划、控制或交付操作错误；不能把超时或缺少后续提交本身当作根因，因此根因层留空。"
+ },
+ "summary_zh": "受评录像中物体始终留在起始柱上，右手持起、交给左手、放上目标柱三个节点均未达成（0/3）；断点为右手持起物体的首步前，现象为没尝试。log step 57 仅向实时服务提交一个全零控制步，而 log step 159 的成功接触与放置结果来自独立本地仿真，未出现在受评录像中。缺少未执行后续实时动作的决策证据与状态日志，根因留空，原生指标不采用本地测试值。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "标注与证据一致：受评局仅 1 个控制步（log step 57、run.json n_actions=1），视频两帧物体仍在起始柱上，三个节点均未达成；断点首步前 + 没尝试，根因层留空并写 gap_zh，符合 FAIL-22 与 D23；log 159 的本地测试结果未被当作受评事实。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:42:53Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--ydvnvvj/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-bimanual-i00-standard--ydvnvvj/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b915383042d8ea3799f7d581e92e19aabe2e66e5ce002ed7ab37d02cf6768976",
+   "video_sha256": "5114182e3c768f781d467ac2efc0f338c13a41aba77051463293636e57482924",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-privileged--346fbz4.json

@@ -0,0 +1,183 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-privileged--346fbz4",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 172,
+   "sim_steps": 172,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 103,
+   "agent_resets": 12396,
+   "agent_sim_steps": 11699,
+   "agent_records": 3,
+   "agent_replays": 9,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.274,
+   "peak_processes": 16,
+   "peak_disk_gb": 0.241,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-privileged--346fbz4/replay.mp4",
+  "duration_s": 6.92,
+  "fps": 25.0,
+  "frames": 173,
+  "sha256": "d13fe45044a18ee63b0e666a50055dab5d7fbbe19f8a5f0d92facb04caaa4cfc"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 白色 MyoLegs 模型位于场地中部，浅蓝圆顶黑色柱状对手位于画面右下，两者明显分离；视频 0.0–5.6 s 模型摆腿并向对手方向移动，但尚未靠近对手。"
+     },
+     {
+      "t": 6.88,
+      "log_step": 129,
+      "text_zh": "观察：视频 5.6–6.88 s 的密集帧显示模型身体下沉并向前倒伏，末帧骨盆与对手仍有明显水平间隔，对手圆顶仍为浅蓝色；成功参考视频 9.64 s 则显示模型骨盆紧邻对手、圆顶变绿。log step 129 对最终交付文件的两次重放均报告 success=false、172 步、相同 state_sha，与评分器 replay_success=0、deterministic=1 一致。"
+     },
+     {
+      "t": 6.88,
+      "log_step": 134,
+      "text_zh": "推断：结合全段视频未出现追及位置关系及原生重放失败结果，追及对手节点始终未达成。log step 134 再次验证交付文件为 172 步且失败；log step 132 的另一组参数测试及 log step 139 对更近样本的自述不代表本次交付视频。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–5.6 s，MyoLegs 双腿摆动、脚部先后落在木质地面承重，推动本体向对手移动；视频约 6.0 s 后支撑失稳、身体下沉。未见本体与柱状对手实际触碰，任务验收也不要求触碰。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 6.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 6.0 s，正在向固定对手行进的 MyoLegs 身体开始明显前倾下沉，视频 6.4–6.88 s 持续倒伏，未恢复支撑；回合结束时仍与对手分离。log step 129 对已交付的 172 步文件确认两次重放均失败，评分器 task_failure=1。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 6.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 5.6–6.88 s，MyoLegs 从尚能以腿支撑的姿态转为躯干前倾、骨盆明显降低及身体倒伏，直到末帧未恢复；模型此时仍在场内且尚未追及对手。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent 的步态控制器未能在向对手推进时维持身体支撑。log step 127 用 best4 参数录制并保存 172 步动作，原生 replay 即报告失败；log step 134 与 log step 129 对该交付重复验证仍失败，视频约 6.0–6.88 s 显示其失稳倒伏。agent 在 log step 139 自述支撑腿近似直杆、未能制动前向运动，这是尚未独立验证的具体机制；因此根因仅记控制，不细分，也不将已修复的早期求解器问题作为本次根因。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志；俯视视频分辨率有限，不能精确给出骨盆高度越过 0.5 米阈值的时刻、逐足接触力或控制失稳的细分类机制。评分器 reward 未报告 distance，故原生距离指标留 null；log step 124 的距离为会话测试输出，不作为评分器指标。视频时钟独立标注，未把 agent 自述的仿真持续时间映射到视频。"
+ },
+ "summary_zh": "本次交付的 MyoLegs 已摆腿向固定对手移动，但始终未达到追及距离，最终完成 0/1 个节点。视频约 6.0 s 起模型失去支撑并倒伏，断点为追及对手的操作中，现象为全身·平衡；日志对同一交付文件的确定性失败重放支持控制未能维持支撑的判断。缺少逐帧状态和接触力，不能确定更细的控制机制，评分器也未报告原生距离数值。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点未达成、断点（追及对手·操作中，约6.0 s开始倒伏）、现象全身·平衡、根因停在控制均有视频与重放日志支持；视频全程模型与对手明显分离，未见错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:03:29Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--346fbz4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--346fbz4/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "43d715927c77ecb185bce10520f9165eac3e8d9facc677bde44d387d8ff95be5",
+   "video_sha256": "d13fe45044a18ee63b0e666a50055dab5d7fbbe19f8a5f0d92facb04caaa4cfc",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-privileged--bgpmfjx.json

@@ -0,0 +1,159 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-privileged--bgpmfjx",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 242,
+   "sim_steps": 241,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 161,
+   "agent_resets": 164,
+   "agent_sim_steps": 15557,
+   "agent_records": 0,
+   "agent_replays": 38,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 2.282,
+   "peak_processes": 20,
+   "peak_disk_gb": 0.983,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-privileged--bgpmfjx/replay.mp4",
+  "duration_s": 9.68,
+  "fps": 25.0,
+  "frames": 242,
+  "sha256": "4bb26d8f44f912cf4fab08f2272dcaea35e8d29f64707e3f5bed21275d86eb13"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [
+     [
+      9.64,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s，白色MyoLegs模型位于场地中央偏远处，带浅蓝圆顶的黑色柱状对手位于其右前方，两者明显分离；视频 0.00–8.08 s，模型用双腿迈步并逐渐靠近固定对手。"
+     },
+     {
+      "t": 9.64,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 9.20–9.64 s，白色模型继续靠近对手，视频 9.60 s 对手顶部仍为浅蓝色，视频 9.64 s 末帧变为绿色，模型已到其近旁。该时间取自视频帧时间戳，未将 agent 工作时间或模拟时间换算为视频时间。"
+     },
+     {
+      "t": 9.64,
+      "log_step": 289,
+      "text_zh": "推断：视频 9.64 s 的变绿与最终成功判定一致，追及对手节点在此达成。inputs/run.json 的评分器字段 success=1、replay_success=1、deterministic=1、instance_ok=1、task_failure=0；log step 250 安装的242行交付文件，在 log step 289 的最终官方重放中两次均于控制步241返回 success=true 且最终状态哈希相同。更晚另存的256行测试候选不属于该交付文件；未用会话测试中的跌倒或距离替代本段视频。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–8.08 s，MyoLegs的双脚交替落在木质场地上支承身体，双腿迈步使骨盆向固定对手移动；实际可见的支承交互为脚与地面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "白色MyoLegs模型从场地中央迈步靠近固定对手，视频 9.64 s 末帧对手顶部变绿；结合评分器成功字段及 log step 289 的最终交付重放，追及对手节点完成，终态为1/1，无失败断点。没有逐帧状态日志，评分器未报告原生distance数值，故不从低分辨率俯视画面估算精确距离，达成时刻以视频变色帧为锚。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "成功回合，追及对手节点在视频 9.64 s 末帧（对手顶部由浅蓝变绿）达成，与评分器 success=1 及 log step 250/289 的官方重放（控制步241）一致；contact_actual 为脚支撑，无失败断点，未发现需修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:01:42Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--bgpmfjx/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--bgpmfjx/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9dcb53e86afd7b0e35180d60c2ae1965294f47b4c2c13c4c6fca06fce83b9c0c",
+   "video_sha256": "4bb26d8f44f912cf4fab08f2272dcaea35e8d29f64707e3f5bed21275d86eb13",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-privileged--dc2fmax.json

@@ -0,0 +1,162 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-privileged--dc2fmax",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 260,
+   "sim_steps": 260,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 46,
+   "agent_resets": 8362,
+   "agent_sim_steps": 72813,
+   "agent_records": 5,
+   "agent_replays": 10,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 1,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.0,
+   "peak_processes": 0.0,
+   "peak_disk_gb": 0.0,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-privileged--dc2fmax/replay.mp4",
+  "duration_s": 10.44,
+  "fps": 25.0,
+  "frames": 261,
+  "sha256": "701c0673438de02a08f9998134753b36166537c2bc647093662e4dfb5a3fea32"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.4,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s，白色 MyoLegs 模型位于场地中部，带浅蓝圆顶的黑色柱状对手在其右前方，两者明显分离；视频 0.00–9.20 s，模型通过双腿迈步逐渐向固定对手接近。"
+     },
+     {
+      "t": 10.4,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 9.20–10.40 s，模型继续靠近对手；视频 10.36 s 圆顶仍为浅蓝色，视频 10.40 s 末帧圆顶变为绿色。成功参考视频的末帧（参考视频 9.64 s）也出现同样的绿色信号；这里只用参考视频解释画面信号，本次完成时刻取自本次视频。"
+     },
+     {
+      "t": 10.4,
+      "log_step": 146,
+      "text_zh": "观察：log step 143 中，最终交付动作的两次新进程重放均报告 success=true、steps=260、state_sha=87fb24c9797074df；log step 146 对该最终文件的渲染重放报告 solved=1、reason=success，哈希相同。inputs/run.json 的评分器字段也为 success=1、replay_success=1、deterministic=1、instance_ok=1，与视频 10.40 s 的绿色信号一致。"
+     },
+     {
+      "t": 10.4,
+      "log_step": null,
+      "text_zh": "推断：结合视频 10.40 s 的达标信号和评分器成功结果，“追及对手”在视频 10.40 s 达到骨盆水平距离不大于 0.5 m 的原生成功条件，并保持至录像结束；本次没有逐帧状态日志，不从低分辨率画面测量精确米数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00、2.00、4.00、6.00、8.00 s 及其间帧中，MyoLegs 的双腿屈伸迈步，脚先后在木质地面落地承重，使身体向对手移动；实际作用端为双脚，接触对象为地面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "MyoLegs 通过双腿迈步接近固定对手，在视频 10.40 s 末帧触发圆顶变绿的成功信号，完成“追及对手”（最终 1/1）；评分器确认成功且最终交付可确定性重放。没有逐帧状态日志，精确距离无法由画面测量，且评分器 reward 未报告 distance 数值，因此该原生指标留空；log step 144 的仿真 2.60 s 为 agent 脚本诊断时间，不作为视频时间。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点“追及对手”在视频 10.40 s 末帧圆顶变绿，与评分器成功字段和 log step 143/146 的重放结果一致；时间、接触（腿/躯干·支撑）和 failure=null 均成立。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:05:08Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--dc2fmax/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--dc2fmax/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "acc1a0406e214b3f2dde0752e2142c183a24d39d657f36533fce0b9613708190",
+   "video_sha256": "701c0673438de02a08f9998134753b36166537c2bc647093662e4dfb5a3fea32",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-privileged--ddgtnpd.json

@@ -0,0 +1,183 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-privileged--ddgtnpd",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 61,
+   "sim_steps": 61,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 29,
+   "agent_resets": 313,
+   "agent_sim_steps": 17046,
+   "agent_records": 6,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.659,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.239,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-privileged--ddgtnpd/replay.mp4",
+  "duration_s": 2.48,
+  "fps": 25.0,
+  "frames": 62,
+  "sha256": "5c8ab98b7c02335f67c3e96aa1368a17ecc4b42930d155e1a5b19e457527d1df"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s中，白色MyoLegs模型位于场地中后部，带浅蓝圆顶的黑色柱状对手位于画面右前方，两者明显分离。"
+     },
+     {
+      "t": 2.44,
+      "log_step": null,
+      "text_zh": "观察：检查视频0.00–2.44 s全程及视频1.28–2.44 s逐帧末段，MyoLegs仅向前移动少许便逐渐倾倒，直到末帧骨盆与对手仍明显分离，从未出现追及所需的近距离位置关系。"
+     },
+     {
+      "t": null,
+      "log_step": 42,
+      "text_zh": "观察：log step 42中，agent脚本重新生成并保存61×80动作的交付文件；基准CLI对该文件两次重放均返回success=false、steps=61，末态哈希相同且instance_ok=true、deterministic=true。评分器reward同时报告replay_success=0、task_failure=1。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–1.20 s中，MyoLegs以脚接触木质场地地面承重，双腿迈动时身体向前移动；随后身体倾倒。视频全程未见模型接触黑色柱状对手。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 1.6,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约1.60 s中，正在前移的MyoLegs身体明显倾斜、躯干与腿的投影逐渐横倒；视频1.60–2.44 s中未恢复直立，仍远离右前方对手。推断：追及过程中发生未修复的失衡，断点为“追及对手”。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频1.28–2.44 s连续帧中，MyoLegs从直立外形逐渐侧倾、身体下沉，末段明显失去直立支撑姿态，始终没有恢复，也未到达对手。log step 42对交付文件的两次重放均以原生failure结束。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：agent的肌肉控制器未能维持前移中的支撑和平衡。log step 42的交付脚本先跟踪修改后的初始关节姿态，再切换步态参考，并将生成的61步动作保存；同一文件两次重放均失败且末态一致，视频约1.60–2.44 s直接显示倾倒。log step 36、41的会话测试在多组参考相位和增益下也出现骨盆下沉并失败，支持控制能力不足这一归因，但不足以证明某个参数或接触几何是唯一原因。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志，俯视画面中的模型较小，无法精确定位骨盆跨过0.5 m高度阈值的瞬间，或区分失衡究竟由参考轨迹、肌肉力分配还是反馈调节引起，故根因仅标“控制”。评分器reward未报告distance数值，native_metrics中留null；会话测试的best_d不充当评分器指标。log step 51自述的仿真约0.61 s与视频2.48 s属于不同时间尺度，本文所有视频锚点均取自所检视帧。"
+ },
+ "summary_zh": "评分轨迹只��移了一小段便倾倒，始终未追及固定对手，最终完成0/1节点。视频约1.60–2.44 s显示未恢复的失衡，交付文件两次确定性重放均失败，根因支持标为控制不足。缺少逐帧状态日志和评分器距离值，无法进一步确定失衡的具体控制机制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点（追及对手·倾倒）、平衡现象与控制根因均与视频一致；视频时间均在片内，未发现需改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:59:04Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--ddgtnpd/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--ddgtnpd/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "986cae30048c93566d2b4d9b89813b56c15fb66051008fd5b3954948d9d0ba33",
+   "video_sha256": "5c8ab98b7c02335f67c3e96aa1368a17ecc4b42930d155e1a5b19e457527d1df",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-privileged--stwgf9z.json

@@ -0,0 +1,164 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-privileged--stwgf9z",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 450,
+   "sim_steps": 450,
+   "task_failure": 0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 36,
+   "agent_resets": 3775,
+   "agent_sim_steps": 4354,
+   "agent_records": 8,
+   "agent_replays": 3,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.334,
+   "peak_processes": 17,
+   "peak_disk_gb": 0.505,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-privileged--stwgf9z/replay.mp4",
+  "duration_s": 18.04,
+  "fps": 25.0,
+  "frames": 451,
+  "sha256": "98bfd85d121a8ab07760ef5f6647b16612ce13213db4d60d2506f563b3b295dc"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [
+     [
+      18.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s，白色MyoLegs肌肉骨骼模型位于场地中部，带浅蓝圆顶的黑色柱状对手位于其右前方，两者明显分离；视频 0–17 s 的逐秒画面显示模型迈步并逐渐转向、接近固定对手。"
+     },
+     {
+      "t": 17.96,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 17.00–18.00 s，模型继续迈步靠近对手；视频 17.96 s 对手圆顶仍为浅蓝色，视频 18.00 s 末帧圆顶变绿，模型骨盆已到对手附近，身体未倒伏。成功参照视频 9.64 s 末帧也出现圆顶变绿，仅用作成功提示的外观对照。"
+     },
+     {
+      "t": 18.0,
+      "log_step": 38,
+      "text_zh": "观察：log step 37 中agent脚本生成并保存最终450步轨迹，输出成功、终点水平距离0.49948371705143674米、骨盆高度0.8652807316460549米及仿真 4.50 s；log step 38 中基准重放工具对该交付文件两次重放均报告success=true、steps=450且终态哈希相同。评分器字段success=1、replay_success=1、deterministic=1，与视频 18.00 s 的绿色成功提示一致。"
+     },
+     {
+      "t": 18.0,
+      "log_step": 38,
+      "text_zh": "推断：结合视频 18.00 s 的成功提示、最终交付文件的重放验收和评分器成功字段，追及条件在视频末帧成立，区间记为[18.00 s, 视频结束]；此时刻直接取自视频帧，不将agent用时或日志中的仿真 4.50 s 当作视频时间。评分器reward未单列distance数值，因此native_metrics的distance保留null。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0–18.00 s，MyoLegs双腿交替迈步，足部落在木质场地地面上支撑身体并推动骨盆接近对手；实际支承交互为脚与地面。视频未显示可确认的身体撞击对手，追及验收也不要求实际触碰对手。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "MyoLegs通过双脚支撑、迈步和转向接近固定对手，在视频 18.00 s 末帧触发圆顶变绿的成功提示，追及对手节点完成，最终为1/1。log step 38 对最终450步交付轨迹的两次重放均成功，与评分器成功结果及视频相符，无失败断点。未提供逐帧状态日志，无法独立量测精确水平距离；评分器未单列distance数值，故原生距离指标保留null，节点达成时刻取自视频成功提示。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "opponent_tagged 在视频 18.00 s 圆顶由浅蓝变绿，与评分器 success=1 及重放验收一致；接触标签（腿/躯干·支撑）合理，无失败断点。无需修改。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:01:50Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--stwgf9z/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--stwgf9z/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b62fa901dea06474eeda780ebb89ffa68354feac2ccd32522bd153fa2058eefa",
+   "video_sha256": "98bfd85d121a8ab07760ef5f6647b16612ce13213db4d60d2506f563b3b295dc",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-privileged--zpujabs.json

@@ -0,0 +1,184 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-privileged--zpujabs",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 88,
+   "sim_steps": 88,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 331,
+   "agent_resets": 6213,
+   "agent_sim_steps": 25214,
+   "agent_records": 1,
+   "agent_replays": 18,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 4.942,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.269,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-privileged--zpujabs/replay.mp4",
+  "duration_s": 3.56,
+  "fps": 25.0,
+  "frames": 89,
+  "sha256": "0fd5f5a1be09961389a59a0c827a05e53a61fb2be5734698daf59620c69fb555"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0–3.52 s的俯视画面中，白色MyoLegs模型从场地中部向前移动，带浅蓝圆顶的黑色柱状对手始终在其右前方，两者地面投影一直明显分离；视频3.52 s结束时模型已后仰下沉，仍未靠近对手。"
+     },
+     {
+      "t": 3.52,
+      "log_step": 438,
+      "text_zh": "观察：log step 438的诊断脚本从初态逐行重放最终88步交付文件，读取pelvis世界位置和对手位置；控制步索引87处骨盆高度0.486 m，水平距离3.620 m。log step 441随后对同一交付文件两次重放均success=false、steps=88、末态哈希相同，与视频末尾下沉且远离对手一致。"
+     },
+     {
+      "t": 3.52,
+      "log_step": null,
+      "text_zh": "推断：本次全程未进入0.5 m追及阈值。作为视图参照，成功oracle视频9.64 s中白色模型骨盆已紧邻黑色对手，圆顶变绿；本次视频3.52 s无此接近关系，且评分记录success=0、task_failure=1。oracle仅用于理解成功外观，不作为本次进度证据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.0–2.8 s，MyoLegs双脚交替落在木质场地上支承身体并迈步移动；随后身体后仰下沉。全程未见模型与柱状对手实际接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 2.8,
+   "log_step": 438,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约2.8–3.52 s，正在迈步接近对手的MyoLegs模型逐渐后仰、骨盆下沉，未恢复直立，结束时与对手仍明显相隔。log step 438重放最终交付文件的末控制步索引87处骨盆高度为0.486 m；log step 441两次确定性重放均在88步失败。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 2.8,
+    "log_step": 438,
+    "evidence_zh": "观察：视频2.8–3.52 s逐帧画面中，MyoLegs身体逐渐后仰并降低，末帧不再保持正常直立支撑姿态，追及尚未完成；log step 438对最终交付轨迹的重放读取末控制步骨盆高度0.486 m，低于DAG所述平地跌倒阈值0.5 m。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 439,
+    "evidence_zh": "推断：log step 439所示交付生成脚本用关节位置/速度反馈、躯干俯仰与侧倾力矩及有界肌肉分配驱动行走，log step 440执行该脚本保存88步，log step 441两次重放一致失败；结合视频2.8–3.52 s的后仰下沉及log step 438的末步骨盆高度0.486 m，表明该肌肉行走控制未能维持支撑和平衡。证据不足以区分目标步态、力矩分配或反馈增益中的具体缺陷。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志，俯视低分辨率画面无法精确恢复足底接触力或失衡起始时刻；log step 438提供的是最终交付文件的稀疏重放诊断，可佐证末态，但不足以定位具体控制子类。评分器reward未报告distance数值，因此native_metrics.distance留null，诊断脚本读取的3.620 m仅作辅助证据。",
+  "unmet": {
+   "opponent_tagged": false
+  }
+ },
+ "summary_zh": "MyoLegs模型尝试迈步接近固定对手，但视频0.0–3.52 s始终未达到追及距离，唯一节点“追及对手”最终未完成（0/1）。视频约2.8 s起后仰下沉，最终交付文件的末步诊断骨盆高度为0.486 m，两次确定性重放均失败，支持平衡丧失及控制层归因。缺少逐帧接触力与控制误差记录，无法确定更具体的控制缺陷。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "唯一节点 opponent_tagged 未达成、final=false、breakpoint 与平衡/控制归因均与视频、log step 438/441 一致（末步骨盆高度 0.486 m<0.5 m、距离 3.62 m，两次确定性重放失败）；未发现需要修改的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:06:46Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--zpujabs/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-privileged--zpujabs/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b562f9642912a3990e09b6f3fac039676f0b7fb5ac5eafc4b6ceaf4f933156b3",
+   "video_sha256": "0fd5f5a1be09961389a59a0c827a05e53a61fb2be5734698daf59620c69fb555",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-standard--4n8f8yi.json

@@ -0,0 +1,184 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-standard--4n8f8yi",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 0,
+   "sim_steps": 0,
+   "physics_steps": 0,
+   "steps": 0,
+   "calls": 2,
+   "episodes": 1,
+   "resets": 0,
+   "reset_budget": 10,
+   "sim_steps_total": 0,
+   "elapsed_s": 2454.9,
+   "video_rendered": 1,
+   "task_failure": 0,
+   "timeout": 1,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.467,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.371,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-standard--4n8f8yi/replay.mp4",
+  "duration_s": 0.04,
+  "fps": 25.0,
+  "frames": 1,
+  "sha256": "a894b7592370159d3a1ba5b7e128c6a55a1080d916a5214612192d2dc7ebf615"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：评分视频仅含视频 0.00 s 的一帧（总长0.04秒）；白色MyoLegs模型位于场地中央附近，带浅蓝圆顶的黑色柱状对手在其右下方，两者地面投影明显分离，未达到追及距离。"
+     },
+     {
+      "t": 0.0,
+      "log_step": 64,
+      "text_zh": "观察：评分器记录 n_actions=0、sim_steps=0、episodes=1、resets=0、live_success=0；agent 在 log step 64 明确自述未在服务中运行控制器。推断：评分回合始终停在初态，追及步骤没有启动；日志中的本地测试不构成该回合进度。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 0.0,
+   "log_step": 64,
+   "where": "首步前",
+   "evidence_zh": "观察：视频 0.00 s 的唯一画面中，MyoLegs与对手仍远离；评分器记录零动作、零仿真步。agent 在 log step 64 自述因本地控制器测试持续跌倒，决定不在服务中启动控制器。推断：评分回合的追及步骤从未开始，断点位于首步前。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.0,
+    "log_step": 64,
+    "evidence_zh": "观察：视频 0.00 s 为初始分离画面，完整评分视频仅这一帧；评分器 n_actions=0、sim_steps=0、resets=0。agent 在 log step 64 自述没有向现场服务运行控制器。推断：MyoLegs未开始追及对手；本地测试的跌倒不是评分回合中的现象。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 60,
+    "source_zh": "日志",
+    "evidence_zh": "观察：agent 的本地仿真测试在 log step 19、30、60 分别于本地仿真 0.61 s、1.32 s、0.47 s 输出 fallen=True、win=False；log step 64 中 agent 自述控制器无法稳定落地，因此没有在服务中启动。推断：控制器尚未形成可用的稳定行走控制，促使agent放弃现场执行；依据是会话测试及其与停止的明确关联，不是从零进度反推。"
+   }
+  ],
+  "gap_zh": "观察：inputs/包含评分视频和完整会话日志，但没有逐帧状态日志；评分视频只有初始一帧，评分器未报告原生distance数值。推断：本地测试支持控制能力不足的宽泛归因，但不能验证该控制器在评分回合中的落地表现，也不能确定肌肉分配、反馈或具体参数中的哪一项是最终原因，故不细分控制子类。"
+ },
+ "summary_zh": "评分回合停留在初态，追及对手未启动，最终完成0/1节点；唯一视频帧中两者明显分离，评分器记录零动作、零仿真步。断点为追及对手的首步前，现象为没尝试；本地测试持续跌倒及agent最终说明支持控制能力不足促使其未启动现场执行的归因。没有逐帧状态日志或现场运动，具体控制机制无法核实，原生距离指标也未报告。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "评分回合0动作、0仿真步，视频仅一帧初态；节点未达、断点首步前/没尝试、根因控制（来自会话本地测试，按FAIL-22）均有证据支持，时钟与推断标注无误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:12:15Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--4n8f8yi/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--4n8f8yi/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9cee0f3839504a5e79506e1f3682971a8ffd01afd5701f5ff7b7353400df55b7",
+   "video_sha256": "a894b7592370159d3a1ba5b7e128c6a55a1080d916a5214612192d2dc7ebf615",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-standard--c6usgpu.json

@@ -0,0 +1,195 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-standard--c6usgpu",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 99,
+   "sim_steps": 99,
+   "physics_steps": 990,
+   "steps": 294,
+   "calls": 15,
+   "episodes": 5,
+   "resets": 4,
+   "reset_budget": 10,
+   "sim_steps_total": 294,
+   "elapsed_s": 2863.8,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.625,
+   "peak_processes": 15,
+   "peak_disk_gb": 0.582,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-standard--c6usgpu/replay.mp4",
+  "duration_s": 4.0,
+  "fps": 25.0,
+  "frames": 100,
+  "sha256": "7748ba0573867e662a34db9f1fd5867948019be2058c9b4efa0f48431ead38b2"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s，白色MyoLegs位于场地中部，浅蓝圆顶黑柱对手在其右前方，两者地面投影明显分离，追及条件起初不成立。"
+     },
+     {
+      "t": 2.0,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–3.96 s，模型双腿迈动并向画面下方移动，但骨盆始终与右下方对手有明显间隔；视频 2.00 s 尚未走到对手旁，视频 3.96 s 身体已大幅倾倒，仍未接近到成功参考末段所示的贴近位置。"
+     },
+     {
+      "t": 3.96,
+      "log_step": 152,
+      "text_zh": "观察：log step 151 的最终部署脚本重置后执行肌肉序列，log step 152 的服务状态为 episode=5、control_steps=99、done=true，返回未追及/出界/跌倒的合并失败原因，末次骨盆root_z=0.48 m；评分字段 graded_by_live=1、live_success=0、n_actions=99 与该最后回合相符。视频 3.96 s 模型仍远离对手，故本节点从未达成；此前本地测试的运动不计入进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.40–2.00 s，MyoLegs双腿迈动，脚在木质地面落下并支承身体；随后视频约 2.80–3.96 s 身体倾倒、支承失稳。未见模型接触对手。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 2.8,
+   "log_step": 152,
+   "where": "操作中",
+   "evidence_zh": "观察：模型在追及移动中，于视频约 2.80 s 已出现明显倾倒，视频 3.00–3.96 s 倾斜继续扩大、身体下沉，始终未靠近右下方对手。log step 152 的最后现场回合返回失败，末次骨盆root_z=0.48 m，之后未恢复或重新执行；断点定位为正在尝试的追及对手节点。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 2.8,
+    "log_step": 152,
+    "evidence_zh": "观察：视频约 2.80–3.96 s，MyoLegs躯干由倾斜转为近水平，骨盆下沉，未能保持站立支承；log step 152 记录最后回合末次root_z=0.48 m，低于任务文本规定的跌倒阈值0.5 m，并返回done=true及失败原因。该倾倒发生在尚未追及对手时。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 145 的seq函数仅按时间和固定参数分配各肌肉激活；log step 151 的最终脚本在执行前生成全部指令，执行中读取的本体状态仅用于打印和判断回合结束，没有依据骨盆高度或身体倾斜调整后续指令。结合视频约 2.80–3.96 s 持续倾倒及log step 152 的低骨盆高度，所交付的肌肉控制序列未能维持行走支承，导致追及中断；证据不足以再区分具体参数或控制子类。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或足底接触力记录，低分辨率俯视视频不能给出精确骨盆距离、接触部位及倾倒起点；断点时间取视频约 2.80 s 的可见明显倾倒处。评分字段未报告distance，因此原生距离指标置null；log step 149 的3.17 m来自本地测试，不是本次评分回合的距离。最终服务回合与本地测试为何分歧无法确定，不归因于仿真或环境；服务控制步的仿真时钟不直接当作视频时钟。"
+ },
+ "summary_zh": "评分采用最后的第5回合：MyoLegs实际迈步并以脚支承地面，但全程未追及固定对手，节点最终为0/1。视频约 2.80 s 开始明显倾倒，至视频 3.96 s 身体下沉，log step 152 的末次骨盆高度为0.48 m；最终按预生成肌肉序列执行的控制未能维持支承。无逐帧状态和接触力记录，无法细分控制失稳机制或给出精确末端距离。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点（追及中跌倒，平衡）与根因（开环肌肉序列，停在控制）均与视频帧和log step 152的服务读数一致；未发现需要修改之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:01:27Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--c6usgpu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--c6usgpu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0a2d1b131c25c2903c73e181c0b22ba9519edfdfc79e0a46a32fa04d1ad0c30a",
+   "video_sha256": "7748ba0573867e662a34db9f1fd5867948019be2058c9b4efa0f48431ead38b2",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-standard--kscoqvt.json

@@ -0,0 +1,179 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-standard--kscoqvt",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 0,
+   "sim_steps": 0,
+   "physics_steps": 0,
+   "steps": 0,
+   "calls": 1,
+   "episodes": 1,
+   "resets": 0,
+   "reset_budget": 10,
+   "sim_steps_total": 0,
+   "elapsed_s": 380.2,
+   "video_rendered": 1,
+   "task_failure": 0,
+   "timeout": 1,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.439,
+   "peak_processes": 6,
+   "peak_disk_gb": 0.325,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-standard--kscoqvt/replay.mp4",
+  "duration_s": 0.04,
+  "fps": 25.0,
+  "frames": 1,
+  "sha256": "a894b7592370159d3a1ba5b7e128c6a55a1080d916a5214612192d2dc7ebf615"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：本次视频只有视频 0.00 s 的初始帧（全片时长0.04秒），白色模型位于场地中央，浅蓝圆顶黑色柱状对手位于右下方，两者地面投影明显相隔，未呈现追及到0.5米以内的关系。"
+     },
+     {
+      "t": null,
+      "log_step": 5,
+      "text_zh": "观察：log step 5 的机器人服务初始观测记录 control_steps=0、physics_steps=0、sim_time_s=0；run.json 的评分字段 n_actions=0、sim_steps=0、resets=0、live_success=0，完整会话未向服务提交动作。"
+     },
+     {
+      "t": null,
+      "log_step": 23,
+      "text_zh": "观察：log step 23 的最后消息为“This request was blocked by Gemini's filters”；此前日志是初始观测、源码检查、对手位置估计和本地模型测试，无已执行并验证的现场追及结果。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "不适用",
+      "how": "不适用",
+      "note_zh": "观察：评分字段 n_actions=0、sim_steps=0；会话在 log step 23 被提供方过滤器中断，未交付运动轨迹，录像仅保存初始画面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "unknown",
+   "t": null,
+   "log_step": 23,
+   "where": "首步前",
+   "evidence_zh": "观察：log step 23 明确报告 Gemini 过滤器拦截；评分字段 n_actions=0、sim_steps=0，视频 0.00 s 唯一帧仍是模型与对手分离的初态。推断：现场追及尚未开始且未形成有效结果便遭提供方中断，按外因中断处理，无法定位物理失败步骤。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [],
+  "root_causes": [],
+  "gap_zh": "观察：inputs/ 中有完整会话记录和仅一帧、时长0.04秒的本次录像，无逐帧状态日志；log step 23 是提供方过滤器拦截信息，未交付任何现场动作，评分记录同时标记 timeout=1。推断：这次中断应重试并排除在 agent 失败分析之外；保持原始 success=0，但不将外因记为没尝试或能力根因。本次没有可供判断步态、平衡或控制质量的运动过程，distance 原生指标也未报告，故现象和根因均留空。"
+ },
+ "summary_zh": "观察：本次视频 0.00 s 唯一帧中模型与对手仍明显分离，评分记录的动作数和仿真步数均为零，追及对手节点未达成（0/1）。观察：log step 23 明确显示 Gemini 过滤器拦截，会话只完成初始观测和本地检查；推断：属于有效现场动作开始前的外因中断，应重试并排除在 agent 失败分析之外。观察：没有运动录像或状态日志，且评分器未报告 distance，无法评价具体运动失败现象或能力根因。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "标签与证据一致：视频仅一帧初态，n_actions=0，log step 23 为 Gemini 过滤器拦截，未交付任何动作；节点未达成，现象和根因留空并有 gap_zh，符合外因中断规则。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:02:39Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--kscoqvt/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--kscoqvt/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7332057f7502705cfdaa3fbd6d0384697da5957460d5b0573daea53d38351b68",
+   "video_sha256": "a894b7592370159d3a1ba5b7e128c6a55a1080d916a5214612192d2dc7ebf615",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-standard--kwb4otb.json

@@ -0,0 +1,190 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-standard--kwb4otb",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 74,
+   "sim_steps": 74,
+   "physics_steps": 740,
+   "steps": 529,
+   "calls": 39,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 579,
+   "elapsed_s": 422.4,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.648,
+   "peak_processes": 9,
+   "peak_disk_gb": 1.167,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-standard--kwb4otb/replay.mp4",
+  "duration_s": 3.0,
+  "fps": 25.0,
+  "frames": 75,
+  "sha256": "8f441038d66b29a055eea09799b197aba50c415bdee550927a6d1a02f4183ddb"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 白色 MyoLegs 位于场地中央附近，带浅蓝圆顶的黑色柱状对手位于画面右下方，两者明显分离。已检查视频 0.0–2.96 s 的全部 75 帧，模型始终远离对手，未出现骨盆投影靠近对手至阈值内的画面。"
+     },
+     {
+      "t": 2.96,
+      "log_step": 61,
+      "text_zh": "观察：视频 2.0–2.96 s 模型的躯干翻转、双腿蜷曲，仍在对手左上方较远处；log step 61 的服务状态为 episode=11、control_steps=74、done=true，reason 为未追及、出界或跌倒的合并失败说明。评分器 live_success=0、graded_by_live=1，与该末回合未追及的画面一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–1.2 s，MyoLegs 以脚部接触木质地面支承身体，双腿改变姿态并移动；随后模型失衡下沉。整个视频未见模型接触柱状对手。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 1.8,
+   "log_step": 60,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 1.8 s，正在以双腿运动的 MyoLegs 已明显倾斜、转体，随后至视频 2.96 s 持续翻转下沉，始终未靠近右下方的对手；log step 60 记录末回合支撑阶段执行后，在仿真 0.74 s 骨盆高度 z=0.497、done=true。视频时间以所看帧为准，不将该仿真时刻直接当作视频时刻。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.8,
+    "log_step": 60,
+    "evidence_zh": "观察：视频 1.8–2.96 s，MyoLegs 躯干从倾斜进一步翻转，身体下沉、双腿蜷曲，没有恢复站立；log step 60 的末回合输出在仿真 0.70 s 记录骨盆高度 0.542，在仿真 0.74 s 记录 0.497 且 done=true，支持失衡跌倒导致回合结束。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：末回合肌肉控制未能建立稳定的腿部支撑。log step 58 说明 agent 选择左右支撑阶段并延长右侧支撑；log steps 59–60 执行该脚本后，输出依次为左支撑、右支撑及保持阶段，骨盆高度从仿真 0.35 s 的 0.828 降至仿真 0.74 s 的 0.497；视频 1.8–2.96 s 同时可见持续翻转下沉。证据支持控制失稳，但不足以区分接触布局、反馈律或具体参数问题。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志；评分器 reward 未报告 distance，故该原生指标为 null。录像只有低分辨率俯视图，无法精确测量骨盆距离或各脚受力；日志中的脚本编辑内容未完整展开，不能把失稳进一步归因于某个肌肉参数或反馈索引。末回合仿真终点为 0.74 s，视频时长为 3.0 s，未建立精确的仿真到视频映射，断��仅按画面作粗略定位。",
+  "unmet": {
+   "opponent_tagged": false
+  }
+ },
+ "summary_zh": "评分采用最后一次重置后的第 11 回合；视频全程未追及固定对手，节点最终为 0/1。MyoLegs 在视频约 1.8 s 已明显失衡，之后翻转下沉，末回合日志也记录骨盆高度降至跌倒阈值以下，根因支持标为控制失稳。缺少逐帧状态及完整末回合控制脚本，不能进一步确认具体控制参数，评分器也未报告最终距离。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点 opponent_tagged=false、断点（opponent_tagged，视频约1.8 s 倾斜失衡）、现象（全身·平衡）与根因（控制，已声明不足以细分）均与视频和 log step 60-61 一致；未发现需修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:00:08Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--kwb4otb/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--kwb4otb/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "25d4d5cb15f58504299f270cccda8d2c8cd9e22d30015a20ce09a72fca94fc34",
+   "video_sha256": "8f441038d66b29a055eea09799b197aba50c415bdee550927a6d1a02f4183ddb",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-standard--rn5emks.json

@@ -0,0 +1,192 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-standard--rn5emks",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 63,
+   "sim_steps": 63,
+   "physics_steps": 630,
+   "steps": 184,
+   "calls": 143,
+   "episodes": 3,
+   "resets": 2,
+   "reset_budget": 10,
+   "sim_steps_total": 184,
+   "elapsed_s": 2546.4,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.701,
+   "peak_processes": 14,
+   "peak_disk_gb": 1.324,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-standard--rn5emks/replay.mp4",
+  "duration_s": 2.56,
+  "fps": 25.0,
+  "frames": 64,
+  "sha256": "f96dc9a84d9b12267ad5879cdc8c3e29efd3cc6aa90e715976f678f1b82dcbc8"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s，白色MyoLegs模型位于场地中央，带浅蓝圆顶的黑色柱状对手位于画面右下方，两者明显分离。逐帧检查视频0.00–2.52 s，模型仅向画面下方移动少许，并逐渐屈腿下沉、倾斜，骨盆始终没有接近对手至0.5米的范围。"
+     },
+     {
+      "t": 2.52,
+      "log_step": 174,
+      "text_zh": "观察：视频2.52 s，模型仍在场内、远离对手；成功参考inputs/oracle/replay.mp4的视频约9.60 s则可见模型紧邻对手。log step 173重置后运行控制脚本，log step 174的服务状态为episode=3、control_steps=63、done=true，reason为未追及、出界或跌倒的合并失败信息；评分字段live_success=0、graded_by_live=1、n_actions=63，与最后回合一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频中MyoLegs以双脚落脚支承木质地面，双腿屈伸尝试推进身体；随后腿部弯曲、身体下沉倾斜。未见模型与柱状对手实际接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 1.6,
+   "log_step": 174,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约1.20–2.52 s，MyoLegs在推进过程中屈腿下沉，身体逐渐倾斜，未能走向右下方对手附近；视频约1.60 s作为失稳过程的导航锚点。log step 173启动最后的第3回合，log step 174报告该回合在63个控制步后失败结束，之后没有再重置或修复。推断：失败发生在追及对手的行走操作中。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.6,
+    "log_step": 174,
+    "evidence_zh": "推断：视频1.20–2.52 s可见MyoLegs腿部屈曲、身体下沉倾斜，且仍在场内并远离对手；log step 174服务报告第3回合失败，任务控制率100 Hz（log step 15），63个控制步只对应仿真0.63 s，尚未达到仿真20 s时限。结合任务的失败条件，符合骨盆过低的跌倒失败，而非正常步态继续行进；视频未延伸到身体完全倒伏。此处仿真时长不作为视频时间。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 1.6,
+    "log_step": 130,
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 130的最终脚本读取本体反馈，以骨盆位置与姿态PD、腿部关节PD和交替迈步目标计算肌肉指令，骨盆高度目标至少为0.95米；log step 173实际运行该脚本，log step 174显示第3回合完成63个控制步后失败，视频1.20–2.52 s仍出现下沉倾斜。控制器没有实现其维持支撑并推进的目标，支持控制层面的归因；没有足够的现场跟踪误差或接触力记录来确认具体子类。"
+   }
+  ],
+  "gap_zh": "无逐帧状态日志。现有320×240俯视视频可判读两者始终分离及模型下沉倾斜，但不能直接量出骨盆高度、逐足接触力或精确水平距离；评分器未报告distance，原生失败原因也是多个条件的合并信息。跌倒判读由画面、场内位置、回合控制步数和失败条件共同推断，断点时间仅为视频约1.60 s的粗略锚点。日志中的本地副本测试属于会话诊断，不能当作第3回合的状态轨迹；具体失稳机制未确定。"
+ },
+ "summary_zh": "最后评分的第3回合中，MyoLegs尝试以双脚支撑推进，但始终没有接近固定对手至成功距离，追及对手节点终态为0/1。视频约1.20–2.52 s显示模型屈腿下沉、身体倾斜，结合log step 174的63个控制步后失败状态，判断其在追及过程中失去平衡，最终控制器未能维持有效支撑。没有逐帧状态日志或评分器距离数值，精确骨盆高度及具体控制失稳机制无法确认。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点 opponent_tagged 未达成、final=0 与视频（0–2.52 s 两者始终分离）及服务读数（第3回合 63 控制步 done=true 失败，0.63 s 仿真内不可能超时/出界，故为跌倒）一致；断点、平衡现象、控制根因均有证据且已标为推断，未发现需改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:08:02Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--rn5emks/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--rn5emks/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "696d86e375e0faf7fff057c566b746c9d701e8aaaca0fd8aaa43ec24f0862481",
+   "video_sha256": "f96dc9a84d9b12267ad5879cdc8c3e29efd3cc6aa90e715976f678f1b82dcbc8",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-standard--tmcuy4h.json

@@ -0,0 +1,187 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-standard--tmcuy4h",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 86,
+   "sim_steps": 86,
+   "physics_steps": 860,
+   "steps": 922,
+   "calls": 931,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 922,
+   "elapsed_s": 3308.8,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-standard--tmcuy4h/replay.mp4",
+  "duration_s": 3.48,
+  "fps": 25.0,
+  "frames": 87,
+  "sha256": "b5967619cd3513b8b4d7b977a92b19aea24de82bab8b14eccf7cf3b909724176"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 118,
+      "text_zh": "观察：log step 118 执行最后一次 reset 后开始控制；log step 137、139 的服务状态均为 episode=11、resets_left=0，评分器记录 n_actions=86、live_matches_replay=1。本次视频从白色模型站在场地中部开始，只呈现这一最后回合，不包含前十回合的测试。"
+     },
+     {
+      "t": 2.4,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.40 s 的俯视画面中，白色模型双腿交替移动，身体略向画面下方推进；黑色柱状、浅蓝圆顶对手一直在其右下方，两者间仍有明显空地。"
+     },
+     {
+      "t": 3.44,
+      "log_step": 139,
+      "text_zh": "观察：已检查视频全程每约 0.20 s 的画面及视频 2.40–3.44 s 的每帧；白色模型末段倾倒，直到视频 3.44 s 仍与对手明显分离，圆顶未变绿。成功参考视频 9.64 s 中模型已经靠到对手旁，圆顶变绿；该参考仅用于辨认达标外观。log step 139 的最后回合服务状态为 done=True、reason=failure，评分器 success=0。"
+     },
+     {
+      "t": 3.44,
+      "log_step": 139,
+      "text_zh": "推断：本视频 87 帧对应初态及最后回合的 86 个控制步，以 25 fps 播放；控制步 k 对应视频 k/25 s，而脚本每步 dt=0.01 对应仿真 k/100 s。log step 139 的 control_steps=86、骨盆高度 0.493 m 与视频 3.44 s 的下沉倾倒末帧吻合，据此核对映射；不能把视频时长当作仿真时长。结合全程可见间距及原生失败结果，“追及对手”从未达成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–2.40 s，MyoLegs 的双脚在木质地面交替落脚并支承身体，双腿带动本体向前移动；视频 2.80–3.44 s 身体倾倒、支撑未能维持。未见模型与柱状对手实际接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 2.8,
+   "log_step": 129,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 2.80 s 起白色模型躯干倾斜加重，随后骨盆持续下沉，视频 3.44 s 结束时仍未靠近对手。log step 129、131、135、137、139 的最后回合读数依次给出骨盆高度 0.774、0.711、0.613、0.510、0.493 m，最终服务判负。推断：机器人正在执行追及导航时失衡，且直到回合结束未恢复或完成该步；断点取“追及对手”。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 2.8,
+    "log_step": 129,
+    "evidence_zh": "观察：视频 2.80–3.44 s 的连续画面中，白色模型由较直立姿态变为明显倾斜、压低的姿态，未恢复站立；log step 129 的骨盆高度为 0.774 m，log step 139 为 0.493 m，后者同时报告最后回合 done=True、failure。推断：模型失去平衡，骨盆下降越过任务的 0.5 m 跌倒阈值而终止，未追及对手。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：肌肉步行控制未能在支撑切换中稳定本体。log step 123、128、130 显示 agent 在最后回合连续修改摆腿伸膝和接触阶段逻辑；log step 136 又在骨盆过低时添加伸膝与髋力矩恢复指令，但 log step 135、137、139 的实际本体读数显示骨盆仍由 0.613 m 降至 0.510 m、0.493 m，视频 3.20–3.44 s 亦显示倾倒继续。上述证据支持控制未能稳定身体，尚不足以把失败唯一归于接触估计、肌肉模型或某一增益参数。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志；输入仅含远距离俯视视频，足底接触细节及真实接触力无法分辨，因而不能验证控制器接触估计的误差或进一步定位失稳机制。评分器未报告原生 distance 数值，故该指标为 null；未用图像估计替代米制距离。"
+ },
+ "summary_zh": "最后第 11 回合中，MyoLegs 通过双脚支撑尝试向前移动，但一直未靠近固定对手，追及节点始终未达成，最终为 0/1。视频约 2.80–3.44 s 模型失衡下沉，最后回合 log step 139 的骨盆高度为 0.493 m 且服务判负，支持肌肉步行控制未能维持平衡。缺少逐帧状态、接触力和原生距离数值，无法进一步确定失稳的具体控制环节。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "标注与视频、日志一致：追及节点未达成（0/1），控制步70对应视频2.8 s（log step 129 的 step 70、骨盆0.774 m）映射成立，断点、平衡现象和控制根因均有证据支持，未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:01:29Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--tmcuy4h/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--tmcuy4h/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "011237f53cf6207fff0df0b29957d04883e35304af6b7440c07506eda750d428",
+   "video_sha256": "b5967619cd3513b8b4d7b977a92b19aea24de82bab8b14eccf7cf3b909724176",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/chase_tag/trajectories/myosuite-chase-tag-i00-standard--uxe89cz.json

@@ -0,0 +1,195 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-chase-tag-i00-standard--uxe89cz",
+ "taskKey": "myosuite/chase_tag",
+ "dag_sha256": "44d1913cbf93a1cace28fb27d008741b9ec7d5862825d6d9128583801a384d16",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 203,
+   "sim_steps": 203,
+   "physics_steps": 2030,
+   "steps": 686,
+   "calls": 23,
+   "episodes": 4,
+   "resets": 3,
+   "reset_budget": 10,
+   "sim_steps_total": 686,
+   "elapsed_s": 2649.3,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.711,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.533,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-chase-tag-i00-standard--uxe89cz/replay.mp4",
+  "duration_s": 8.16,
+  "fps": 25.0,
+  "frames": 204,
+  "sha256": "550a6409faa61f4cd3fa8b206008e1b4088b97040cffcec9024a75eb627efc0f"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "opponent_tagged": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 白色MyoLegs模型位于场地中央，浅蓝圆顶的黑色柱状对手在画面右下方，两者明显分离；视频 2.0、4.0、6.0 s 模型虽有向前位移，骨盆仍未接近对手。"
+     },
+     {
+      "t": 8.12,
+      "log_step": 260,
+      "text_zh": "观察：视频 7.2、7.4、7.6、7.8、8.0、8.04、8.08、8.12 s 模型逐渐倾倒，直到最后一帧仍与对手隔着明显地面间隙；对手圆顶未出现参考成功录像末帧的绿色提示。log step 260 明确重置为 Episode 4，服务最终 control_steps=203、done=True，返回失败原因，服务骨盆读数为(0.481,-1.651,0.494) m；评分器字段 live_success=0、n_actions=203 与该末局相符。"
+     },
+     {
+      "t": 8.12,
+      "log_step": 260,
+      "text_zh": "推断：评分回合是 log step 260 的 Episode 4，前几局测试不计入本轨迹。该局203个控制步加初始画面与视频204帧相符，末帧203/25=视频 8.12 s 的下沉姿态与服务骨盆高度0.494 m相互核对；控制步dt=0.01对应仿真 2.03 s，不能把该仿真时长直接当视频时长。脚本以服务骨盆读数和估计对手坐标计算末端距离3.324 m，结合全程画面始终明显分离，追及距离未达0.5 m。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "推断：视频 2.0、4.0、6.0 s 的落脚与身体位移显示，MyoLegs以双腿和双脚接触木质地面、承重推进骨盆；视频约 7.6 s 起明显倾倒，未见模型接触黑色柱状对手。没有手部或轮式底盘参与。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "opponent_tagged",
+   "t": 7.6,
+   "log_step": 260,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 6.8–8.12 s，正在向对手移动的MyoLegs躯干持续倾斜、骨盆下沉，视频约 7.6 s 已明显倾倒，末帧仍与对手明显分离；log step 260 的 Episode 4 最后服务骨盆高度为0.494 m并返回done=True及失败原因。该局未恢复支撑或完成追及，断点为追及对手。"
+  },
+  "unmet": {
+   "opponent_tagged": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 7.6,
+    "log_step": 260,
+    "evidence_zh": "观察：视频 7.0、7.2、7.4、7.6、7.8、8.0及8.12 s，MyoLegs从倾斜站姿逐渐转为躯干接近横向、骨盆下降的倾倒姿态，没有恢复直立；log step 260 的最终服务骨盆高度0.494 m低于任务0.5 m跌倒阈值，此时回合结束。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 254 可见agent脚本先在本地仿真生成肌肉动作序列，再按最多50步一批提交；服务返回的骨盆读数用于打印距离和检查结束，未用于重算后续动作。log step 259 将生成长度改为205步，log step 260 运行该脚本时，最后一局服务骨盆高度从150控制步的0.874 m降至200控制步的0.560 m，再到203控制步的0.494 m；视频 6.0–8.12 s 同步显示逐渐倾倒，说明所执行肌肉控制未能维持追及过程所需支撑。现有证据只支持控制层面的归因，不能确定具体肌群、反馈增益或接触布局的问题。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或评分器原生distance数值；log step 260 的3.324 m由agent脚本结合服务骨盆读数与估计对手坐标计算，故native_metrics.distance留空。俯视录像分辨率较低，不能解析逐足接触力或具体关节失稳过程；log step 254 的生成函数中段在输入记录中已被截断，根因不细分为反馈控制、轨迹或逆解或参数调试。"
+ },
+ "summary_zh": "评分采用最后一局 Episode 4（log step 260），MyoLegs以双腿落脚推进，但直到视频 8.12 s 结束仍与固定对手明显分离，追及对手节点全程未达成，最终0/1。视频约 7.6 s 出现明显倾倒，最终服务骨盆高度0.494 m，所执行肌肉控制未能维持支撑。没有逐帧状态日志及评分器原生距离值，具体控制失稳机制无法进一步定位。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "评分回合为 Episode 4（log step 260，203 控制步对应视频 204 帧），追及节点全程未达成，终局骨盆 0.494 m 触发跌倒；断点、平衡现象与停在“控制”的根因均有证据支持，未发现需修改之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:04:10Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--uxe89cz/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-chase-tag-i00-standard--uxe89cz/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "98d8f191437a047df32ed7a389d6f4d28d8286e1ed352f9470639c5275e034ee",
+   "video_sha256": "550a6409faa61f4cd3fa8b206008e1b4088b97040cffcec9024a75eb627efc0f",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-privileged--5ybpkdi.json

@@ -0,0 +1,202 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-privileged--5ybpkdi",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 99,
+   "sim_steps": 36,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 136,
+   "agent_resets": 21398,
+   "agent_sim_steps": 472278,
+   "agent_records": 0,
+   "agent_replays": 3,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.623,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.841,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-privileged--5ybpkdi/replay.mp4",
+  "duration_s": 1.48,
+  "fps": 25.0,
+  "frames": 37,
+  "sha256": "7982fccadcf4e9b1ee5a5805ffde0d4616a1a433349eef123530f7fcec0249f9"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.4,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–1.44 s 的全部37帧；MyoHand手指屈曲包围飞机的机身及翼根，飞机只在桌面附近小幅移动和改变朝向，没有出现随手在空中飞行的完整过程。"
+     },
+     {
+      "t": 1.44,
+      "log_step": 223,
+      "text_zh": "观察：视频 1.04–1.44 s 飞机仍贴近桌面，手附近的绿色标记逐渐远离飞机；log step 222 记录提交文件含99行动作，log step 223 对该文件两次重放均仅执行36个控制步，原生终止原因均为物体距参考或手腕超过0.25米，且终态哈希一致；评分器字段sim_steps=36、replay_success=0与提前结束一致。"
+     },
+     {
+      "t": null,
+      "log_step": 223,
+      "text_zh": "推断：该单一event节点要求走完99个控制步，提交文件实际只重放36步便终止，结合视频中没有完成空中运动，飞机循迹飞行未达成；没有把会话中其他候选轨迹的局部成绩计入本次进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约 0.32–1.44 s，MyoHand单手弯曲的手指从机身及翼根周围包围、夹持玩具飞机，使其小幅移动；飞机未被带入完整的空中飞行。此处记录可见合手接触，不据此声称稳定承载或成功抬起。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 1.08,
+   "log_step": 223,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 1.08 s 起，MyoHand仍在飞机周围弯曲手指，飞机停留在桌面附近，未随远离的绿色标记升空；视频直到 1.44 s 仍未恢复。log step 223 的提交文件重放在第36控制步因原生距离条件终止，未走完99步。断点锚定可见偏离开始明显的画面，不将控制步或agent工作时间换算成视频时间。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 1.08,
+    "log_step": 223,
+    "evidence_zh": "观察：视频 1.04–1.44 s，机器人手指继续围住飞机，但飞机只在桌面附近移动、转动，没有跟随所需的空中飞行过程；绿色标记与飞机明显分离。log step 223 对最终文件的两次重放均因物体距参考或手腕超过0.25米提前终止。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 194–195 的会话搜索用前臂目标偏移与肌肉命令优化持续对向接触，输出虽报告接触评分提高，飞机相对初态的竖直位移仍接近零；这些仅作控制问题的会���证据。视频 0.40–1.44 s 中，交付动作也只能使手指围住飞机并小幅移动，未产生所需的升空运动；结合log step 223 的一致重放失败，支持控制层失败，尚不能确定具体控制子类。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志或成功示范；评分器reward未报告track_pos_err、track_rot_err、track_ok_fraction及track_max，故四项原生指标均为null，未借用其他会话候选的数值。视频可见合手接触和未完成飞行，但无法精确确认各指接触力、承载分布及每步误差；不能进一步区分抓取和接触设计、轨迹或逆解、反馈控制。"
+ },
+ "summary_zh": "评分轨迹中，MyoHand合手接触飞机并使其小幅移动，但飞机始终贴近桌面，未完成空中循迹，唯一节点未达成（0/1）。视频约 1.08 s 后偏离明显，最终提交在两次重放中均于第36控制步触发原生距离终止；证据支持控制层失败。没有状态日志和成功示范，四项精确循迹指标未报告，无法细分控制根因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点、位姿不准现象与控制根因（停在大类并写明推断）与视频和评分器字段一致，未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:55:04Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--5ybpkdi/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--5ybpkdi/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9ba7f87bf4566fa1396f871b029667f1a1671276b278500da226a1232f0971dc",
+   "video_sha256": "7982fccadcf4e9b1ee5a5805ffde0d4616a1a433349eef123530f7fcec0249f9",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-privileged--biszf8k.json

@@ -0,0 +1,210 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-privileged--biszf8k",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 37,
+   "sim_steps": 37,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 348,
+   "agent_resets": 1103,
+   "agent_sim_steps": 10390,
+   "agent_records": 4,
+   "agent_replays": 10,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.583,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.269,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-privileged--biszf8k/replay.mp4",
+  "duration_s": 1.52,
+  "fps": 25.0,
+  "frames": 38,
+  "sha256": "f92ac7f23cd0d309deb34693826910b46dcbb6e09fce1bdb55695e4af799db1c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.96 s 的两个视角中，MyoHand手指贴近并触及银色玩具飞机上侧，手部小幅调整；飞机一直贴着桌面，未见被夹住带起。已检查全部38帧。"
+     },
+     {
+      "t": 1.08,
+      "log_step": null,
+      "text_zh": "观察：视频 1.04–1.48 s，绿色参考标记向上、向外移动，银色飞机仍在桌面原处附近，未随手形成空中运动或跟随参考。"
+     },
+     {
+      "t": 1.48,
+      "log_step": 442,
+      "text_zh": "观察：log step 441生成最终37行、45维动作并重放，track_ok=26、终态哈希a07d006b9acee9f6；log step 442官方评分器两次重放同一提交均在37步失败，原因是物体距参考或手腕超过0.25 m，哈希一致。视频末帧（视频 1.48 s）飞机仍在桌面；未完成DAG要求的99步循迹。"
+     },
+     {
+      "t": null,
+      "log_step": 438,
+      "text_zh": "观察：log step 438对已保存的交付动作逐步重放，零基控制步26位移误差43.14 mm，控制步27为60.72 mm，控制步36为289.88 mm、转动误差0.5189 rad并终止，飞机质心高度始终约17–18 mm。这是交付文件诊断，log step 435和442的重放哈希相同；不把这些控制步或agent工作时间换算成视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频 0.00–1.48 s，MyoHand单手指端在飞机上侧贴触、擦动；未见手指从相对两侧夹住飞机，也未把飞机带离桌面，因此不记抓握。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better"
+  },
+  {
+   "name": "track_rot_err",
+   "value": null,
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  },
+  {
+   "name": "track_ok_fraction",
+   "value": null,
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better"
+  },
+  {
+   "name": "track_max",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 1.04,
+   "log_step": 442,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 1.04 s 起，参考标记开始明显离开桌上飞机，而MyoHand手指没有夹持、带起飞机；至视频 1.48 s 均未补成抓取或空中循迹。log step 442的官方重放在37步以物体距参考或手腕超过0.25 m失败。断点为正在尝试的飞机循迹飞行，视频时间仅为未能带起飞机的粗略导航锚点。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 1.04,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.00–1.48 s，MyoHand指端虽触碰飞机上侧，却没有形成能把飞机带起的包围或对向夹持；视频 1.04–1.48 s 参考向上离开时飞机仍贴桌。缺的是承重抓握接触，未见已持起后再滑落的过程。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 440的最终生成脚本每步用rest_action初始化45维动作，只改前6个前臂执行器，手指肌肉保持静息动作；log step 441生成的37行动作与log step 442两次失败重放的终态哈希一致。结合视频 0.00–1.48 s 指端停留于飞机上侧、飞机始终未被带起，所提交控制只调整前臂而未形成承重夹持，接触安排不足以将前臂运动传给飞机。没有成功抓取示范或已验证的替代抓法，替代可行性未知。"
+   }
+  ],
+  "gap_zh": "没有状态日志及成功示范，无法核对逐步接触力或验证替代抓法。run.json的reward及官方重放结果未报告DAG四项原生指标的数值，native_metrics均置null；log step 438含交付文件的舍入诊断误差，只作辅助证据。视频不能精确给出厘米、弧度误差或达标比例；断点时间依据可见的参考离开而粗定，未将20毫秒控制步映射到25 fps视频。"
+ },
+ "summary_zh": "提交轨迹只执行37步，飞机始终留在桌面，飞机循迹飞行节点未完成，最终为0/1。视频约 1.04 s 起参考明显离开，手指未形成承重夹持；最终脚本只控制前臂、手指保持静息动作，支持抓取和接触设计不足，官方两次重放均以距离超限失败。无状态日志或成功示范，精确接触力和替代抓法尚不可确认，评分器未报告的原生指标保留为空。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与视频、日志一致：飞机始终留在桌面，节点未达成，断点、现象（没建立接触）和根因（抓取和接触设计，推断标注）有证据支持，时钟未换算。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:53:53Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--biszf8k/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--biszf8k/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7dcdf517919e07f91e7295a7484e2b3a78e5c82b8774ef847f2778d9b2dde086",
+   "video_sha256": "f92ac7f23cd0d309deb34693826910b46dcbb6e09fce1bdb55695e4af799db1c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-privileged--efg8rkx.json

@@ -0,0 +1,205 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-privileged--efg8rkx",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 36,
+   "sim_steps": 36,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 52,
+   "agent_resets": 319,
+   "agent_sim_steps": 273,
+   "agent_records": 1,
+   "agent_replays": 12,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.979,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.985,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-privileged--efg8rkx/replay.mp4",
+  "duration_s": 1.48,
+  "fps": 25.0,
+  "frames": 37,
+  "sha256": "0e528e76beb57fd29c0c0da83095ac5187dc0186a33ec1bdfa095071f36bc97b"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 玩具飞机在桌面上，MyoHand单手位于飞机上侧；初始接近目标不等于完成整段循迹。"
+     },
+     {
+      "t": 0.96,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频 0.00–1.44 s 的全部37帧；手指在飞机上侧屈曲、触拨，但飞机一直留在桌面，没有随手形成空中位移和转向；视频 0.96–1.44 s 仍未带起飞机。"
+     },
+     {
+      "t": 1.44,
+      "log_step": 128,
+      "text_zh": "观察：视频 1.44 s 末帧飞机仍在桌面；log step 126 生成并保存36行控制指令，服务返回done=True、track_steps=99、track_ok=24及原生距离终止原因。log step 128 对该交付文件的两次官方重放均在控制步36失败且终态哈希相同，评分器reward也报告n_actions=36、replay_success=0、deterministic=1。未完成99步循迹；不将控制步时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "观察：视频 0.00–1.44 s，MyoHand单手指端在玩具飞机机翼及机身上侧触拨，飞机仍由桌面承托；未见拇指与其余手指形成对向夹持，实际接触不记为抓握。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.96,
+   "log_step": 128,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 0.96 s，MyoHand仍在飞机上侧屈指，飞机留在桌面；至视频 1.44 s 仍无有效带起。log step 126 保存的交付只有36个控制步，log step 128 的官方重放确认该文件因飞机距参考或手腕超过原生阈值而提前终止，飞机循迹飞行始终未完成。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 0.96,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.00–1.44 s 中，MyoHand指端虽触拨飞机上侧，但未包住或夹住飞机，飞机始终留在桌面。所需的承载抓持从未建立，未见先提起再滑落的过程。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "evidence_zh": "推断：视频 0.64–1.44 s 的手指作用位置集中在飞机上侧，未形成能带起飞机的对向承载接触；log step 125 的交付生成脚本直接跟踪参考手臂及手指关节角，没有建立抓持的确认或针对飞机接触的修正，log step 126 随即记录提前终止。结合飞机始终未离桌，当前抓持接触布局未奏效；未验证可��替代，不能由会话搜索失败断言任务无解。",
+    "source_zh": "视频 / 日志"
+   }
+  ],
+  "gap_zh": "没有状态日志或成功示范，无法确定具体指骨接触对、接触力及可行替代抓法。评分器reward未单独报告DAG四项原生指标，native_metrics值均留null；log step 126 仅给出所保存运行在控制步36终止时的服务读数track_pos_err=0.26702 m、track_rot_err=0.569 rad、track_ok=24，不能据此补出99步完整达标比例或精确历史最大误差。视频逐帧检查足以判断未带起和未完成，不作仿真与视频时钟的等速假设。"
+ },
+ "summary_zh": "观察：评分视频全段中，MyoHand仅在玩具飞机上侧触拨、屈指，飞机始终留在桌面，飞机循迹飞行未完成，节点终态为0/1。推断：当前接触布局未建立有效抓持，无法带动飞机飞行；log step 128 的两次官方重放均在控制步36原生终止，支持物理失败而非交付不一致。缺少状态日志、成功示范及评分器完整循迹指标，无法细分接触力或验证替代抓法。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "视频 0–1.44 s 飞机始终在桌面（绿球为参考标记），节点未完成、断点、接触（推/扫非抓握）、现象与根因均有证据支持；时钟未把控制步换算成视频时间。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:50:12Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--efg8rkx/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--efg8rkx/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "ddde06e13533a14be83be13ba4e9ed58f26953ac5e2b9e18102830f024759f05",
+   "video_sha256": "0e528e76beb57fd29c0c0da83095ac5187dc0186a33ec1bdfa095071f36bc97b",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-privileged--p6hl228.json

@@ -0,0 +1,215 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-privileged--p6hl228",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 99,
+   "sim_steps": 25,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 104,
+   "agent_resets": 247,
+   "agent_sim_steps": 6776,
+   "agent_records": 3,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.998,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.257,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-privileged--p6hl228/replay.mp4",
+  "duration_s": 1.04,
+  "fps": 25.0,
+  "frames": 26,
+  "sha256": "061e8136c533839b105b25408e7ae7016228e86a7929518372e64e95e55bf6ff"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s的两个视角中，单只红色肌腱骨骼手位于飞机上方，飞机靠近桌面；初态相近不代表整段参考运动已完成。"
+     },
+     {
+      "t": 0.76,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频0.00–1.00 s（全部26帧）；手指逐渐弯曲，视频0.60–0.80 s拇指与其余弯曲手指从两侧包围飞机机身，飞机被带动横移，但仍贴近桌面。"
+     },
+     {
+      "t": 0.88,
+      "log_step": null,
+      "text_zh": "观察：视频0.80–1.00 s，飞机仍在红色手指下方、靠近桌面，绿色参考标记与飞机的间隔持续拉大；末帧未出现飞机随参考在空中飞行的结果。"
+     },
+     {
+      "t": null,
+      "log_step": 135,
+      "text_zh": "观察：log step 135运行最后一次生成脚本并保存交付文件，输出在控制步25提前终止，原因是飞机距参考或手腕超过0.25 m；该会话报告控制步20位移误差0.0597 m，控制步25位移误差0.25538 m、转动误差0.2737 rad、track_ok=19。评分器reward独立报告sim_steps=25、n_actions=99、task_failure=1、deterministic=1。"
+     },
+     {
+      "t": 1.0,
+      "log_step": 135,
+      "text_zh": "推断：log step 134的脚本将提前结束的动作以最后一行补至99行，log step 135保存后无后续覆盖记录；它的25步终止与评分器及视频短暂的桌面附近运动一致，因此认定为本次交付来源。结合视频中未完成的空中运动与评分器实际只执行25步，完整99步循迹节点从未达到；日志数值是生成会话的佐证，未将agent工作时间或控制步的仿真时间当作视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频0.60–1.00 s，MyoHand单只手的拇指与其余弯曲手指从相对两侧包围、夹住玩具飞机机身并带动其移动；前段手指先从上方接触、压擦飞机表面。此处记录可见的合手接触，不据此认定飞机完成抬升或循迹；半透明白色手形是参考显示，不是第二只操作手。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.8,
+   "log_step": 135,
+   "where": "操作中",
+   "evidence_zh": "观察：视频0.80–1.00 s，红色手指合拢并带动飞机在桌面附近移动，飞机与绿色参考标记明显分离，直至视频结束没有修复。log step 135最后保存的生成会话在控制步25因飞机距参考或手腕超过0.25 m而终止，评分器实际执行步数也为25；失败发生在飞机循迹操作中。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 0.8,
+    "log_step": 135,
+    "evidence_zh": "观察：视频0.80–1.00 s，实际飞机保持在红色手下方的桌面附近，朝向和位置没有随分离的绿色参考显示共同变化，未形成要求的空中飞行过程。log step 135最终生成会话的位移误差由控制步20的0.0597 m增至控制步25的0.25538 m，与视频后段的空间偏离相符。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 134,
+    "evidence_zh": "推断：log step 134的最后交付脚本以参考位姿生成前臂逆解和腕部、手指肌肉控制，并增加屈肌激活来合手；log step 135执行该脚本时，飞机仅小幅升高（控制步24相对初态14.8 mm），位移误差却迅速增加并提前终止。视频0.60–1.00 s也显示合手后飞机只在桌面附近移动，支持控制未能产生要求的持物循迹运动；缺少本次逐步关节与接触数据，不能细分接触设计、逆解或反馈控制。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "未提供状态日志和成功示范；视频仅640×240且叠加参考手形，无法独立测量厘米级误差、接触力和各指承载。输入reward未报告DAG列出的四项原生循迹数值，故native_metrics均为null；log step 135中的误差和计数属于最后生成会话，未冒充评分重放的原生数值。最后控制脚本的日志中间部分被截断，也无法据此定位具体控制子类。"
+ },
+ "summary_zh": "单手合拢夹住飞机并带动其在桌面附近移动，但未完成飞机循迹飞行，节点最终0/1。视频0.80–1.00 s显示飞机与参考标记持续分离，最后保存的生成会话在控制步25提前终止，评分器同样只执行25步；证据支持位姿不准及一般控制失败。缺少状态日志、接触力与评分重放的精确循迹指标，不能细分控制根因。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "单节点未达成、断点在该节点、位姿不准与停在“控制”的根因均有视频和评分器25步终止支持；时钟未把控制步换算成视频时间，未发现需改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:58:09Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--p6hl228/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--p6hl228/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "02ccee890d7f4fd0ba8158f56323dc87d2b7ad2212f6fee6aebd7788d13d99c6",
+   "video_sha256": "061e8136c533839b105b25408e7ae7016228e86a7929518372e64e95e55bf6ff",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-privileged--pawfrco.json

@@ -0,0 +1,199 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-privileged--pawfrco",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 99,
+   "sim_steps": 38,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 141,
+   "agent_resets": 5034,
+   "agent_sim_steps": 5115,
+   "agent_records": 103,
+   "agent_replays": 3,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 1.515,
+   "peak_processes": 17,
+   "peak_disk_gb": 0.243,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-privileged--pawfrco/replay.mp4",
+  "duration_s": 1.56,
+  "fps": 25.0,
+  "frames": 39,
+  "sha256": "e1e646924f83c4ced3ac1a174be7b8faea34397ba23461185577b903b673297c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.64,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.68 s，MyoHand单手向桌面上的白色玩具飞机合拢，飞机随后随手离开原来的桌面位置；视频 1.20–1.52 s，飞机仍夹在弯曲的手指之间并随手转动。"
+     },
+     {
+      "t": 1.36,
+      "log_step": 93,
+      "text_zh": "观察：视频 1.20–1.52 s，白色飞机仍在手中，但绿色参考标记逐渐向画面左下方远离飞机；log step 93 对最终提交文件进行两次独立重放，均只执行38/99个控制步，原因均为飞机距参考或手腕超过0.25米，终态哈希相同；评分器reward亦报告sim_steps=38、deterministic=1、task_failure=1。完整99步循迹未完成，不能因开局误差小而记为完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 0.56–0.68 s及1.20–1.52 s，MyoHand单手的拇指和弯曲手指围住飞机的翼根、机身后部附近，飞机随手抬起、转动；推断：由合拢包夹及持物共同行动判断为抓握，具体接触点没有状态日志可精确核对。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 1.28,
+   "log_step": 93,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.20–1.52 s，机器人已把飞机带离桌面并继续持握，但飞机与运动中的绿色参考标记明显分离，且差距持续扩大至片尾；log step 93 验证的最终提交在控制步38提前终止，未走完99步。推断：失败发生在持物循迹过程中，视频约1.28 s为开始明显拉开差距的导航锚点，并非从agent时间换算。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 1.28,
+    "log_step": 93,
+    "evidence_zh": "观察：视频 1.20 s绿色参考标记还在白色飞机附近，视频 1.28–1.52 s标记向画面左下方持续移走，机器人持握的飞机仍停留在手旁，未跟上参考的移动；片尾没有看见飞机脱手。log step 93 的最终文件重放因飞机距参考或手腕超过0.25米提前终止。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：agent脚本的运动控制未能带着已抓住的飞机跟上参考。log step 79 的脚本在抓取后将参考臂腕增量加到抓取姿态上，并按物体误差反复修正；log step 88 的修正试验仍在控制步38左右终止，log step 89 保存并重放的最终动作仅26步共同达标、控制步38终止，log step 93 两次独立重放复现该结果，与视频 1.28–1.52 s持物滞后的现象一致。会话诊断log step 83–84还显示指令与实际臂部运动存在偏差，但它们不是评分状态日志，因此不据此细分反馈控制、逆解或参数问题。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有状态日志或成功参考视频；右视角在视频 0.72–1.16 s被桌面遮挡，左视角仍能看到手与飞机，但不能精确还原各指接触力。reward未报告track_pos_err、track_rot_err、track_ok_fraction和track_max，故四项原生指标留null；无法定位首次越过0.05米的精确视频时刻，断点仅用可见分离作粗略锚点。log step 95 关于载荷、抓法后续不可达的解释是agent自述，会话诊断也未排除其他合法策略，不能据此认定任务无解或进一步确定控制子类。"
+ },
+ "summary_zh": "机器人单手夹住并抬起玩具飞机，但唯一的“飞机循迹飞行”节点未完成，最终为0/1。视频 1.28–1.52 s可见飞机仍在手中而参考标记持续远离；log step 93确认最终文件两次重放均在控制步38/99提前终止，支持运动控制未能跟上动态参考。缺少状态日志和评分器原始循迹指标，不能精确核对误差越限时刻或细分控制根因。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未完成(0/1)、单手抓握飞机后参考标记远离、最终文件重放控制步38/99终止，均与视频及log step 93一致；断点、动态时机现象和控制根因表述克制，未发现需要修改的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:54:31Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--pawfrco/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--pawfrco/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "41a5fd54d7076e6e653568be0a60561768843429cb60b32f2233aa714a9633e5",
+   "video_sha256": "e1e646924f83c4ced3ac1a174be7b8faea34397ba23461185577b903b673297c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-privileged--wnelj7s.json

@@ -0,0 +1,206 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-privileged--wnelj7s",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "instance_ok": 1,
+   "n_actions": 37,
+   "sim_steps": 37,
+   "task_failure": 1,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 102,
+   "agent_resets": 400,
+   "agent_sim_steps": 12963,
+   "agent_records": 0,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "infra_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 2.435,
+   "peak_processes": 21,
+   "peak_disk_gb": 0.251,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-privileged--wnelj7s/replay.mp4",
+  "duration_s": 1.52,
+  "fps": 25.0,
+  "frames": 38,
+  "sha256": "596f6c1af80333ffe478b3e75ff6973023a5d07761cdc37c5315bc6d7561fad6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 双视角显示玩具飞机在桌面上，MyoHand 单手手指位于飞机上侧。逐帧检查视频 0.00–1.48 s，飞机只有少量桌面内位移，没有被手带起或在空中飞行。"
+     },
+     {
+      "t": 1.08,
+      "log_step": null,
+      "text_zh": "观察：视频 1.00–1.48 s，手指继续屈曲并贴压飞机上侧，飞机仍贴着桌面；未出现包住机身并带起飞机的夹持。"
+     },
+     {
+      "t": null,
+      "log_step": 171,
+      "text_zh": "观察：log step 162 的生成命令保存了 37×45 的动作数组；log step 171 对该提交文件进行两次重放，均在 37 个控制步后返回 success=false，原因是飞机距参考或手腕超过 0.25 m，终态哈希均为 89627a70296b910b。该结果与视频中飞机未起飞一致，也未走完 DAG 要求的 99 个控制步。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频 0.80–1.48 s，MyoHand 单手弯曲的手指贴压玩具飞机上侧及尾部附近；飞机仍由桌面承托，未观察到对向包夹并抬升飞机的抓握。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 1.08,
+   "log_step": 171,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.00–1.48 s，MyoHand 正在飞机上方屈指，飞机始终留在桌面，没有被带起；这一尝试直到片尾都未修复。log step 171 的提交文件重放在 37 个控制步后因飞机距参考或手腕超过 0.25 m 提前结束。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 1.08,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.80–1.48 s，MyoHand 的手指在飞机上侧弯曲、压触，飞机仍在桌面；没有形成使飞机离桌并随手运动的夹持。这里缺少的是承载飞机所需的抓握接触，未观察到先拿稳再掉落的过程。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": null,
+    "log_step": 161,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 161 的提交生成脚本以参考手关节位置误差和关节速度计算力矩，再用非负最小二乘分配肌肉激励；脚本未依据飞机运动或接触状态调整夹持。视频 0.80–1.48 s 显示执行时手指只有上侧压触，未使飞机离桌，log step 171 又确认该提交可确定性重现同一失败，支持将未形成有效夹持归于控制；不足以确定是具体接触布局、肌力分配还是关节跟踪所致。"
+   }
+  ],
+  "gap_zh": "没有状态日志和成功参考视频；现有双视角可判定飞机未被带起，但不能精确恢复每根手指的接触力和关节跟踪误差。评分器 reward 字段及提交文件重放结果未报告 track_pos_err、track_rot_err、track_ok_fraction、track_max，故这四项数值留空；会话中其他测试的数值不作为提交轨迹的指标。控制根因的具体子类仍不确定。"
+ },
+ "summary_zh": "观察：逐帧检查视频 0.00–1.48 s，MyoHand 手指在飞机上侧屈曲、压触，但飞机始终留在桌面，“飞机循迹飞行”未完成，最终进度为 0/1。推断：尝试在建立可带起飞机的夹持时失败，归为控制问题；log step 171 确认已提交的 37 步轨迹两次确定性重放均提前失败。无状态日志、成功示范或提交轨迹的精确循迹指标，尚不能细分控制原因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点、没建立接触、控制根因均与视频和log相符；未发现需改动的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:59:18Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--wnelj7s/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-privileged--wnelj7s/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "4a51487f7f963963b6309ccb6b306da5c15c444e8d549f804f3d07ad7a1ed19c",
+   "video_sha256": "596f6c1af80333ffe478b3e75ff6973023a5d07761cdc37c5315bc6d7561fad6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-standard--bnj3e9h.json

@@ -0,0 +1,221 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-standard--bnj3e9h",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 38,
+   "sim_steps": 38,
+   "physics_steps": 380,
+   "steps": 85,
+   "calls": 8,
+   "episodes": 3,
+   "resets": 2,
+   "reset_budget": 10,
+   "sim_steps_total": 85,
+   "elapsed_s": 2207.0,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.088,
+   "peak_processes": 17,
+   "peak_disk_gb": 0.444,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-standard--bnj3e9h/replay.mp4",
+  "duration_s": 1.56,
+  "fps": 25.0,
+  "frames": 39,
+  "sha256": "1b24d42ff48ffdef8c204c1e3cb0fd11aa42df6ebf5ef4ab68f8931aca546135"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.8,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频0.00–1.52 s的全部39帧；单只MyoHand的拇指与弯曲手指围住白色飞机，飞机随手移动并转向，因而本步确实开始尝试。"
+     },
+     {
+      "t": 1.4,
+      "log_step": null,
+      "text_zh": "观察：视频1.28–1.52 s，绿色参考标记继续向画面左下方移走，手及所带飞机未同步跟上，末段明显分离；未看到完整循迹完成。"
+     },
+     {
+      "t": null,
+      "log_step": 63,
+      "text_zh": "观察：log step 63的服务返回episode=3、control_steps=38、done=True，原因为飞机距参考或手腕超过0.25米；inputs/run.json评分字段n_actions=38、sim_steps=38、success=0。log step 58属于此前的episode=2，不用于本局进度。"
+     },
+     {
+      "t": null,
+      "log_step": 63,
+      "text_zh": "推断：最后评分局仅执行38个控制步即原生终止，未满足DAG要求的完整99步，因此该整段运动事件始终未完成；这不是仅由零分推断未尝试。视频锚点直接取所查帧，不将agent工作时间或20毫秒控制步直接当作视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频0.00–1.28 s，MyoHand单手拇指与弯曲手指从相对两侧围夹飞机，随手移动及转向；并非双手协作。视频局部分辨率和遮挡不足以逐指确认整个末段的接触维持情况。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better"
+  },
+  {
+   "name": "track_rot_err",
+   "value": null,
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  },
+  {
+   "name": "track_ok_fraction",
+   "value": null,
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better"
+  },
+  {
+   "name": "track_max",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 1.28,
+   "log_step": 63,
+   "where": "操作中",
+   "evidence_zh": "观察：视频1.28–1.52 s，机器人仍在带动飞机尝试循迹，绿色参考标记与手中飞机逐渐拉开，未恢复跟随；log step 63确认最后评分局episode=3在control_steps=38因飞机距参考或手腕超过0.25米而终止。推断：失败发生在飞机循迹飞行的操作中；视频1.28 s为明显失配的粗略导航锚点，不是精确阈值越界时刻。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 1.28,
+    "log_step": 63,
+    "evidence_zh": "观察：视频1.12–1.28 s绿色参考标记仍在飞机附近，视频1.36–1.52 s标记继续移向左下而机器人手及飞机的运动未同步，间隔明显扩大；log step 63服务返回距离超限终止。���断：这是对随时间变化的参考运动未跟上，表现为动态时机失配。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "观察：log step 47生成的执行脚本将离线优化的整批执行器指令以dt=0.02提交给服务，没有依据执行中的物体误差进行修正；log step 63实际执行这一路径并在38个控制步后失败。log step 51的离线抓持测试中，后段物体升高约0.048米、腕升高约0.110米且相对腕位置偏移约0.043米；这些是会话诊断而非评分局状态。推断：结合视频1.28–1.52 s的跟随失配，所交付的肌肉与前臂控制未能可靠带动飞机跟上参考；证据不能分清抓持与轨迹跟踪的贡献，根因停在控制层。"
+   }
+  ],
+  "gap_zh": "无状态日志、成功示范或评分局逐控制步的未舍入误差；评分字段未报告track_pos_err、track_rot_err、track_ok_fraction、track_max，四项指标均填null。双视角存在过曝、遮挡及相机运动，可辨认尝试与后段跟随分离，但不能精确确定失配阈值的首次越界、逐指接触或原生终止究竟由参考距离还是腕距触发。agent在log step 65自述抓持滑移及竖直执行器能力不足，不能据此证实环境不可解，也不将其离线诊断当作本局精确状态。"
+ },
+ "summary_zh": "最后评分局中MyoHand抓住并带动飞机移动、转向，但视频后段未跟上参考，完整的“飞机循迹飞行”节点未达成（终态0/1）。log step 63确认该局在38/99个控制步时触发原生距离终止，现象为动态时机失配，根因保守归为控制。缺少状态日志和原生误差数值，无法确定抓持滑移与前臂跟踪各自的影响或精确阈值越界时刻。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标签与证据一致：最后评分局（episode 3，log id 63，38/99 步原生终止）里手抓住并带动飞机，后段未跟上参考；节点未达成、断点、动态时机现象与保守的控制根因均有支撑，时钟未把agent时间转换为视频时间。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:52:45Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--bnj3e9h/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--bnj3e9h/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "a9cbb3654ab5b395b6a87ffb13ac17aaac055a784523d1e0ed8c6582d6c61236",
+   "video_sha256": "1b24d42ff48ffdef8c204c1e3cb0fd11aa42df6ebf5ef4ab68f8931aca546135",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-standard--edde464.json

@@ -0,0 +1,210 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-standard--edde464",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 19,
+   "sim_steps": 19,
+   "physics_steps": 190,
+   "steps": 290,
+   "calls": 81,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 290,
+   "elapsed_s": 1184.7,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.337,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.84,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-standard--edde464/replay.mp4",
+  "duration_s": 0.8,
+  "fps": 25.0,
+  "frames": 20,
+  "sha256": "40a18b6d3b6273d9f66c948cc43020a261cc156f61d044f0f54f21069d8ad894"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频0.00–0.76 s的全部20帧；开局白色飞机在桌面上，MyoHand指端位于机身上方。视频0.16–0.32 s手指卷曲，视频0.56–0.76 s右侧视图中空手与飞机明显分离，未见飞机被夹持并带起完成空中循迹。"
+     },
+     {
+      "t": 0.76,
+      "log_step": 80,
+      "text_zh": "观察：log step 80明确运行最后的episode 11，服务最终状态为control_steps=19、done=true，reason为物体距参考或手腕超过0.25 m；评分器n_actions=19、live_matches_replay=1、success=0。该局未走完节点要求的99个控制步；视频中的空手闭合与此局的抓取失败一致。未将日志仿真0.38 s换算为视频时刻。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频0.00–0.76 s中，MyoHand单手指端在飞机上方闭合后空手移离，未见形成对向夹持。低分辨率及初始指端重叠使瞬时轻触及其方式无法确认，故实际接触记未知，不将抓取指令记为已发生的抓握。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better"
+  },
+  {
+   "name": "track_rot_err",
+   "value": null,
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better"
+  },
+  {
+   "name": "track_ok_fraction",
+   "value": null,
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better"
+  },
+  {
+   "name": "track_max",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.32,
+   "log_step": 80,
+   "where": "操作中",
+   "evidence_zh": "观察：视频0.16–0.32 s，MyoHand手指逐渐闭合，但白色飞机仍在手下方的桌面上；视频0.56–0.76 s手中仍无飞机，未恢复抓取。log step 80是最后一局，服务在control_steps=19时以物体距参考或手腕超过0.25 m终止，该局没有完成飞机循迹飞行。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 0.32,
+    "log_step": 80,
+    "evidence_zh": "观察：视频0.00–0.76 s中，MyoHand手指从飞机上方的张开姿态卷成空握，飞机未随手离开桌面；视频0.32 s已能看到指端与机身间的间隙，后续逐帧至视频0.76 s始终未见包夹。所需的承载抓握没有建立，无证据表明曾抓住后滑落。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log steps 78–79的最终控制器用固定的手指肌肉激励闭合，并按初始手与物体的相对变换计算掌部目标；log step 80中，控制器在尚未抓起飞机时继续执行掌部运动，例如仿真0.20 s的q前两项为[-0.042,-0.076]而q_cmd为[-0.158,-0.038]。结合视频0.16–0.76 s手指在空中闭合并与飞机分离，说明抓取及手部运动控制未实现所需接合；不足以进一步区分接触设计、逆解或动态跟踪，故停在控制层级。"
+   }
+  ],
+  "gap_zh": "观察：inputs/提供完整视频和运行日志，但无状态日志、成功示范或评分器的track_pos_err、track_rot_err、track_ok_fraction、track_max数值，四项原生指标均留null。初始指端重叠且视频分辨率低，无法确认短暂轻触或具体接触方式；能确认未建立承载抓握。视频20帧共0.80 s，而log step 80报告19个20毫秒控制步、仿真0.38 s；没有经核对的两者时间映射，所有导航时刻直接取视频帧，根因不细分，脚本打印的plane估计不作为飞机真实位姿。"
+ },
+ "summary_zh": "最后评分局中，MyoHand在飞机上方闭合手指后空手移离，未抓起飞机，飞机循迹飞行节点未完成（0/1）。断点在该步操作中，现象为没建立接触；最终控制器的抓取及掌部运动未实现所需接合，根因保留为控制，服务在19个控制步后原生终止。无状态日志、成功示范及精确循迹指标，初始轻触方式和控制子类无法确定，视频与日志仿真时长也未建立映射。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点、没建立接触与控制根因均与视频帧一致；时间均取视频帧，未换算日志时间。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T04:11:56Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--edde464/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--edde464/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "88e70d74efea6cb6370baf4ed068c2c9214d5b27dfad16fe4fc6180021f1e029",
+   "video_sha256": "40a18b6d3b6273d9f66c948cc43020a261cc156f61d044f0f54f21069d8ad894",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-standard--emurq9d.json

@@ -0,0 +1,219 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-standard--emurq9d",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 37,
+   "sim_steps": 37,
+   "physics_steps": 370,
+   "steps": 214,
+   "calls": 196,
+   "episodes": 6,
+   "resets": 5,
+   "reset_budget": 10,
+   "sim_steps_total": 214,
+   "elapsed_s": 3566.1,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.253,
+   "peak_processes": 15,
+   "peak_disk_gb": 1.761,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-standard--emurq9d/replay.mp4",
+  "duration_s": 1.52,
+  "fps": 25.0,
+  "frames": 38,
+  "sha256": "6ee0c58dcf0fe0cbfca294b5ce283004044e647bde02cde5ccc2b8377b051bba"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 两个视角显示玩具飞机在桌面，MyoHand位于飞机一侧上方；初态不代表整段循迹飞行已完成。"
+     },
+     {
+      "t": 0.8,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–1.48 s；MyoHand手指逐渐收拢，视频 0.72–0.96 s 的闭合手指仍在机身一侧上方，未包住机身；视频 1.12–1.48 s 手指贴近机翼，但飞机仍贴着桌面，没有被手带起飞行。"
+     },
+     {
+      "t": 1.48,
+      "log_step": 239,
+      "text_zh": "观察：视频 1.48 s 飞机仍在桌面；log step 236 重新运行修正后的控制器，log step 237 返回执行37个控制步后原生终止，原因是物体距参考或手腕超过0.25 m，log step 239 的服务状态确认episode=6、control_steps=37、done=True。评分器reward.graded_by_live=1、sim_steps=37，判定最后一局失败；未走完所需99步。日志中的控制步未换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "推断：视频 1.12–1.48 s，MyoHand单手收拢的手指贴在飞机一侧机翼上方，按贴压归类；飞机保持桌面支承，未见对向夹住机身或从下方托起，不记抓握。无接触对或力传感记录，具体触点无法精确确认。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.8,
+   "log_step": 237,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 0.80 s，MyoHand手指已经收拢在飞机一侧上方，飞机仍留在桌面；随后视频 1.12–1.48 s 也未被抓起。log step 237 记录最后一局在37个控制步后因物体距参考或手腕超过0.25 m终止，log step 239 确认这是第6局最终服务状态；飞机循迹飞行未完成，前面的抓取失败始终未修复。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 0.8,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.72–0.96 s，MyoHand手指在飞机机身一侧上方闭合，飞机没有进入手指围成的夹持区域；视频 1.12–1.48 s 虽有贴压机翼附近的动作，飞机始终未随手离开桌面。推断：循迹所需的承载抓握未建立，不是先抓起后滑落。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "t": 0.8,
+    "log_step": 236,
+    "evidence_zh": "推断：视频 0.72–1.48 s 显示手指在飞机一侧上方闭合，后段贴压机翼附近，未在机身两侧形成可承载的夹持。log step 231 的控制代码先调整前臂、再以同组屈指肌逐渐收拢，随后直接切换参考轨迹；log step 236 仅删除越界索引后重置并执行此策略，log step 237 记录正常执行至原生物体距离终止。结合可见的空夹和飞机未抬起，所用接触布局未抓住飞机；没有在本局验证成功的替代握法。",
+    "source_zh": "视频 / 日志"
+   }
+  ],
+  "gap_zh": "无状态日志、接触对或接触力记录，贴压方式由可见接近与闭合几何推断，不能精确确定各指触点。没有成功示范；评分器reward未报告track_pos_err、track_rot_err、track_ok_fraction、track_max，四项原生指标均留null，不能从视频量出精确误差。agent在log step 240 将问题归于物体位姿估计，但现有证据不足以独立确认错误估计如何导致本局动作，故不另标感知根因。"
+ },
+ "summary_zh": "评分的第6局中，MyoHand收指并在飞机一侧贴压，但未建立能带起飞机的抓握，飞机始终留在桌面，飞机循迹飞行完成0/1。断点位于抓取过程，推断闭合位置与接触布局不适配；服务在37个控制步后因物体距参考或手腕超过0.25 m终止。无状态日志和精确循迹指标，具体触点及替代握法可行性仍不确定。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "视频逐帧显示手指在飞机一侧收拢、飞机始终留在桌面、绿色参考目标逐渐远离，节点未达成、断点与现象（没建立接触）、根因均有证据支持，未发现需改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:52:40Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--emurq9d/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--emurq9d/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f4a69ef5ec16d4a1bc9c262361a3269eb073202c6220e117a86c6f08d423a77d",
+   "video_sha256": "6ee0c58dcf0fe0cbfca294b5ce283004044e647bde02cde5ccc2b8377b051bba",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-standard--erawpc2.json

@@ -0,0 +1,215 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-standard--erawpc2",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 11,
+   "sim_steps": 11,
+   "physics_steps": 110,
+   "steps": 11,
+   "calls": 2,
+   "episodes": 1,
+   "resets": 0,
+   "reset_budget": 10,
+   "sim_steps_total": 11,
+   "elapsed_s": 3590.8,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.702,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.327,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-standard--erawpc2/replay.mp4",
+  "duration_s": 0.48,
+  "fps": 25.0,
+  "frames": 12,
+  "sha256": "8297f8d0a02e2bfdc3e040199a5f2ff4001a875c8f5c9cc6a3a73427d408c7f7"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s，MyoHand单手位于桌面附近的玩具飞机上方，飞机尚未被带起。"
+     },
+     {
+      "t": 0.28,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频 0.00–0.44 s 的全部12帧；手指屈伸靠近飞机，视频 0.24–0.44 s 左侧视图中飞机向手的左侧移开，末帧仍未见飞机被手稳定抓持并带起飞行。"
+     },
+     {
+      "t": 0.44,
+      "log_step": 153,
+      "text_zh": "观察：log step 153 中，机器人服务在提交前报告control_steps=0；提交99行执行器命令后只执行11步，返回done=True及“failure: object more than 0.25 m from the reference (or from the wrist)”。评分器字段n_actions=11、task_failure=1、live_matches_replay=1；评分局未走完节点所需的99步。此处视频时间仅定位末帧，不把服务的仿真 0.22 s 换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "推断：视频 0.12–0.44 s，MyoHand单手手指在飞机上方及侧缘屈伸，飞机随后向手外侧移开，表现为手指拨动飞机；未见相向夹持后稳定带起，故不记抓握。具体接触手指与短暂接触的先后无法从低分辨率画面确定。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.24,
+   "log_step": 153,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.24–0.44 s，机器人手指在飞机附近屈伸，而飞机向手外侧移开，没有被稳定抓持并带起。log step 153 的机器人服务在执行11步后因飞机距参考或手腕过远而终止；该抓取及循迹尝试没有后续修复。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 0.24,
+    "log_step": 153,
+    "evidence_zh": "推断：视频 0.00–0.44 s 的逐帧画面中，机器人手指虽靠近并拨动飞机，但未形成能让飞机随手共同运动的稳定抓握；飞机向手外侧移开，未见先被持稳再掉落。结合log step 153 服务在11步后以飞机距参考或手腕过远终止，判断抬起所需的抓握没有建立；这里指所需抓握未建立，并非完全没有物理触碰。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 153 的agent脚本在本地MuJoCo中计算99行肌肉及前臂执行器命令，再一次性提交机器人服务；同一步的本地测试已显示控制步30飞机质心高度0.0182米、参考高度0.0835米，控制步50仍为0.0182米而参考为0.4467米，未实现有效带起。实际评分视频 0.00–0.44 s 同样未形成稳定抓持，服务只执行11步即终止，支持执行器控制未实现抓取这一宽泛归因。本地测试仅作为根因旁证，不用于评分节点进度；现有证据不能区分接触设计、模型差异或跟踪控制的具体责任。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有状态日志或成功示范；视频分辨率与局部遮挡限制了具体接触手指、接触力及瞬时夹持的判定，推/扫是依据手指动作与飞机移开的上下文推断。log step 153 服务返回的仿真时长为0.22秒，视频时长为0.48秒，缺少可核对的逐帧控制步映射，未按控制步换算视频时间。评分器reward未报告track_pos_err、track_rot_err、track_ok_fraction、track_max，服务输出亦有截断，四项数值均保留null；不能用agent本地仿真的误差替代实际评分指标，也不能进一步确认控制根因子类。"
+ },
+ "summary_zh": "评分尝试仅执行11个控制步，飞机循迹飞行节点未完成，最终进度0/1。视频 0.00–0.44 s 显示手指未建立能稳定带起飞机的抓握，飞机向手外侧移开；log step 153 的机器人服务因飞机距参考或手腕过远而终止，根因证据支持宽泛的控制问题。缺少状态日志、精确循迹指标及仿真到视频的可靠时间映射，无法细分接触和控制机制。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点未达成、最终0/1、断点在操作中均与视频和评分器字段一致：视频0.24–0.44 s飞机被手指拨开并飞离，服务在11步后因飞机距参考或手腕超过0.25 m终止；没建立接触符合定义，控制根因标为宽泛归因且已声明本地仿真只作旁证。未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:54:42Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--erawpc2/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--erawpc2/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "82e4bf0b3154e315b00831763e77229ec3771bb644992725811546e69969c453",
+   "video_sha256": "8297f8d0a02e2bfdc3e040199a5f2ff4001a875c8f5c9cc6a3a73427d408c7f7",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-standard--hyul5i7.json

@@ -0,0 +1,221 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-standard--hyul5i7",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 20,
+   "sim_steps": 20,
+   "physics_steps": 200,
+   "steps": 293,
+   "calls": 113,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 293,
+   "elapsed_s": 2767.1,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 2.164,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.523,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-standard--hyul5i7/replay.mp4",
+  "duration_s": 0.84,
+  "fps": 25.0,
+  "frames": 21,
+  "sha256": "46e296878c49b02d641a2eab8dd8664a121f209f495d076d41c20a0359b5157e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 两个视角显示玩具飞机在桌面附近，MyoHand手指位于飞机上方；右侧近景未见手指围住机身。"
+     },
+     {
+      "t": 0.16,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–0.80 s 的全部21帧，视频 0.16 s 起手指与飞机的间隙增大，随后手部上移；飞机留在桌面附近，至视频 0.80 s 仍未随手抬起或进行空中循迹。"
+     },
+     {
+      "t": null,
+      "log_step": 368,
+      "text_zh": "观察：log step 368 中agent执行最终脚本，服务从第10局重置为第11局，接收99步提交但仅执行20个控制步即done=True，原因为飞机距参考或手腕超过0.25米；log step 373确认同一终态。评分字段graded_by_live=1、n_actions=20、success=0。"
+     },
+     {
+      "t": null,
+      "log_step": 368,
+      "text_zh": "推断：结合视频中飞机始终未被带起和log step 368的第20控制步提前终止，本局没有完成节点要求的99步参考运动；不能把agent在log step 375自述的运动学误差达标当作飞机实际循迹成功。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "未知",
+      "note_zh": "观察：视频 0.00–0.80 s 中MyoHand单手在飞机上方屈指并上移，未见对向夹紧机身或带起飞机。开头指尖与飞机尾部局部投影接近，缺少接触对记录，不能确认是否有瞬时碰触及其方式；不据脚本的抓握注释认定实际抓握。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.16,
+   "log_step": 368,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.16 s 手部开始明显向上离开飞机，视频 0.32–0.80 s 飞机仍在桌面附近，没有随手升起；log step 368显示最终第11局在第20控制步触发原生距离终止，未完成99步运动。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 0.16,
+    "log_step": 368,
+    "evidence_zh": "观察：视频 0.00–0.16 s 未见手指合拢夹住飞机；视频 0.16–0.80 s 手部屈指上移，飞机留在桌面附近。推断：循迹所需的承载抓握未建立，未见已经持有飞机后再滑脱的过程；log step 368的原生距离终止与手、飞机分离的画面一致。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "观察：log step 363的最终脚本将target_q[:6]-q0[:6]写入六个根执行器命令，并对指定肌肉使用整段固定激活，一次提交99步；log step 12说明这六个执行器接收关节位置目标。log step 368确认执行该脚本的第11局仅运行20步即终止。推断：结合视频 0.16–0.80 s 的空手上移与飞机留在桌面附近，这组控制命令未能在移动前建立并维持承载夹持；现有证据不能区分接触几何、肌肉激活和根执行器目标换算的各自贡献，故停在控制大类。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志、成功示范或最终第11局的逐步位置误差、转动误差、达标比例与最大误差，四项原生指标均未报告；开头指尖与飞机局部投影接近，无法确认瞬时碰触，但完整视频可确认没有抓住并带起飞机。log step 368记录20个控制步、命令dt=0.02 s，对应仿真时长0.40 s，而视频时长0.84 s，未建立两者映射，所有视频锚点均直接来自已检查的帧。"
+ },
+ "summary_zh": "观察：最终评分为log step 368重置后的第11局，视频中MyoHand手部屈指后上移，飞机仍留在桌面附近，飞机循迹飞行节点未完成，终态0/1。推断：失败断在操作中的承载抓握未建立，归入控制；服务在第20控制步因飞机距参考或手腕超过0.25米终止，而非执行完99步。缺少状态日志及逐步循迹误差，不能细分控制根因，也不能确认开头是否有短暂碰触。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "逐帧核对视频：手指上移，飞机始终留在桌面，节点未达成、断点与现象与证据一致，时间均为视频钟，未转换agent时间。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:57:08Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--hyul5i7/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--hyul5i7/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "32a2f7dccc4d29e26635f37643a9e8e11d9e4b83561637eb5275b25cddbe4d57",
+   "video_sha256": "46e296878c49b02d641a2eab8dd8664a121f209f495d076d41c20a0359b5157e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-standard--yknseap.json

@@ -0,0 +1,217 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-standard--yknseap",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 37,
+   "sim_steps": 37,
+   "physics_steps": 370,
+   "steps": 392,
+   "calls": 421,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 392,
+   "elapsed_s": 3485.6,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.668,
+   "peak_processes": 17,
+   "peak_disk_gb": 1.387,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-standard--yknseap/replay.mp4",
+  "duration_s": 1.52,
+  "fps": 25.0,
+  "frames": 38,
+  "sha256": "d0849eaa8c3d3497cebeaba27722f681609b0ca639f4c21369a07db9b730aff0"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.48,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频 0.00–1.48 s 的全部38帧；机器人单手弯曲手指并移动，右侧视角中飞机仍在桌面，视频 0.48–0.96 s 手与飞机之间的间隔增大，未见手指夹持飞机并将其带起。"
+     },
+     {
+      "t": 1.48,
+      "log_step": 143,
+      "text_zh": "观察：视频 1.20–1.48 s 飞机仍在桌面，手指在其上方空置；log step 143 的机器人服务状态为 episode=11、control_steps=37、done=true，原因为飞机距参考或手腕超过0.25米。该局未走完99个控制步，与视频中未完成抬起和循迹一致；这里只并列两种证据，不把控制步或agent工作时间换成视频时间。"
+     },
+     {
+      "t": null,
+      "log_step": 137,
+      "text_zh": "观察：log step 137 调用 final_run(stop=24, reset=True) 启动最后一局；log step 140 的服务状态确认 episode=11、resets_left=0；log step 142 继续该局，log step 144 再次确认其在control_steps=37失败。inputs/run.json 的reward记录 graded_by_live=1、n_actions=37、episodes=11，故此前重置掉的试验不计为本次节点进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "未知",
+      "note_zh": "观察：视频 0.00–1.48 s，MyoHand单手在飞机尾部附近弯曲手指、尝试捏取，未见包围或夹住飞机，不能记为已发生的抓握。开局指尖与尾部投影接近，画面分辨率不足以排除短暂擦碰，且无接触对日志，因此实际接触方式记未知。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.6,
+   "log_step": 141,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 0.48–0.96 s，机器人手指弯曲后随前臂移动，飞机仍留在桌面，没有形成可带起飞机的夹持；到视频 1.48 s 仍未修复。log step 141 检查的最终局控制步24显示实际关节位置偏离计划；log step 143 的服务状态随后在控制步37以飞机距参考或手腕超过0.25米终止。断点取正在执行的飞机循迹飞行，视频 0.6 s 仅作抓取未建立的粗略导航锚点。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 0.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.00–1.48 s，机器人单手在飞机附近弯曲手���，但飞机没有随手离开桌面；视频 0.60、0.96、1.20及1.48 s 的右侧视角显示手指与机体分离，未建立抬起所需的夹持接触。未见曾抓住飞机后再掉落的过程。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 141,
+    "evidence_zh": "推断：最终局的执行未把手送到计划的捏取位置。log step 141 对比该局控制步24的计划与本体感觉关节位置，第二个平移分量计划为-0.03078米、实际为-0.08147米；该脚本用手模型计算的实际拇指/食指尖高度约0.0969/0.1099米，计划约0.0626/0.0656米。这些模型计算不是物体状态真值，但与视频 0.48–1.48 s 手指空置、飞机留桌一致。log step 142 的控制脚本随后尝试修正捏点，log step 143 仍显示未跟上其目标。证据支持控制执行偏差，尚不足以确定是逆解、反馈结构或具体参数所致。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "无模拟器状态日志或接触对记录；开局指尖与机尾邻近处不足以确认是否有短暂擦碰，因此contact_actual的接触方式为未知，但全片足以确认没有建立有效夹持。评分器reward未报告track_pos_err、track_rot_err、track_ok_fraction或track_max的原始值，四项原生指标均置null；无成功示范，不能进一步验证具体控制子类或替代抓法的可行性。视频时钟直接按25 fps取帧，未把50 Hz控制步或agent工作时间直接映射为视频时间。"
+ },
+ "summary_zh": "最终评分局中，机器人单手尝试捏取飞机，但飞机始终留在桌面，没有完成抬起及参考循迹，节点终值为0/1。断点为飞机循迹飞行操作中未建立有效夹持；最终局日志显示控制执行偏离计划，服务在控制步37因飞机距参考或手腕超限而终止。无状态接触记录和原始循迹误差，短暂擦碰及更细控制根因无法确认。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未达成、断点、没建立接触与控制根因均与视频（1.5 s 内飞机始终在桌面、手指空置）及最终局日志一致；时钟未换算，未发现需修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:59:10Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--yknseap/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--yknseap/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "6562e3673eed1adfe66b6e9b6273768430375d6b1bccad4f1ffd7cd3b35ae97c",
+   "video_sha256": "d0849eaa8c3d3497cebeaba27722f681609b0ca639f4c21369a07db9b730aff0",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/myosuite/dex_airplane_fly/trajectories/myosuite-dex-airplane-fly-i00-standard--yyrojfo.json

@@ -0,0 +1,216 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "myosuite-dex-airplane-fly-i00-standard--yyrojfo",
+ "taskKey": "myosuite/dex_airplane_fly",
+ "dag_sha256": "211c82c127093021391297ba04609fe900a6c8513c288e6a7fb1169216a9dbdd",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "n_actions": 35,
+   "sim_steps": 35,
+   "physics_steps": 350,
+   "steps": 361,
+   "calls": 27,
+   "episodes": 11,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_steps_total": 368,
+   "elapsed_s": 1336.9,
+   "video_rendered": 1,
+   "task_failure": 1,
+   "timeout": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.538,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.581,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/myosuite-dex-airplane-fly-i00-standard--yyrojfo/replay.mp4",
+  "duration_s": 1.44,
+  "fps": 25.0,
+  "frames": 36,
+  "sha256": "f8ef8c2ed2468b5ef53c368587dc41f2e2061e795d2413c58fbd044d7232b121"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "airplane_reference_motion_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 的右侧外部视角显示飞机在桌面，MyoHand手指位于飞机上方；逐帧检查视频 0.00–1.40 s，飞机未随手升起。"
+     },
+     {
+      "t": 0.6,
+      "log_step": null,
+      "text_zh": "观察：视频 0.32–0.60 s 手指逐渐收拢，手与前臂向上离开飞机；视频 0.84–1.40 s 右侧视角中手已明显空悬，飞机仍在桌面，未出现持机飞行或转向。"
+     },
+     {
+      "t": 1.4,
+      "log_step": 73,
+      "text_zh": "观察：log step 73 最后一次reset进入episode 11、resets_left=0，服务返回control_steps=35、done=true，原因为物体距参考或手腕超过0.25米；run.json评分字段sim_steps=35、task_failure=1。该局没有走完要求的99个控制步，且视频末帧仍未带起飞机。这里不将控制步直接换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "未知",
+      "note_zh": "观察：视频 0.00–0.60 s，MyoHand单手手指在飞机上方及尾部附近收拢，随后空手抬离；局部手指与机体重叠且没有接触日志，无法确认短暂触碰的具体方式，不能据闭指动作认定已抓握飞机。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "track_pos_err",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_rot_err",
+   "unit": "rad",
+   "threshold": 0.5,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "track_ok_fraction",
+   "unit": "ratio",
+   "threshold": 0.9,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "track_max",
+   "unit": "m",
+   "threshold": 0.05,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "airplane_reference_motion_completed",
+   "t": 0.6,
+   "log_step": 73,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 0.60 s，MyoHand已收拢手指并开始与桌面上的飞机分离，后续视频 0.84–1.40 s 空手悬在飞机上方，未修复抓取。log step 73 记录评分的最后一局在35个控制步时原生失败终止，尚未完成99步循迹运动。"
+  },
+  "unmet": {
+   "airplane_reference_motion_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 0.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.32–0.60 s 手指闭合后向上离开，飞机没有被夹持并带离桌面；视频 0.84–1.40 s 手和飞机之间有清晰空隙，飞机仍留在桌面。推断：飞行所需的有效抓持未建立，不能把靠近或短暂擦碰视为先抓住再滑落。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "观察：log step 73 的评分局脚本在控制步索引7–14递增FDS、FDP、FPL、OP屈指肌命令，索引15以后保持；前28行维持预抓取基部目标，随后执行事先计算的IK轨迹，提交中没有依据抓持观测修正动作。视频 0.32–1.40 s 显示闭指和抬手未能带起飞机。推断：这套肌肉与前臂动作未实现有效抓持，属控制失败；现有证据不足以区分具体接触几何、肌肉参数或跟踪误差。"
+   }
+  ],
+  "gap_zh": "无状态日志或接触对，视频初段的手指与飞机局部重叠，无法确定短暂触碰的方式，contact_actual的how保留未知；run.json评分字段未报告track_pos_err、track_rot_err、track_ok_fraction、track_max，四项数值均留null。无成功示范。agent在log steps 78–79自述飞机已被抓起并携带，但评分视频 0.00–1.40 s 中飞机仍留在桌面，节点和现象以视频为准；不采纳该自述中的抓起高度或可达性判断。"
+ },
+ "summary_zh": "评分的第11局中，MyoHand在飞机附近闭指并抬手，但飞机始终留在桌面，飞机循迹飞行节点未完成，最终为0/1。断点在抓持并带起飞机的操作中：有效抓持未建立，服务在35个控制步时因物体距参考或手腕超过0.25米提前终止。视频与agent自述的“已抓起携带”不符；缺少接触日志和精确循迹指标，接触方式及控制子类型仍不能确定。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点未达成、断点在抓持并带起飞机的操作中、现象没建立接触、根因停在控制，均与视频及 log step 73 的服务读数一致；控制步未换算为视频时间，时钟合规。补充观察：视频 0.6–1.4 s 右侧视角中飞机在桌面上被推动并转向，但仍在桌面，未被带起，不改变标签结论。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:57:12Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--yyrojfo/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/myosuite-dex-airplane-fly-i00-standard--yyrojfo/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "ebe24fef539d8494fe64e55f7067be428013c9e60347fd3ccdcfb8b29a9bd3b8",
+   "video_sha256": "f8ef8c2ed2468b5ef53c368587dc41f2e2061e795d2413c58fbd044d7232b121",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}