Add 114 trajectories (phase 2, remaining runs; batch S35)

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-privileged--5pq9urq.json

@@ -0,0 +1,238 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-privileged--5pq9urq",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 419,
+   "submitted_actions": 444,
+   "sim_steps": 419,
+   "physics_steps": 4184,
+   "sim_time_s": 10.46000000000008,
+   "final_reward": 1.0,
+   "replay_wall_s": 435.1,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 419,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 1,
+   "agent_resets": 6,
+   "agent_sim_steps": 0,
+   "agent_records": 2,
+   "agent_replays": 1,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-blanket-flatten-08-i01-privileged--5pq9urq/replay.mp4",
+  "duration_s": 10.533,
+  "fps": 15.0,
+  "frames": 158,
+  "sha256": "08b10779d2d47ddd135b8b0db48d8fbc87608ddd2bb5fb496708cc87d237e2ab"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.47,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，桌面蓝色方格毯的中部有横贯毯面的翻折重叠层，尚未展平。"
+     },
+     {
+      "t": 3.6,
+      "log_step": null,
+      "text_zh": "观察：视频 1.67–2.40 s，机器人左夹爪夹住毯子左侧局部并拉动，布片随夹爪移动；视频 3.20–4.33 s，左夹爪再次拉展左侧布片，中部重叠逐渐展开，右夹爪停在毯外。"
+     },
+     {
+      "t": 8.73,
+      "log_step": null,
+      "text_zh": "观察：视频 6.20–7.13 s，机器人右夹爪在毯子右侧夹取并移动布片；视频 7.60–9.27 s，右夹爪再次夹取、抬起并铺回右侧折角，布边随夹爪抬升后落下，左夹爪保持在毯外上方。"
+     },
+     {
+      "t": 10.47,
+      "log_step": null,
+      "text_zh": "观察：视频 9.33–10.47 s，原先横向重叠层已消失，毯子呈近矩形并贴在桌面；右夹爪张开后向上撤离，末段两爪均与毯子分离，毯面没有明显继续滑动。inputs/run.json 的评分器 reward 报告 success=1、all_success=1、replays=2、strict_goal_reached=1、replay_hold_steps=0、replay_tolerance_needed=0。"
+     },
+     {
+      "t": 10.47,
+      "log_step": 144,
+      "text_zh": "推断：结合视频 10.47 s 终态与评分器严格成功字段，将包含释放、撤离及连续确认的整体展平节点约定在视频末尾约 10.47 s 达成；毫米级清距、粒子速度及材料伸长不能从画面独立量测，精确达成时刻不确定。log step 144 记录 agent 保存了419行原动作并补至444行，评分器实际使用419行动作；此处未将 agent 工作时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：机器人左夹爪在视频约 1.67–4.33 s 先后夹住毯子左侧布片并拉展；右夹爪在视频约 6.20–9.27 s 夹取右侧布片、抬起并铺回折角。两爪依次操作，单爪夹持时另一爪在毯外。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人先用左夹爪拉展毯子左侧与中部，再用右夹爪展开右侧折角，最终完成1/1个节点。视频末段毯子平铺、两爪松开并撤离，结合评分器两次回放均严格成功且无需补持或宽松容差，将整体节点达成时间估计为视频约 10.47 s。无状态日志，评分器 reward 未直接报告八项原生指标数值，毫米级几何及速度条件依赖严格成功字段确认，精确达成时刻无法独立复核。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "成功运行，单节点 blanket_flattened 在视频约 10.47 s 达成，与严格成功字段（strict_goal_row=419、sim_time 10.46 s）及末段画面一致；接触标注（单爪依次抓握）与画面相符，无失败层。",
+    "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": "2026-10-30T23:48:36Z",
+  "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/roboarc-blanket-flatten-08-i01-privileged--5pq9urq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--5pq9urq/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "54f909f91e49ea451b92291f8cac81b0e9afda9910bb738541fc5b9e686530cb",
+   "video_sha256": "08b10779d2d47ddd135b8b0db48d8fbc87608ddd2bb5fb496708cc87d237e2ab",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-privileged--7ppv6e8.json

@@ -0,0 +1,243 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-privileged--7ppv6e8",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 2839,
+   "submitted_actions": 2919,
+   "sim_steps": 2839,
+   "physics_steps": 28387,
+   "sim_time_s": 70.967499999999,
+   "final_reward": 1.0,
+   "replay_wall_s": 3012.8,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 2839,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 1,
+   "agent_resets": 4,
+   "agent_sim_steps": 0,
+   "agent_records": 1,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-blanket-flatten-08-i01-privileged--7ppv6e8/replay.mp4",
+  "duration_s": 71.067,
+  "fps": 15.0,
+  "frames": 1066,
+  "sha256": "858a9213af5a7f2a4b431b0d6b97fd4af795ea04c10b2f6659132e3bb3138ac1"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [
+     [
+      70.9,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 蓝色方格毯位于桌面，横贯毯面的翻折重叠层清楚可见，初态尚未展平。"
+     },
+     {
+      "t": 29.333,
+      "log_step": null,
+      "text_zh": "观察：视频 28.0–32.7 s 左右夹爪分别夹住毯子两侧的局部布片并配合向外拉，布边随夹爪移动，投影轮廓逐渐展开；中央仍有明显折脊，不能记为完成。"
+     },
+     {
+      "t": 62.333,
+      "log_step": null,
+      "text_zh": "观察：视频 60.0–63.7 s 机器人放低上边缘、先后松手撤离，毯面已大幅展开，但靠左上部仍可见局部褶皱；视频 67.0–69.5 s 左夹爪再次夹住该区域，布面随其下放而展开，随后左爪张开并上撤。"
+     },
+     {
+      "t": 71.0,
+      "log_step": null,
+      "text_zh": "观察：视频 69.6–71.0 s 双爪离开毯子并抬高，开局横向重叠和后续斜向折脊消失；末帧毯子近矩形平铺桌面，侧视没有明显悬空，末段未见继续滑移。"
+     },
+     {
+      "t": 70.9,
+      "log_step": 289,
+      "text_zh": "推断：结合视频 70.4–71.0 s 的平铺、撤离终态与 inputs/run.json 的评分器字段 success=1、all_success=1、strict_goal_reached=1、replay_hold_steps=0、replay_tolerance_needed=0，唯一目标最终成立；将包含连续确认的完成时刻估计在视频 70.9 s 附近，误差约0.5秒，不把此前仅外观展开记为完成。log step 257、289 记录交付动作表的两次 grading replay 均严格成功、同为2839个控制步且最终 state_sha 相同，支持这是交付回放的结果。"
+     },
+     {
+      "t": 71.0,
+      "log_step": 289,
+      "text_zh": "观察：log step 289 在第二次交付回放完成后输出任务 evaluate_state 的终态指标：coverage_ratio=0.9993868909118968、maximum_height_m=0.01138058155775068、maximum_stretch_ratio=1.002861550440994、maximum_speed_m_s=0.0005700039892057798。log step 9 的 report 定义表明这些值由任务评估函数读取，非 agent 的口头估计；未报告的最小高度、释放清距、最大漂移和确认时长数值保持 null。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 28.0–32.7 s 两个二指夹爪各自夹住毯子不同布片并配合向外拉展，不是两臂从外侧挤住毯子；后续换握继续展开。视频 67.0–69.5 s 改由左夹爪单独夹住残余折脊附近布片、放低修整，右爪已撤离，随后左爪松开上撤。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": 0.9993868909118968
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": 0.01138058155775068
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": 1.002861550440994
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": 0.0005700039892057798
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人通过双爪换握、配合拉展和左爪最后修整折脊，将毯子铺成近矩形后松手撤离，唯一节点“毯子展平”在视频约70.9 s完成，最终1/1。评分器两次回放均严格成功，无额外补时或阈值放宽，末段画面与参考成功视频的平铺撤离外观一致。没有状态日志，毫米级联合条件和准确完成时刻不能独立逐帧复算；四项连续指标来自 log step 289 对交付第二次回放终态的任务评估输出，其余四项数值未报告。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "成功运行，唯一节点毯子展平末帧近矩形平铺、双爪撤离，评分器字段支持成功；时间与接触标注未发现错误。",
+    "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": "2026-10-30T23:57:24Z",
+  "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/roboarc-blanket-flatten-08-i01-privileged--7ppv6e8/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--7ppv6e8/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "5ab5ff86dd059f7df61794fa4597125ab65e25d088782e5bf7107940ff9e1188",
+   "video_sha256": "858a9213af5a7f2a4b431b0d6b97fd4af795ea04c10b2f6659132e3bb3138ac1",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-privileged--booovbq.json

@@ -0,0 +1,243 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-privileged--booovbq",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 494,
+   "submitted_actions": 534,
+   "sim_steps": 494,
+   "physics_steps": 4940,
+   "sim_time_s": 12.350000000000108,
+   "final_reward": 1.0,
+   "replay_wall_s": 523.3,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 494,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 4,
+   "agent_resets": 12,
+   "agent_sim_steps": 988,
+   "agent_records": 2,
+   "agent_replays": 5,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-blanket-flatten-08-i01-privileged--booovbq/replay.mp4",
+  "duration_s": 12.4,
+  "fps": 15.0,
+  "frames": 186,
+  "sha256": "f75e67f57623a020ef5d0c81cecc6dfddc08b7b791de8a0d43c5562085320d44"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [
+     [
+      12.33,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的蓝色方格毯横贯中部存在翻折重叠层，外形较窄，展平节点初始不成立。"
+     },
+     {
+      "t": 2.0,
+      "log_step": null,
+      "text_zh": "观察：视频 1.2–3.4 s，机器人右夹爪闭合夹住毯子右侧局部，抬起后向外铺下，该侧毯边随夹爪移动；视频 5.2–7.4 s，左夹爪以同样方式夹起并铺开左侧折角。此时中部横折仍在。"
+     },
+     {
+      "t": 10.0,
+      "log_step": null,
+      "text_zh": "观察：视频 8.9–10.4 s，机器人两夹爪各夹住毯子近侧的一处材料，协同抬起，侧视可见毯边离桌并下垂展开；视频 10.5–11.33 s，两爪将毯边铺下，中部重叠层消失，毯子成为近矩形。"
+     },
+     {
+      "t": 12.33,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 11.0–12.33 s，两夹爪松开后向上撤离，末帧侧视可见毯面贴桌，腕部视图可见张开的爪指与毯边分离；毯子保持平铺，未再次折回。"
+     },
+     {
+      "t": 12.33,
+      "log_step": 101,
+      "text_zh": "观察：inputs/run.json 的 reward 报告 success=1、all_success=1、deterministic=1、strict_goal_reached=1、strict_goal_row=494，replay_hold_steps=0、replay_tolerance_needed=0；log step 101 的评分器命令行输出也确认提交文件两次回放严格成功，确认窗口最低覆盖率为0.9835513647205534、最高粒子速度为0.027424147098699156 m/s。"
+     },
+     {
+      "t": 12.33,
+      "log_step": 101,
+      "text_zh": "推断：log step 3 的物理步长0.0025 s及控制抽取10给出每动作行0.025 s；log step 94–95 表明重新生成的提交动作与保存版本一致，log step 101 报告该文件在控制行494完成，即仿真12.35 s，与视频12.4 s时长一致。动作行451后的松手及行482前的撤离对应视频11.27–12.07 s，已在末段逐帧核对；因此将首次完成锚定到视频最后可见帧12.33 s，约有一帧误差。毫米级清距、材料伸长和完整半秒联合确认由严格评分结果支持。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 8.9–11.33 s，左右二指夹爪各自闭合夹住毯子近侧不同材料片区，共同抬起、展开并铺下，再松开撤离。此前视频 1.2–3.4 s 先由右夹爪单独处理右侧折角，视频 5.2–7.4 s 再由左夹爪单独处理左侧折角；均为夹住材料后带动毯面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": 0.9835513647205534
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": 0.027424147098699156
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人依次抓取展开两侧折角，再双爪协同抬起并铺开中部折叠，松手撤离后于视频约12.33 s完成毯子展平，最终节点1/1，无失败断点。评分器两次回放均严格成功，未用补时或容差；原生覆盖率与速度取log step 101评分器确认窗口最差值，其余指标未见评分器数值报告而留空。无状态日志，毫米级高度、清距、伸长及持续确认无法单靠视频精确复测，最终联合验收依据严格成功字段。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "毯子展平节点在视频约12.33 s（末帧，与sim 12.35 s一致）达成，双臂协作抓握及无失败断点均与帧和评分字段一致。",
+    "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": "2026-10-31T00:05:47Z",
+  "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/roboarc-blanket-flatten-08-i01-privileged--booovbq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--booovbq/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c83cd57a342b9a5f4bdaa6f9473f3741e7fec9ee3db0e7b82fcca2ef29545ab8",
+   "video_sha256": "f75e67f57623a020ef5d0c81cecc6dfddc08b7b791de8a0d43c5562085320d44",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-privileged--cz2ulht.json

@@ -0,0 +1,292 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-privileged--cz2ulht",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 2589,
+   "submitted_actions": 2589,
+   "sim_steps": 2669,
+   "physics_steps": 26690,
+   "sim_time_s": 66.72499999999805,
+   "final_reward": 0.0,
+   "replay_wall_s": 2312.4,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 80,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 1,
+   "agent_resets": 7,
+   "agent_sim_steps": 0,
+   "agent_records": 58,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-blanket-flatten-08-i01-privileged--cz2ulht/replay.mp4",
+  "duration_s": 66.8,
+  "fps": 15.0,
+  "frames": 1002,
+  "sha256": "34195002c8ad4a636381ca7731d6305893d4e4f06aa9105fa27d5872902d8391"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的俯视及侧视中，蓝色方格毯横贯中部有翻折重叠层，开局未展平。"
+     },
+     {
+      "t": 12.0,
+      "log_step": null,
+      "text_zh": "观察：视频 12.0 s 两只夹爪分别夹住毯子左右不同片区；视频约 12–44.5 s 机器人多次换握和配合拉动，毯子铺开的范围逐渐增大，但横向折层仍在。"
+     },
+     {
+      "t": 50.0,
+      "log_step": null,
+      "text_zh": "观察：视频 50.0 s 主画面显示毯子已大幅展开，右侧仍翘起，侧视可见右侧毯面被右爪提离桌面，未达到整体贴桌并撤离的联合条件。"
+     },
+     {
+      "t": 64.5,
+      "log_step": null,
+      "text_zh": "观察：视频 61.5–64.5 s 右爪在毯面中部夹着局部褶脊向下整形，毯面仍有浅脊和弯曲方格，右爪没有打开撤离。"
+     },
+     {
+      "t": 66.73,
+      "log_step": null,
+      "text_zh": "观察：视频 65.75–66.73 s（含末帧）右爪仍闭合停在毯面中部，腕部视图中布面进入两指之间；毯子中下部仍有浅褶脊，没有出现松手撤离后的完整平铺状态。参照成功视频 10.47 s 可见毯面近矩形、双爪均已张开并离开。"
+     },
+     {
+      "t": null,
+      "log_step": 435,
+      "text_zh": "观察：log step 348 重置后清空动作数组并开始本次录制；log step 435 保存的最终动作数组为 2589 行，与评分器 n_actions=2589 一致。评分器 strict_goal_reached=0、success=0；结合视频中折叠、翘起以及最后仍夹住毯子的过程，联合展平节点未达到。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 12.0 s 及约 32–40 s，左右二指夹爪各夹住毯子的不同片区，配合向外或沿桌面拉开，布面随夹爪移动；其间依次松开换握。视频约 50–64.5 s 主要由右爪夹住翘起片区和中部褶脊逐步降低，左爪后段张开停在边缘上方；视频 66.73 s 右爪仍未撤离。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 63.0,
+   "log_step": 422,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 61.5–64.5 s 机器人用右爪夹住中部褶脊向下整形，褶脊只部分减小，右爪随后一直闭合停在布面上，直至视频 66.73 s 仍未完成松手撤离。log step 421–422 记录最后数次降低褶脊动作，log step 435 保存该动作序列；展平步骤最终未修复，评分器 strict_goal_reached=0。"
+  },
+  "unmet": {
+   "blanket_flattened": false
+  },
+  "phenomena": [
+   {
+    "cat": "其他",
+    "item": "残余褶皱",
+    "t": 63.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 58.0–64.5 s 右爪从右侧转向毯面中部处理褶脊；视频 64.5 s 和 66.73 s 主画面中下部仍有浅脊，方格沿该带弯曲、压缩，未像成功参照视频 10.47 s 那样展开成连续平整的方格面。 视频 62.0–66.73 s 右爪仍夹住毯面中部褶脊，未见最终松手撤离；结合 log step 408、420–422 的抓持下降动作，这仍是局部整形过程，不能据此判为释放失败。",
+    "note_zh": "候选名称“残余褶皱”：这是柔性片材未充分展开的可见结果，现有现象条目未直接覆盖；不能仅凭褶皱认定脱手、阻挡碰撞或接触力异常。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 63.0,
+    "log_step": 422,
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 408 定义的局部整形控制保持右爪抓住褶脊，仅逐次降低末端；log step 421–422 每次请求降低 6 毫米，而记录的实际 TCP 高度从约 0.95725 米降至 0.95645 米，后几次位移很小，局部伸长比例同时上升。视频 61.5–66.73 s 也显示夹爪在褶脊处低位停留，未把残余褶皱充分展开；因此最后的局部整形控制未实现目标。现有证据不足以进一步区分接触几何、轨迹跟踪或反馈控制子类。"
+   }
+  ],
+  "gap_zh": "没有状态日志；评分器 reward 未报告 DAG 中八项连续指标，native_metrics 均为 null。log step 440 的约 0.9433 覆盖率是 agent 对会话状态的计算，发生在最终保存后额外保持动作之后，不能直接作为评分回放终态指标；log step 433 的 met 及 log step 438 的 P 也有旧变量问题，不用于节点判定。视频足以确认末段仍夹住毯子及残余褶皱，但不能复算毫米级清距、投影并集面积和所有粒子速度；控制根因的细分机制未记录。 运行墙钟为 10803 s（inputs/run.json wall_s），触及 10800 s 上限（time_limit_s）；会话在整形后的检查阶段结束，最后 log step 440 为指标计算，无最终收尾汇报。agent 在 log step 423 自述还要继续小幅降低，随后才松手撤离并验证；log step 435 保存的 2589 行动作序列末段尚未包含最终松手撤离，保存后仅额外 hold(4)，视频末段也未见释放。墙钟截断仅交代运行结束背景，不作为失败现象或根因，亦不换算成视频时间。"
+ },
+ "summary_zh": "机器人用双爪多次换握拉开毯子，并在后段用右爪降低翘起片区和中部褶脊，但“毯子展平”节点始终未达到，终态为 0/1。视频约 63.0 s 后局部整形趋于停滞，视频 66.73 s 仍有浅褶脊且右爪未松开撤离；日志支持最后的整形控制未能完成目标。缺少状态日志和评分回放连续指标，无法精确分解各毫米级验收条件或进一步确定控制子类。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点未达成、断点（展平·操作中）、时间与接触标签与视频一致，开局折叠、中段多次换握、末段右爪夹住褶脊均已核对。问题在于：把“没松开”列为失败现象，而 agent 当时仍在整形（尚未进入松手阶段），覆盖率本身也未达标；并且 gap_zh 没有交代该运行在 10800 s 墙钟上限（wall_s=10803）被截断，agent 自己计划“整形后再松手撤离”。",
+    "n_issues": 2,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成（视频末帧毯子仍呈歪斜四边形、右爪仍闭合夹在中部，对照成功视频 10.4 s 近矩形且双爪张开）、断点（展平·操作中）、接触标签与单一现象“残余褶皱”及根因“控制”均有视频与日志支持；上一轮的两处问题已修正，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": "2026-10-31T00:17:06Z",
+  "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/roboarc-blanket-flatten-08-i01-privileged--cz2ulht/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--cz2ulht/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--cz2ulht/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--cz2ulht/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "49b71c39edf62fd117b0d7fb2ebd139a6af7f2f20aa5169560f280788e1c1ea9",
+   "video_sha256": "34195002c8ad4a636381ca7731d6305893d4e4f06aa9105fa27d5872902d8391",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-privileged--jipgzjf.json

@@ -0,0 +1,280 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-privileged--jipgzjf",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 698,
+   "submitted_actions": 698,
+   "sim_steps": 778,
+   "physics_steps": 7780,
+   "sim_time_s": 19.449999999999964,
+   "final_reward": 0.0,
+   "replay_wall_s": 664.5,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 80,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 1,
+   "agent_resets": 23,
+   "agent_sim_steps": 0,
+   "agent_records": 4,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.639,
+   "peak_memory_gb": 2.116,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.637,
+   "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/roboarc-blanket-flatten-08-i01-privileged--jipgzjf/replay.mp4",
+  "duration_s": 19.533,
+  "fps": 15.0,
+  "frames": 293,
+  "sha256": "fa763de30b7d8ee6eaf50012cd68882f41753c21e4d9620b4aedd7e01b8d5df5"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的蓝色方格毯在桌面上呈横向折叠，中央有贯穿毯面的重叠层，初始并未展平。"
+     },
+     {
+      "t": 3.25,
+      "log_step": null,
+      "text_zh": "观察：视频 1.75–3.5 s，机器人右夹爪夹住远侧毯边并提起，毯边随夹爪升高，随后松手落回桌面；横向重叠层仍在。"
+     },
+     {
+      "t": 11.5,
+      "log_step": null,
+      "text_zh": "观察：视频 4.5–7.5 s 和视频 8.5–11.5 s，机器人左夹爪先后夹住近侧毯边的两个位置向外拉，毯边随之移动、投影范围增大，但两次拉动后中央仍有连续折叠重叠。"
+     },
+     {
+      "t": 19.4,
+      "log_step": null,
+      "text_zh": "观察：视频 12.0–14.5 s 两臂张爪撤回；视频 14.5–19.4 s 毯子基本不动，横贯毯面的翻折层一直没有消失。成功参考视频约 4.0–10.4 s 可见同一主视角下重叠层消失、毯面完整铺开，仅用于核对外观。"
+     },
+     {
+      "t": null,
+      "log_step": 169,
+      "text_zh": "观察：log step 166–168 记录并保存最终 698 行动作，顺序为右侧提拉、左侧两次外拉、松开撤离，与本次视频一致；log step 169 对该文件回放报告 success=False、strict=False。评分器 reward 的 n_actions=698、strict_goal_reached=0，与可见残余折叠一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约 1.75–3.5 s 机器人右二指夹爪夹住远侧毯边提拉，视频约 4.5–7.5 s、8.5–11.5 s 左二指夹爪依次捏住近侧毯边两个位置外拉；夹口处毯边随手爪移动，随后分别松开。两爪依次操作，未见同时夹持拉展。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 11.5,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 11.5–12.0 s，机器人结束第二次左爪外拉并开始松开，毯面仍有贯穿中部的重叠折层；视频 14.5–19.4 s 撤离后的静止画面显示该折层未被修复，因此“毯子展平”未完成。"
+  },
+  "unmet": {
+   "blanket_flattened": false
+  },
+  "phenomena": [
+   {
+    "cat": "其他",
+    "item": "残留折叠",
+    "t": 11.5,
+    "log_step": null,
+    "evidence_zh": "观察：视频约 11.5 s 左爪最后一次拉动后，毯子中部仍有一条长的翻折重叠层；视频 14.5–19.4 s 双爪停止操作后该重叠层持续存在，毯面未成为无重叠的完整铺展面。",
+    "note_zh": "候选名称：残留折叠。现有现象条目没有描述柔性材料展开后仍保留折叠重叠的形变结果；未见足以判定夹持丢失、碰撞阻断或异常接触力的证据，按 CASES.md OPEN-19（D25）使用其他并具体描述。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "无状态日志；评分器 reward 未报告 DAG 所列八项连续指标，因此 native_metrics 均留 null，不能将 log step 167 录制会话中脚本计算的覆盖率 0.7799、伸长比 1.0187 当作评分回放的原生测量。视频足以确认残余折叠和实际夹持拉动，但无法独立复算投影并集面积、全部粒子高度与速度、材料边长比、毫米级清距和位移。log step 160、163 的 23、47 毫米逆解误差来自会话内重置测试，不能逐路点定位到评分回放；这些诊断可作根因线索，但尚无证据证明它们导致本次回放的残留折叠。log step 166–167 的最终录制覆盖率有所增加、log step 169 的回放验收仍失败，结合视频只能确认展开不足，无法区分拉展策略、接触设计与轨迹执行各自的贡献，故根因留空。"
+ },
+ "summary_zh": "机器人右爪先提拉远侧毯边，随后左爪在近侧两个位置依次夹持外拉，完成松手撤离，但视频约 11.5 s 最后一次外拉后仍留有横贯毯面的重叠折层，至视频 19.4 s 仍未消除，唯一节点“毯子展平”始终未达成。失败发生在展平操作中，现象为残留折叠；会话测试的逆解失败尚不能与本次评分回放的失败建立因果联系，根因留空。缺少粒子状态和评分回放的连续指标，无法进一步区分策略与执行因素或精确量测联合验收阈值。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "WHERE 标签（节点未达成、接触、断点 11.5 s）与视频一致：0 s 横向折叠，3.25/6/9/11.5/14/19 s 帧中折层始终存在。问题在根因：把会话测试中的 IK 失败当作评分回放的控制缺陷，证据不足；该提拉程序在会话中实际把覆盖率拉到 0.78，更像拉展策略不足而非控制失败。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、接触、断点（视频 11.5 s）与视频一致：11.5 s 与 19 s 帧中折层仍在；第一轮意见已采纳，根因留空并写明 gap_zh，没有新的错误。",
+    "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": "2026-10-31T00:16:48Z",
+  "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/roboarc-blanket-flatten-08-i01-privileged--jipgzjf/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--jipgzjf/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--jipgzjf/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--jipgzjf/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "36a116ea5afc79147c36b4522ed8d4e7eed00349b76418396662b158520188bc",
+   "video_sha256": "fa763de30b7d8ee6eaf50012cd68882f41753c21e4d9620b4aedd7e01b8d5df5",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-privileged--maymsw4.json

@@ -0,0 +1,268 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-privileged--maymsw4",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 1310,
+   "submitted_actions": 1310,
+   "sim_steps": 1390,
+   "physics_steps": 13900,
+   "sim_time_s": 34.7499999999991,
+   "final_reward": 0.0,
+   "replay_wall_s": 1003.3,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 80,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 23,
+   "agent_resets": 21,
+   "agent_sim_steps": 19334,
+   "agent_records": 3,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 1.078,
+   "peak_memory_gb": 3.437,
+   "peak_processes": 13,
+   "peak_disk_gb": 1.296,
+   "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/roboarc-blanket-flatten-08-i01-privileged--maymsw4/replay.mp4",
+  "duration_s": 34.8,
+  "fps": 15.0,
+  "frames": 522,
+  "sha256": "250f258a79f375215b53f41c614186e9350865e655911e8761c72a38a8821321"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的主画面显示蓝色方格毯横贯中部有翻折重叠带，侧视也有叠层，开局未展平。"
+     },
+     {
+      "t": 3.2,
+      "log_step": null,
+      "text_zh": "观察：视频 1.6–4.3 s 机器人右夹爪夹住布边并向外拉，布边随之移动；视频 6.4–9.1 s 左夹爪在另一处作同类夹拉，随后还有交替夹拉。所检视频 0.0–26.4 s 各段始终可见横向折层，夹拉只使毯子部分展开。"
+     },
+     {
+      "t": 26.4,
+      "log_step": null,
+      "text_zh": "观察：视频 23.8–26.4 s 机器人左夹爪完成最后一处夹拉并松开抬离，主画面中横贯毯面的凸起重叠条仍未消失。"
+     },
+     {
+      "t": 34.73,
+      "log_step": null,
+      "text_zh": "观察：视频 27.2–34.73 s 两爪已抬离毯子，毯子外形基本不再改变，但中部横向折层和侧视叠层一直保留；视频 33.53–34.73 s 已逐帧核对。成功参考视频 10.0 s 同一主视角中的毯面则连续平铺、无该折层，仅用于外观比较。"
+     },
+     {
+      "t": null,
+      "log_step": 212,
+      "text_zh": "观察：log step 205 的最终生成程序依次执行六次单爪 peel 后撤离；log step 212 确认保存 1310 行最终动作，和评分器 reward.n_actions=1310 一致。评分器 strict_goal_reached=0、success=0；这些字段与视频所见未展平相符，未用早期测试的结果替代本次回放。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：机器人左右二指夹爪依次夹住蓝色毯子局部布边并短距离外拉，再松开撤回。视频 1.6–4.3 s 右爪和视频 6.4–9.1 s 左爪的夹持及布边随动可见，视频 11.2–26.4 s 继续交替操作；未见两爪同时夹持配合拉布。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 26.4,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 23.8–26.4 s 最后一次左爪夹拉结束，夹爪松开抬离时横贯毯面的重叠折层仍在，随后视频 27.2–34.73 s 未再修复。机器人已操作并部分展开毯子，但从未形成该节点要求的平铺形态；以最后一次未能消除折层的操作结束处作为导航锚点。"
+  },
+  "unmet": {
+   "blanket_flattened": false
+  },
+  "phenomena": [
+   {
+    "cat": "其他",
+    "item": "残余折叠",
+    "t": 26.4,
+    "log_step": null,
+    "evidence_zh": "观察：视频 26.4 s 左夹爪离开时，蓝色毯面仍有一条几乎横贯全宽的凸起重叠带；视频 34.73 s 主视图与侧视图仍显示该折层，未形成连续平铺的布面。",
+    "note_zh": "候选名称：残余折叠。可见问题是柔性材料的重叠折层未被展开；现有接触或运动条目没有直接描述这一材料形态缺陷，未观察到可据以归为滑脱、碰撞或放后不稳的事件。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 26.4,
+    "log_step": 205,
+    "evidence_zh": "推断：log step 205 的交付程序采用六次局部单爪短拉；log step 206 显示该次生成的六轮结束后自测覆盖率仅 0.8200，log step 212 的最终自测仍为 0.8177。视频 1.6–26.4 s 可见夹爪确实带动布边，但中部折层未被拉开，因此支持所用柔性材料操作控制不足以实现完整展开；这里的自测仅作控制归因佐证，不当作评分器指标或视频时间。",
+    "source_zh": "日志",
+    "note_zh": "只标控制大类；证据不足以进一步区分抓点和接触几何、运动幅度或反馈控制问题。"
+   }
+  ],
+  "gap_zh": "本目录没有状态日志；评分器 reward 未报告 DAG 中八项连续指标，native_metrics.value 均留 null。视频足以确认残余折叠和单爪交替夹拉，但不能精确测量材料覆盖率、粒子高度、伸长、速度、机器人清距、确认位移和连续确认时长，也不能细分具体控制缺陷。log step 212 的数值来自最终生成会话自测，未冒充评分器数值；早期会话中的双爪试验、高抬布料及抓取失败不作为本次回放现象。"
+ },
+ "summary_zh": "机器人用左右夹爪交替夹拉布边，使毯子部分展开并最终松手撤离，但横贯中部的重叠折层始终保留，“毯子展平”节点未达到，最终为 0/1。断点锚定视频 26.4 s 最后一次夹拉结束处；所用局部夹拉控制未能消除折层，评分器 strict_goal_reached=0 与视频一致。缺少状态日志和评分器连续指标，无法复算毫米级验收条件或细分控制原因。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "视频 0/10/20/26.4/34.5 s 帧与成功参考视频 10 s 对比，毯子始终保留横向折层，节点未达成、断点与接触标注与画面相符；根因仅标控制大类且为推断，已如实标注。",
+    "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": "2026-10-31T00:04:48Z",
+  "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/roboarc-blanket-flatten-08-i01-privileged--maymsw4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--maymsw4/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "4d32b27cdf2393e70bafb6609874368a3423ba4503146f2ab597fc5912c7deae",
+   "video_sha256": "250f258a79f375215b53f41c614186e9350865e655911e8761c72a38a8821321",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-privileged--z3woo9f.json

@@ -0,0 +1,282 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-privileged--z3woo9f",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 1435,
+   "submitted_actions": 1435,
+   "sim_steps": 1515,
+   "physics_steps": 15150,
+   "sim_time_s": 37.87499999999892,
+   "final_reward": 0.0,
+   "replay_wall_s": 2397.0,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 80,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 1,
+   "agent_resets": 22,
+   "agent_sim_steps": 0,
+   "agent_records": 1,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 1.168,
+   "peak_memory_gb": 2.174,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.639,
+   "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/roboarc-blanket-flatten-08-i01-privileged--z3woo9f/replay.mp4",
+  "duration_s": 37.933,
+  "fps": 15.0,
+  "frames": 569,
+  "sha256": "350dc01a4d6f1cdd221169e0cadca7aac39f7e32cfc1a0b159dca73aaf7bde66"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频开局蓝色方格毯横向折叠、重叠，尚未展开；视频约 1–10 s，机器人先用左夹爪捏起并向后拉毯面，再换右夹爪拉另一侧，大部分横向折叠逐渐打开。"
+     },
+     {
+      "t": 6.7,
+      "log_step": 137,
+      "text_zh": "观察：视频 6.2–7.4 s，左夹爪在毯子后缘上方移动，毯面局部隆起后回落；视频 15.0–17.4 s 的左侧折角修正仍留下折回边，帧中不能直接辨认脱手瞬间。log step 137 的026录制输出为 arch grasp: blanket、arch pull: None，稍后 flap grasp: blanket、flap pendulum: None。推断：log step 80 的 drag_log 明确返回 p[\"object_in_hand\"](arm)，log step 131 在该返回之后才执行 release_arm，因此这些输出支持拉动末端已不再持有的推断；视频约 6.7 s 仅为对应拉动阶段的粗略导航锚点。"
+     },
+     {
+      "t": 28.0,
+      "log_step": null,
+      "text_zh": "观察：视频约 19–23 s，右夹爪捏住前侧边缘向前拉，视频约 26.5–28 s 两爪张开并撤离；视频 28–37.87 s，包括最后一秒，毯子左后侧仍有内缩的斜边及折回区域，未成为完全展开的矩形。成功参考视频 10.4 s 的同一主视角显示展开后更完整的方格面和边角，仅作外观比较。"
+     },
+     {
+      "t": 37.8,
+      "log_step": 150,
+      "text_zh": "推断：log step 150 报告026保存了1435行动作，log step 154 核对交付数组为1435行，与评分器 n_actions=1435 一致；该脚本的单爪交替拉动、松手撤离次序也与视频吻合，因此后续排队的029双爪方案不计入本轨迹。结合视频残余折叠和评分器 strict_goal_reached=0、success=0，毯子展平节点全程未达到联合验收。log step 151 中026保存前的脚本诊断 coverage=0.945 仅为辅助证据，并非评分器原生指标读数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约 1–7 s 左二指夹爪捏住毯子后侧局部并提拉；视频约 9–10 s 换右夹爪拉后侧，视频约 14–17 s 左夹爪再次捏拉左侧折回处，视频约 20–23 s 右夹爪捏拉前侧边缘。两只夹爪依次操作，未见同时夹持毯子协同拉动。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 6.7,
+   "log_step": 137,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 6.2–7.4 s，左夹爪拉动后毯面局部回落，不能直接辨认抓持脱离瞬间；视频 15–17.4 s 再次修正左侧折回处后仍有残留，随后机器人处理前缘并撤离，视频 37.8 s 左侧仍呈内缩斜边。推断：log step 137 的 arch grasp: blanket → arch pull: None，结合 log step 80 中 drag_log 返回 object_in_hand、log step 131 在返回后才松手的顺序，支持此轮拉动未保持抓持；视频约 6.7 s 是该阶段的粗略锚点，不是由日志换算的脱手时刻。展平节点后续未修复至达标。"
+  },
+  "unmet": {
+   "blanket_flattened": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 6.7,
+    "log_step": 137,
+    "evidence_zh": "观察：视频 6.2–7.4 s 只见左夹爪在后缘上方移动、毯面隆起后回落，未直接看清脱手瞬间；视频 17.4 s 及 37.8 s 左侧仍有折回和内缩斜边。log step 137 的026输出为 arch grasp: blanket、arch pull: None，稍后 flap grasp: blanket、flap pendulum: None。推断：log step 80 的 drag_log 显式返回持物查询 p[\"object_in_hand\"](arm)，故 None 并非函数缺少返回值；log step 131 在返回之后才调用 release_arm，支持拉动期间未保持抓持。视频约 6.7 s 只是粗略定位，具体脱离时刻与机制未获直接观察。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 131,
+    "source_zh": "日志",
+    "evidence_zh": "推断：控制归因来自 log step 80、131、137 的代码与执行输出，不能把视频约 6.7 s 脱手视为直接观察。log step 80 的 drag_log 明确返回持物查询 p[\"object_in_hand\"](arm)；log step 131 的026脚本在左侧折角修正中只检查初始 hc，拉动后不检查返回值或展平结果，执行松手后即置 m2=True 并跳出候选循环。log step 137 实际输出 flap grasp: blanket、flap pendulum: None、M2 done: True 后转向右侧前缘，支持脚本没有针对拉动末端未持有继续重抓修正的判断。观察：视频 15–17.4 s 及 37.8 s 可见左侧残余折回；这些帧不能直接证明抓持为何丢失。推断：该控制流程未闭环修正残余折角，可能使其保留；不足以细分为夹持几何、载荷或轨迹参数问题。"
+   }
+  ],
+  "gap_zh": "无状态日志；评分器未提供DAG列出的8项连续指标，故 native_metrics.value 均为 null，不能用会话脚本的自算结果替代。视频 6.2–7.4 s 显示毯面隆起后回落，未直接显示抓持脱离瞬间；约视频 6.7 s 的失持及控制归因是结合已交付026代码、持物查询返回值和残余折回作出的推断，具体脱离时刻、机制及其对最终缺陷的贡献无法独立确定。视频也不能精确测量95%投影面积、所有粒子高度与伸长、毫米级清距或确认位移。节点判定依据实际残余折叠及严格目标未达到的评分字段；其他重置试验与未交付的029方案不作为进度证据。"
+ },
+ "summary_zh": "机器人用左右夹爪依次提拉，打开了大部分横向折叠，最后张爪撤离，但视频 37.8 s 左侧仍有折回和内缩斜边，毯子展平节点最终为0/1，评分器 strict_goal_reached=0。推断：log step 80、131、137 的代码与持物查询输出支持拉动末端未保持抓持、脚本未据此继续修正；视频约 6.7 s 仅定位相关拉动阶段，帧中未直接看清脱手瞬间。因关键失持事件及控制归因仍有推断成分，缺少状态日志和评分器连续指标，conf 降为5。",
+ "conf": 5,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "进度节点（未达成、final=false）、接触（左右夹爪依次抓握）、n_actions 与视频的对应均与帧一致；末帧毯子左后/左前仍有斜边，评分器 strict_goal_reached=0，支持未展平。问题在于失败原因的证据强度：6.7 s 断点与『没保持���/控制』主要靠 026 脚本的 None 返回值（agent 会话自带输出）推断，视频帧并不能显示夹爪脱手，整体 conf=7 偏高。",
+    "n_issues": 1,
+    "n_major": 0
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未达成、接触（左右夹爪依次抓握）、断点引用与推断/观察区分均与帧和日志一致；第1轮的置信度问题已处理，无新证据。",
+    "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": "2026-10-31T00:15:55Z",
+  "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/roboarc-blanket-flatten-08-i01-privileged--z3woo9f/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--z3woo9f/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--z3woo9f/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-privileged--z3woo9f/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "39a99198741bf9358501ad6ff100aab8630f83db11e7647afaa11f930b9a2f87",
+   "video_sha256": "350dc01a4d6f1cdd221169e0cadca7aac39f7e32cfc1a0b159dca73aaf7bde66",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-standard--cfkgyzq.json

@@ -0,0 +1,250 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-standard--cfkgyzq",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 12517,
+   "live_rows": 12517,
+   "sim_steps": 12517,
+   "physics_steps": 125170,
+   "steps": 3029,
+   "elapsed_s": 10773.8,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-blanket-flatten-08-i01-standard--cfkgyzq/replay.mp4",
+  "duration_s": 313.0,
+  "fps": 15.0,
+  "frames": 4695,
+  "sha256": "a4f102cc6f690a66d15383eec9bc779c7489b034d1f77deaf4471e265d87c688"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，机器人两臂位于蓝色方格毯两侧，毯面横向翻折形成明显重叠层，初态未展平。"
+     },
+     {
+      "t": 122.0,
+      "log_step": null,
+      "text_zh": "观察：视频 112.0–129.0 s，机器人把毯子拉成较宽的近矩形，但后缘仍有折叠条，两爪仍在布边持布；视频 142.0–144.0 s 双爪离开时，毯子的右前角仍内折，轮廓有明显斜切角，未成为完整展开的矩形。"
+     },
+     {
+      "t": 310.4,
+      "log_step": 466,
+      "text_zh": "观察：视频 308.0–310.0 s，机器人左夹爪向远端翘起的布尖靠近并闭合；视频 310.4–312.93 s 左爪退向左后方时没有带走布尖，毯子仍呈多层折叠、隆起的非矩形。agent 也在 log step 466 自述“far-tip grasp missed”，随后提出的重新对齐动作未出现在本次视频中。"
+     },
+     {
+      "t": 312.93,
+      "log_step": null,
+      "text_zh": "观察：视频最后一帧 312.93 s，主视图可见毯子中央与左下方仍有重叠折层，侧视显示毯面隆起；评分器字段 live_success=0、success=0，且 graded_by_live=1、resumes=0。成功参考视频 10.466 s 则显示毯子平铺成完整近矩形、双爪撤离；参考仅用于外观比较。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频约 35.0–40.0 s，两只二指夹爪分别夹住毯子两侧材料并共同提起；视频 105.0–113.0 s 左爪固定一角、右爪牵拉另一侧。之后也交替使用单爪夹住布边拉动，例如视频 294.0–298.0 s 左爪低位牵拉；末次视频 310.4–312.93 s 左爪空拉不算成功抓握。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 310.4,
+   "log_step": 461,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 308.0–310.0 s 机器人左爪靠近毯子的远端布尖，视频 310.4–312.93 s 随后空手向左后方拉开，布尖留在原位，毯子仍有重叠和隆起。log step 461 的控制脚本要夹住该角后外拉，log step 465 返回该拉动已执行但 done=false；整个展平步骤未在后续修复。"
+  },
+  "unmet": {
+   "blanket_flattened": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 310.4,
+    "log_step": 465,
+    "evidence_zh": "观察：视频 309.8 s 左夹爪闭合在远端布尖附近，视频 310.4–312.93 s 夹爪退离但布尖没有随爪移动，末次外拉没有建立能牵动布角的夹持。log step 465 的服务返回左爪 opening=0.0；agent 在 log step 466 自述该抓取未命中，与视频的空拉一致。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 461 的脚本命令左爪闭合下降后向外拉，log steps 463、465 的机器人服务返回这两段动作均无逆解失败，左爪开口为零；视频 309.8–312.93 s 却显示空爪离开、布角未跟随。结合视频 146.0–153.0 s 先前夹爪闭合和试提也未带起目标折角，归因为有效抓持未建立的控制问题；证据不足以进一步区分抓取接触几何与末端跟踪问题。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有模拟器状态日志，也没有评分器报告的八项原生连续指标；无法从画面精确复算投影并集面积、粒子高度、材料伸长、速度、毫米清距和确认漂移，native_metrics 均为 null。视频可直接判断终态残留折叠和末次空拉；缺少接触对及逐刻目标/实际位姿，控制根因不再细分。视频 290.8–291.0 s 布爪分离与随后主动张爪相邻，未据 agent 在 log step 441 的自述另判“没保持住”。"
+ },
+ "summary_zh": "机器人多次以双爪夹持提起、单爪牵拉并换抓，曾将毯子大部分铺开，但始终未完成“毯子展平”节点，最终为 0/1。末次远角抓取在视频约 310.4 s 出现空拉，直到视频 312.93 s 毯子仍有明显重叠与隆起，失败归于抓持建立不可靠的控制问题。缺少状态日志和原生指标，无法核验毫米级联合验收或进一步细分控制原因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "终态仍有折叠隆起（视频 312.9 s），final=false、断点在展平步骤、末次远角空拉与视频一致；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": "2026-10-31T00:04:23Z",
+  "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/roboarc-blanket-flatten-08-i01-standard--cfkgyzq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--cfkgyzq/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "a477e265d703c970dfb6eaf5b20501c32db1d997accc660dc553993cef701ccd",
+   "video_sha256": "a4f102cc6f690a66d15383eec9bc779c7489b034d1f77deaf4471e265d87c688",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-standard--mt2rkas.json

@@ -0,0 +1,279 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-standard--mt2rkas",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 1,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 16247,
+   "live_rows": 16247,
+   "sim_steps": 16247,
+   "physics_steps": 162470,
+   "steps": 3840,
+   "elapsed_s": 10094.8,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.358,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.352,
+   "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/roboarc-blanket-flatten-08-i01-standard--mt2rkas/replay.mp4",
+  "duration_s": 406.267,
+  "fps": 15.0,
+  "frames": 6094,
+  "sha256": "609b67374167bc90bdfc698d31141d46e61d0c58d5f87b3a4c566ce896f187f8"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 主画面中的蓝色方格毯横向翻折，中部有贯穿毯面的重叠层，开局未展平。"
+     },
+     {
+      "t": 28.8,
+      "log_step": null,
+      "text_zh": "观察：视频 16.0–28.8 s，机器人双爪接触并拉动毯边，随后抬开；视频 28.8 s 毯面仍有贯穿左右的横向折叠，未出现完全展开的矩形毯面。"
+     },
+     {
+      "t": 39.5,
+      "log_step": 104,
+      "text_zh": "观察：视频 33.6–42.7 s，机器人两只夹爪分别夹住中部折层并提起，毯面随爪升高，原先铺在桌面的部分向中间收拢，侧视出现高起的折壁；视频 45.9 s 放下后仍呈狭窄的多层折叠。log step 104 的脚本指定双爪夹取中部卷折并竖直抬升，log step 110 的 agent 也自述提拉使毯子聚成一束。"
+     },
+     {
+      "t": 406.2,
+      "log_step": 265,
+      "text_zh": "观察：视频全程每约5秒抽查，并对视频 27.2–45.9 s 和最后一秒加密查看；后续反复提拉、推拨和单爪甩拉后仍保留折叠。视频 405.0–406.2 s 双爪已撤离，毯子仍呈不规则收拢形状，中部与左侧有高起折脊及重叠，未平铺成矩形。log step 265 的机器人服务终态为 done=False，评分器 live_success=0；据可见形状判定该节点始终未达成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 33.6–42.7 s，左右二指夹爪各夹住毯子中部的一处材料并一起提拉，毯面随两爪升起；此前两爪也共同拉动毯边。后期依次换用单爪夹住边部、提起并向外甩拉。"
+     },
+     {
+      "who": "双臂协作",
+      "how": "推/扫",
+      "note_zh": "观察：机器人也用闭合夹爪抵着折叠毯面向前推拨，使隆起折层移位、聚拢；log step 177 执行双爪推毯脚本，log step 179 返回实际推行终点，log step 182–184 接着用闭合夹爪推动隆起部分倾倒。此处记录夹爪外侧对毯面的推拨接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 39.5,
+   "log_step": 104,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 39.5 s 左右，机器人双爪提起中部折层，毯子被聚拢成悬起的折壁；视频 45.9 s 降下后折层仍叠在一起。后续继续尝试拉边、推倒和甩拉，直到视频 406.2 s 仍未展平，失败发生在毯子展平操作中。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 39.5,
+    "log_step": 107,
+    "evidence_zh": "观察：log step 104 的机器人脚本指定双爪下降至抓取高度 z=0.013 m；log step 107 中由机器人服务 proprio 读数打印的下降后及抓取后左右 TCP z 分别为 0.038/0.037 m，未到指令高度，闭合后夹口分别为 0.0035/0.0051 m。视频 33.6–42.7 s 可见双爪夹住中部折层的局部并提起，毯面随之向中间收拢成悬起的折壁；视频 45.9 s 放下后仍呈多层窄条。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 39.5,
+    "log_step": 107,
+    "evidence_zh": "推断：log step 104 指定抓取高度 z=0.013 m，而 log step 107 返回下降后及抓取后左右 TCP z=0.038/0.037 m，夹口为 0.0035/0.0051 m；视频 33.6–42.7 s 可见两爪夹住折层局部提起，毯面向内聚拢。抓取执行未达到计划位姿，随后提拉没有展开重叠层，支持控制层面的归因；现有证据不能区分折层接触阻挡、轨迹跟踪或抓取流程问题，也不能证明按指令高度执行的接触方案本身不可行，因此不选具体控制子类。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志；视频能确定明显折叠及终态未展平，但不能独立量测投影并集面积、所有粒子高度、材料伸长、速度、释放清距和确认漂移。评分器未报告 DAG 所列八项连续指标，故 native_metrics 的 value 均为 null；log step 98、108、266 的面积和高度是 agent 自写深度处理脚本的估计，未当作评分器原生指标。log step 104 与 log step 107 能确定指令抓取高度和服务返回位姿存在偏差，但缺少接触力、材料层数和轨迹跟踪记录，不能区分接触阻挡、跟踪或抓取流程问题；按指令执行是否能展开该折层及替代抓法均未验证，具体失败机理仍有推断。",
+  "unmet": {
+   "blanket_flattened": false
+  }
+ },
+ "summary_zh": "机器人双爪及单爪多次夹取、拉动毯子，也推拨过隆起折层，但“毯子展平”始终未达成，最终为0/1节点，评分器判定失败。视频约39.5 s 双爪夹提中部折层使毯面聚拢；log step 104 指定抓取高度 z=0.013 m，log step 107 返回抓取后左右 TCP z=0.038/0.037 m，支持位姿未到位，视频406.2 s 毯子仍有明显折脊和重叠。根因仅保留“控制”，因无法区分具体控制机理，且缺少状态日志和原生连续指标、无法复算毫米级联合验收，将置信度由8降至6。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点未达成、final=false、断点（视频约39.5 s 双爪提拉中部折层，log step 104）、接触标注与终态观察均与视频相符（视频 0/20/30/38/42/46/60/120/200/300/380/405 s 抽查）。问题在 step 104 的失败机理：服务读数显示夹爪只下到 z≈0.038（指令 0.013），只捏到折层顶部，现象与根因没写这一点，抓取设计的归因缺证据、conf 偏高。",
+    "n_issues": 3,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成（视频406 s 毯子仍成不规则折皱、非矩形）、断点（约39.5 s 双爪提拉中部折层，log step 104/107）、接触标注与终态均与视频和服务读数相符；第1轮意见已采纳，现象“位姿不准”与根因“控制”（不选子类、标为推断）有证据支持，未发现需改内容。",
+    "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": "2026-10-31T00:18:22Z",
+  "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/roboarc-blanket-flatten-08-i01-standard--mt2rkas/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--mt2rkas/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--mt2rkas/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--mt2rkas/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "bd5083555506fd0312fc1e9bb3546b521abc67dc53381b9a43480bedc4f2b1f9",
+   "video_sha256": "609b67374167bc90bdfc698d31141d46e61d0c58d5f87b3a4c566ce896f187f8",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-standard--xfhppg7.json

@@ -0,0 +1,283 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-standard--xfhppg7",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 20226,
+   "live_rows": 20226,
+   "sim_steps": 20226,
+   "physics_steps": 202260,
+   "steps": 4930,
+   "elapsed_s": 10740.4,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.372,
+   "peak_processes": 10,
+   "peak_disk_gb": 1.029,
+   "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/roboarc-blanket-flatten-08-i01-standard--xfhppg7/replay.mp4",
+  "duration_s": 505.733,
+  "fps": 15.0,
+  "frames": 7586,
+  "sha256": "b050b0546f23c4013223bedcc32b232e78aece863f1fc25caa86cfad7a41f9bf"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，蓝色方格毯横向折回，上下层重叠且折边贯穿毯面，展平条件开局不成立。"
+     },
+     {
+      "t": 71.0,
+      "log_step": null,
+      "text_zh": "观察：视频 46–84 s，机器人左夹爪多次下降、收拢并抬起；视频 67.0 s、71.0 s 的主视图和腕部视图仍显示毯子留在桌面、横向重叠折边保留，未见布片被稳定夹住并提起。"
+     },
+     {
+      "t": 253.0,
+      "log_step": null,
+      "text_zh": "观察：视频 226–254 s，左夹爪留在毯子左侧近桌面处，右夹爪压拖右半毯面，毯子的轮廓扩大并转动；视频 253.0 s 仍可见左侧局部折痕与内缩边缘，尚未形成完整展开的连续矩形。"
+     },
+     {
+      "t": 418.0,
+      "log_step": null,
+      "text_zh": "观察：视频 418.0 s，两夹爪已经抬离毯面，毯子大部分贴桌，主视图左缘略有内凹，是否仍有叠层看不清。成功参考视频 10.4 s 显示近矩形且格线展开的毯面；参考只用于外观比较，不能证明本回合哪项连续指标未通过。"
+     },
+     {
+      "t": 462.0,
+      "log_step": 191,
+      "text_zh": "观察：视频 454、458、462、466 s，机器人左夹爪在毯子左部近桌面压拖，毯子随之平移；左缘有轻微内凹，不能从画面确认该处存在折回重叠层。log step 191 的脚本命令左夹爪保持闭合，在近桌面高度进行横向拖动。"
+     },
+     {
+      "t": 505.6,
+      "log_step": null,
+      "text_zh": "观察：视频 504.8、505.2、505.6 s，机器人夹爪在毯子上方移动，毯子保持大部分贴桌、左缘略内凹的外观，未见明显进一步展开；inputs/run.json 的评分器字段 live_success=0、graded_by_live=1、resumes=0，表明本次实时回合未通过联合验收。"
+     },
+     {
+      "t": 505.6,
+      "log_step": 194,
+      "text_zh": "推断：开局明显翻折、后续压拖及撤离处的画面与 inputs/run.json 的实时评分结果相符；本任务仅有一个联合验收目标，live_success=0、graded_by_live=1、resumes=0 支持该节点无达成区间且终值为假。log step 194 中 agent 自编图像测量程序估计投影面积为0.2606平方米，但这不是评分器实测值，不能据此确定精确覆盖率或哪项联合验收条件未满足。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "压/擦",
+      "note_zh": "观察：视频 226–254 s，机器人左夹爪指尖留在毯子左侧近桌面处，右夹爪压住并拖动另一侧毯面；此阶段两爪同时作用于毯子。视频 160–210 s 及454–466 s 还可见左爪单独压拖，其他时段右爪单独压拖；这些接触依次发生。未见两指夹住布片后稳定提离桌面的抓握，不按DAG典型抓握方案填报。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 462.0,
+   "log_step": 191,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 454、458、462、466 s，机器人左夹爪进行末次近桌面压拖，毯子随之平移；视频 468、480 s 及最后一秒画面显示后续抬爪和撤离，毯子大部分贴桌，左缘仍略内凹，但叠层不可确认。log step 191 明确命令这次横向拖动，评分器最终 live_success=0；断点定位于“毯子展平”操作中，不能进一步确定未过的具体验收条件。"
+  },
+  "unmet": {
+   "blanket_flattened": false
+  },
+  "phenomena": [],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "t": 462.0,
+    "log_step": 191,
+    "evidence_zh": "推断：log step 175、191 的脚本均让单爪闭合后在约13–14毫米TCP高度压拖，没有夹持并分离布层的动作，也没有让另一爪固定下层。视频 360、380、400、418 s 可见压拖伴随毯子整体转动，视频 454、458、462、466 s 又显示毯子随左爪平移；这些动作与实时联合验收始终未通过相结合，支持接触方式未能充分控制毯子形变的可能归因，但不能证明残余折层存在或覆盖不足就是失败原因。agent 在 log step 202 自述因捏取未可靠建立而改用压拖，仅作为与画面一致的补充陈述。",
+    "source_zh": "日志 / 视频",
+    "note_zh": "没有状态接触对、层间受力或评分器连续指标记录，接触设计归因为推断，不能确定摩擦、法向压力或具体夹取位置的各自影响，也不能确定该方式使哪项联合验收条件未通过；本回合未验证修正后的可行接触方案。"
+   }
+  ],
+  "gap_zh": "无状态日志，评分器未报告DAG所列八项连续指标；不能精确量测投影并集覆盖率、所有粒子高度、材料伸长、速度、毫米级释放清距、漂移及连续确认时间。log step 152、188、194 的面积和高度为agent自编视觉估计，不能填作native_metrics。视频 418、440、480 s 及最后一秒画面只显示毯子大部分平整、左缘轻微内凹，是否仍有叠层看不清；未能确定覆盖率、折叠、伸长或其他联合验收项中具体哪项未过，故phenomena留空，不把节点未达成改写为现象。评分器支持唯一联合验收节点未达成；缺少接触力与布层状态，根因仅作接触设计层面的推断。"
+ },
+ "summary_zh": "机器人用夹爪压拖展开了开局的大段横向翻折，末段毯子大部分贴桌、左缘略内凹，实时评分器未通过联合验收，“毯子展平”节点最终为0/1。断点位于展平操作中；压拖伴随毯子整体转动和平移，结合脚本保留抓取和接触设计的根因推断。画面不能确认末段残留叠层，故现象留空；无状态日志或评分器连续指标，不能确定具体未过的验收条件，置信度维持6。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点未达成、终值假、断点在展平操作中，均可接受；但现象“其他·残留折叠”只是把节点未达成改写成现象，且 418–505 s 帧里毯子已基本平整、左缘仅轻微内凹，“折叠仍在”被写得过强；contact_actual 把大多依次进行的单爪压拖标成双臂协作。",
+    "n_issues": 2,
+    "n_major": 2
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未达成、终值假（评分器 live_success=0、graded_by_live=1）、断点在展平操作中、现象留空并写 gap、接触设计根因标为推断，均与视频和评分器字段相符；第1轮已处理的问题已修正，双臂协作��接触标注有 226–254 s 同时压拖的画面支持，已在上一轮用规则回应，我不重复提出。",
+    "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": "2026-10-31T00:15:53Z",
+  "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/roboarc-blanket-flatten-08-i01-standard--xfhppg7/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--xfhppg7/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--xfhppg7/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--xfhppg7/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "89d5401e51cbf58c1ecc9dc5b38d015d527ed9bc720b469d2be28929f4e32151",
+   "video_sha256": "b050b0546f23c4013223bedcc32b232e78aece863f1fc25caa86cfad7a41f9bf",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/blanket_flatten_08/trajectories/roboarc-blanket-flatten-08-i01-standard--xyxazvd.json

@@ -0,0 +1,250 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-blanket-flatten-08-i01-standard--xyxazvd",
+ "taskKey": "roboarc/blanket_flatten_08",
+ "dag_sha256": "c3fc53614904199eb02e9a01b27e313faf3d74ced543558049fefd1f67ad0c95",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 17945,
+   "live_rows": 17945,
+   "sim_steps": 17945,
+   "physics_steps": 179450,
+   "steps": 8214,
+   "elapsed_s": 10748.5,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.592,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.514,
+   "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/roboarc-blanket-flatten-08-i01-standard--xyxazvd/replay.mp4",
+  "duration_s": 448.733,
+  "fps": 15.0,
+  "frames": 6731,
+  "sha256": "e311cd3ce3de46c08d4e1be5dd3dc0bd620e48ec7c7f39a3ff9632d3cb65ea2e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "blanket_flattened": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的蓝色方格毯存在横贯中部的翻折重叠层，起始未展平。"
+     },
+     {
+      "t": 72.0,
+      "log_step": null,
+      "text_zh": "观察：视频 70.0–75.3 s，机器人左右夹爪分别夹住毯面片区共同拉动，毯面移动且后侧折角仍翘起；视频 88.0–105.0 s，右爪拉开前侧后大部分毯面贴桌，但轮廓仍不规则，不能据此认定联合验收完成。"
+     },
+     {
+      "t": 194.0,
+      "log_step": 258,
+      "text_zh": "观察：视频 193.0–200.0 s，左夹爪抬起撤回时毯子持续随爪离桌并卷成悬垂折叠；log step 258 的机器人控制脚本将这一阶段写为降下、张爪、抬起和撤回，实际未得到干净释放的结果。"
+     },
+     {
+      "t": 448.6,
+      "log_step": null,
+      "text_zh": "观察：视频 447.8–448.6 s，毯子集中在桌面左侧，左边有明显隆起及重叠，未恢复为完整平铺的矩形；评分器字段 success=0、live_success=0、graded_by_live=1，记录仅一次启动且无重置续接。推断：结合全程抽帧、关键阶段密集帧和实时验收未成功，唯一展平节点从未达到完整联合判据，最终为假。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "视频 70.0–75.3 s，左右二指夹爪分别夹住毯子不同片区配合拉开；随后交替以单爪调整，视频 88.0–96.0 s 右爪拉动前侧，视频 186.0–200.0 s 左爪夹持拉动并在撤回时带起毯子。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "coverage_ratio",
+   "unit": "ratio",
+   "threshold": 0.95,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_height_m",
+   "unit": "m",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "minimum_height_m",
+   "unit": "m",
+   "threshold": -0.008,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_stretch_ratio",
+   "unit": "ratio",
+   "threshold": 1.15,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.03,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "release_clearance_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.003,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "confirmation_time_s",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "blanket_flattened",
+   "t": 194.0,
+   "log_step": 259,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 193.0–200.0 s，机器人左爪从较平的毯面抬��撤回，却把毯子带离桌面并形成大幅悬垂折叠；log step 258 的脚本明确安排先降下、张爪再撤离，log step 259 启动该脚本。此后机器人多次尝试解开和拉展，视频 447.8–448.6 s 仍有明显重叠隆起，展平步骤始终未完成。"
+  },
+  "unmet": {
+   "blanket_flattened": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没松开",
+    "t": 194.0,
+    "log_step": 258,
+    "evidence_zh": "观察：log step 258 的释放段要求左爪张开到原生驱动坐标 0.0368 后抬起；视频 193.0–197.7 s 左爪上升时毯面仍紧随末端，视频 198.0–200.0 s 毯子悬垂在左爪上，原先铺开的材料被重新带起卷折。这里的释放失败由毯子的实际随动判断。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "观察：log step 258 的机器人脚本在同一段轨迹中连续执行 lower、open、lift_up、swing_back，完成整段后才读取毯面并检查结果；视频 193.0–200.0 s，毯子未脱离左爪，机械臂仍继续升高、撤回并把毯面卷起。推断：释放与撤离的实际控制未实现所需的脱离，且未在带起毯子时及时修正，导致展平结果显著退化；现有证据不足以进一步归入具体控制子类。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志；评分器 reward 未报告 DAG 的八项连续指标，因此 native_metrics 均为 null，不能用 agent 自写点云估计替代评分器测量。视频能判定明显折叠和撤离带布，但无法精确核对材料伸长、所有粒子的速度、毫米级清距及连续确认窗口，也不能分清释放失败源于夹爪执行、局部缠挂还是接触几何。日志末尾 log step 477 的执行调用没有返回结果，以视频末态为准。"
+ },
+ "summary_zh": "机器人以双爪配合及单爪交替拉动展开毯子，一度使大部分毯面贴桌，但完整展平节点未达成，最终为 0/1。视频约 194.0 s，左爪在释放撤离阶段仍带起毯面，重新造成大幅折叠，后续补救至视频 448.6 s 仍未恢复。没有状态日志及评分器连续指标，释放失败的具体控制子类和毫米级验收量无法确定。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未达成、断点约视频194 s（左爪撤离时带起毯子）、现象“没松开”与根因“控制”均与帧一致；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": "2026-10-30T23:52:21Z",
+  "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/roboarc-blanket-flatten-08-i01-standard--xyxazvd/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-blanket-flatten-08-i01-standard--xyxazvd/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "3f0c734ac5114e511884461fa69b17654d74867e9580a2e509e138f9e2e0b2f1",
+   "video_sha256": "e311cd3ce3de46c08d4e1be5dd3dc0bd620e48ec7c7f39a3ff9632d3cb65ea2e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-privileged--5wxanqn.json

@@ -0,0 +1,219 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-privileged--5wxanqn",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 380,
+   "submitted_actions": 380,
+   "sim_steps": 380,
+   "physics_steps": 3800,
+   "sim_time_s": 9.500000000000068,
+   "final_reward": 0.0,
+   "replay_wall_s": 32.2,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 24,
+   "agent_resets": 27,
+   "agent_sim_steps": 4812,
+   "agent_records": 5,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-buzz-wire-3d-01-i01-privileged--5wxanqn/replay.mp4",
+  "duration_s": 9.6,
+  "fps": 15.0,
+  "frames": 144,
+  "sha256": "5dc36d04c7b2c53ae90e481e48bb70add1fce6d8cf2e55b5d23a228637861c15"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，金属环和木柄工具平放在左侧导线脚旁的起点木台上。"
+     },
+     {
+      "t": 9.53,
+      "log_step": 104,
+      "text_zh": "观察：已检查视频全程的半秒采样及视频 8.8–9.53 s 的逐帧末段；机器人先抬高、平移，再张开夹爪下降接近左侧木柄，工具始终留在起点，未被提起沿上升段、横跨段和下降段移动。log step 104 的最终生成脚本输出记录 380 个动作、success=False，与评分器 n_actions=380 及录像动作相符；此前测试不作为本次节点进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "其他",
+      "note_zh": "推断：视频 9.2–9.53 s，单臂夹爪上部/腕部与导线上方横段相抵，log step 104 的本次录制输出报告接近时接触力 767 N、随后下降时 1519 N 并停止。候选方式为“碰撞”，现有接触方式没有对应项；夹爪保持张开，没有形成对木柄的双侧夹持，不能标抓握或手持工具避开。具体接触连杆及指尖是否另有轻触木柄无法从画面判清。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 9.53,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–9.53 s，工具一直位于左侧起点，右侧终点线脚旁木台没有工具；视频 9.53 s 的主视角和斜视角仍可见起点的金属环。机器人没有携带工具到右端或实施终点释放；因此不满足终点支撑、释放静置的联合验收。参考成功视频 78.0 s 可见工具在右侧线脚旁平躺、夹爪张开退出的外观，仅用于判据对照。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better"
+  },
+  {
+   "name": "finish_angle_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 9.3,
+   "log_step": 104,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 9.3 s，机器人正以竖直张开的夹爪接近起点木柄，夹爪上部/腕部与上方导线重合相抵，工具仍在木台上；到视频 9.53 s 没有抓起或开始遍历。log step 103 的最终脚本以起点抓取作为遍历的准备动作，log step 104 的该次录制在 slideY、descGrip 段报告接触后停止，保存 380 个动作；此后没有修复该步骤。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 9.3,
+    "log_step": 104,
+    "evidence_zh": "推断：视频 8.8–9.53 s，机器人保持竖直夹爪靠近起点，末段夹爪上部/腕部与导线上段相抵，未能完成接近并夹起木柄。对应本次交付的 log step 104 输出先为 descFront 的 0 N，继而 slideY 接触 767 N、descGrip 接触 1519 N 并停止；与末段画面结合支持抓取接近被固定导线阻挡，而非已经抓住工具后滑脱。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "已验证有",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 103 的交付脚本固定采用 top_down_quat 的竖直抓柄姿态，先在前方下降，再保持该姿态横移到柄上方并下降；log step 104 显示这种进场几何在横移、下降阶段与障碍发生接触，ok 变为假后跳过合爪，视频 9.53 s 亦显示工具未被夹住。问题在所选抓取姿态与进场空间不适配；参考成功视频 10.0 s 显示较水平的夹爪接近并随后于参考视频 15–45 s 携工具沿线运动，验证该任务存在替代抓取方案。agent 在 log step 103、105 自述“几何不可行”不足以证明所有合法方案被排除。"
+   }
+  ],
+  "gap_zh": "本次没有状态日志或逐连杆接触对，末段画面不能精确区分夹爪基座与腕部哪个碰撞几何先触线，也不能确认指尖是否轻触木柄；碰撞时间取视频约 9.3 s 的粗导航锚点，未把 agent 工作时间换算为视频时间。评分器未报告 distance_to_finish_m、finish_angle_deg，两个原生指标保持 null。"
+ },
+ "summary_zh": "机器人在起点进行抓柄接近，工具全程留在左侧木台，沿线遍历和终点放置均未完成，最终进度为 0/2。视频末段及 log step 104 的最终录制表明，竖直抓取的进场几何受到上方导线阻挡，脚本停止并跳过合爪；参考成功视频验证存在替代方案，不支持 agent 的全局不可行判断。缺少逐连杆状态日志，具体先触线部位和精确接触时刻仍有不确定性。",
+ "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（均未达成）、断点 wire_traversed、被挡/碰撞现象与「抓取和接触设计（alt_feasible 已验证有）」根因均与视频帧及 log step 104 的 767/1519 N 接触读数一致；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": "2026-10-31T00:11:07Z",
+  "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/roboarc-buzz-wire-3d-01-i01-privileged--5wxanqn/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-privileged--5wxanqn/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0882ace81a20c6099530cfdf619d3f8a0d20ff788b3b190e02dee6c66a30cda4",
+   "video_sha256": "5dc36d04c7b2c53ae90e481e48bb70add1fce6d8cf2e55b5d23a228637861c15",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-privileged--7k3tawp.json

@@ -0,0 +1,234 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-privileged--7k3tawp",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 3138,
+   "submitted_actions": 3138,
+   "sim_steps": 3138,
+   "physics_steps": 31376,
+   "sim_time_s": 78.44000000000071,
+   "final_reward": 1.0,
+   "replay_wall_s": 367.0,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 3138,
+   "strict_replay_required": 0,
+   "agent_sessions": 7,
+   "agent_resets": 47,
+   "agent_sim_steps": 12552,
+   "agent_records": 14,
+   "agent_replays": 4,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.542,
+   "peak_memory_gb": 1.274,
+   "peak_processes": 19,
+   "peak_disk_gb": 0.487,
+   "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/roboarc-buzz-wire-3d-01-i01-privileged--7k3tawp/replay.mp4",
+  "duration_s": 78.533,
+  "fps": 15.0,
+  "frames": 1178,
+  "sha256": "bb08c92ebce04860749758c50898049223ba37daaf0febe8782dc4fc9c49d882"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 2,
+  "final": 2,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [
+     [
+      47.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 10.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 工具位于左侧线脚的木台上；视频 8.0–10.0 s 单臂二指夹爪从木柄两侧合拢，随后木柄和金属环随夹爪抬起，离开原来的台面位置。"
+     },
+     {
+      "t": 38.0,
+      "log_step": null,
+      "text_zh": "观察：视频 14.0、18.0、22.0、26.0、30.0、34.0、38.0、42.0 s 的主视角、斜视角及腕部视角依次显示机器人持环沿左侧上升段、上方横段和右侧下降段前进，导线穿过环的开口。"
+     },
+     {
+      "t": 47.0,
+      "log_step": 93,
+      "text_zh": "推断：视频 46.53–47.60 s 机器人持环沿右侧导线末段下降至终点线脚附近，约视频 47.0 s 进入末端10毫米范围，时间估计误差约0.5秒；log step 93 是所交付首段的执行记录，末项 traversal_complete=True，log step 110 将该段接入完整轨迹并从重置状态回放成功；结合评分器 strict_goal_reached=1，判定连续全线遍历完成。没有逐物理步状态日志，唯一穿线及连续性细阈值由成功判定支持。"
+     },
+     {
+      "t": 48.4,
+      "log_step": 94,
+      "text_zh": "观察：视频 47.8–48.4 s 机器人张开夹爪，工具短距离落在终点木台上；视频 50.0–62.0 s 工具仍留在右侧线脚附近，机械臂空手退开并调整姿态。log step 94 同一落台段的任务读数 traversal_complete=True，后续已完成的遍历保留。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 8.0–10.0 s 单臂二指夹爪从两侧夹住工具木柄，随后持柄抬起并带环沿整根导线移动，直至视频约47.6 s。"
+     },
+     {
+      "who": "手持工具",
+      "how": "避开",
+      "note_zh": "推断：视频 14.0–47.6 s 机器人所持金属环沿导线移动，导线处于环内开口；结合评分器整体成功，判定工具金属环及杆避开固定金属导线。画面不能独立排除所有物理子步的瞬时微接触。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [
+     [
+      78.47,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 48.4,
+      "log_step": 94,
+      "text_zh": "观察：视频 48.0–48.4 s 工具先落到终点线脚旁的浅色木台上，木柄朝向斜偏；视频 50.0–62.0 s 保持这一临时落台位置。此时尚未完成最终朝向与释放静置验收。"
+     },
+     {
+      "t": 73.0,
+      "log_step": 108,
+      "text_zh": "观察：视频 65.0–69.0 s 单臂夹爪重新靠近并夹住木柄；视频 71.0–76.0 s 工具随夹爪转向、移到右���线脚并降低至木台，原先斜偏的木柄改为朝台内。log step 108 记录该放置段，log step 110 将它接入交付轨迹并回放成功。"
+     },
+     {
+      "t": 78.47,
+      "log_step": 109,
+      "text_zh": "观察：视频 77.20、77.33、77.47、77.60、77.73、77.87、78.00、78.13、78.27、78.40、78.47 s 的腕部视角显示金属环平躺在右侧线脚附近的木台上，木柄朝台内，张开的两指逐渐退离，工具相对木台未见移动；末帧主视角也显示工具已脱离夹爪留在台面。"
+     },
+     {
+      "t": 78.47,
+      "log_step": 110,
+      "text_zh": "推断：将一秒共同确认计入该放置节点，在视频末帧约78.47 s 完成验收；评分器 success=1、all_success=1、deterministic=1、replays=2、strict_goal_reached=1、strict_goal_row=3138，支持终点位姿、支撑接触、辅助抓握解除、机器人间隙及静止阈值均已通过。log step 110 的完整3138动作回放成功，与视频所见末态一致；精确阈值及确认起点不能仅由画面量出。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 47.8–48.4 s 机器人先松开木柄让工具短落终点木台；视频 65.0–71.0 s 二指夹爪再次从两侧夹住木柄，随后持柄调整工具位姿并放到台面，视频 77.2–78.47 s 张开并退离。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人约在视频 47.0 s 完成金属环沿整根导线的连续遍历，随后在终点短落台、重新抓取调整朝向，最终于视频约78.47 s 完成工具释放及一秒静置验收，两节点最终均成立。评分器记录两次确定性回放成功，视频末帧可见工具平躺在终点木台、夹爪张开退离，没有失败断点。缺少逐物理步状态日志，微小金属接触、精确位姿及速度阈值由成功判定支持；两项原生指标未在评分器字段中报告，保留为null，未以会话内自测数值代替。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两节点终态成立，遍历约47 s完成、终点放置在视频末帧验收，与帧和评分器字段一致；无失败断点，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": "2026-10-31T00:04:24Z",
+  "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/roboarc-buzz-wire-3d-01-i01-privileged--7k3tawp/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-privileged--7k3tawp/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "3f47845aa4732faee1f5d0ff9f8d1cacc27d3feb0df4654176da63afd0883899",
+   "video_sha256": "bb08c92ebce04860749758c50898049223ba37daaf0febe8782dc4fc9c49d882",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-privileged--dr6f5x5.json

@@ -0,0 +1,224 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-privileged--dr6f5x5",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 5793,
+   "submitted_actions": 5793,
+   "sim_steps": 5793,
+   "physics_steps": 57922,
+   "sim_time_s": 144.80500000001578,
+   "final_reward": 1.0,
+   "replay_wall_s": 671.7,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 5793,
+   "strict_replay_required": 0,
+   "agent_sessions": 0,
+   "agent_resets": 0,
+   "agent_sim_steps": 0,
+   "agent_records": 0,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-buzz-wire-3d-01-i01-privileged--dr6f5x5/replay.mp4",
+  "duration_s": 144.933,
+  "fps": 15.0,
+  "frames": 2174,
+  "sha256": "c6e0babaa5742f9af4f0337596c1425cb89947659c8bfcf267e9ac43b15e9414"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 2,
+  "final": 2,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [
+     [
+      105.6,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 33.0,
+      "log_step": null,
+      "text_zh": "观察：视频 26.0–33.0 s，机器人二指夹爪在左侧起点合拢夹住木柄，工具随后离开原先的木台位置，金属环沿左侧竖直导线上升；视频 42–60 s 环随夹爪经过左侧弯曲段和上方横跨段，视频 72–102 s 继续经过右侧弯曲段并沿右侧导线下降。"
+     },
+     {
+      "t": 105.6,
+      "log_step": 138,
+      "text_zh": "推断：视频 104.8–106.4 s，工具在右侧线脚附近继续下降，随后木柄从夹爪中脱出，环落向终点木台；结合 log step 137 的 traversal_complete=False 与 log step 138 的 traversal_complete=True，将连续遍历首次完成定位为视频约 105.6 s（约半秒不确定性）。这里的日志是最终交付轨迹所包含的下降和释放段，log step 152 的生成器保留此段，log step 176 对同一交付文件的两次回放均报告 traversal_complete=True，且视频动作顺序一致；未将会话早期碰线试验计入本集。"
+     },
+     {
+      "t": 144.8667,
+      "log_step": 176,
+      "text_zh": "观察：视频 106.4–109.0 s 可见工具留在右侧终点木台上，之后机器人仅在终点重新抓取并调整；视频 144.8667 s 工具仍平放在右侧线脚处。log step 176 的两次交付文件回放评估均报告 traversal_complete=True、episode_contact_coverage_complete=True；评分器 reward.success=1、deterministic=1、strict_goal_reached=1。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 29.0–33.0 s，机器人单臂的二指夹爪从木柄两侧闭合夹持，并带动工具上升；之后持柄转动和横移工具，视频约 105.6 s 末段工具脱离夹爪。"
+     },
+     {
+      "who": "手持工具",
+      "how": "避开",
+      "note_zh": "观察：视频 33–102 s，机器人持木柄使金属环沿导线通过上升段、横跨段及下降段，导线穿过环内开口。推断：结合视频末段到达终点及评分器严格成功，工具金属环和杆在本集避开了固定金属导线；视频不能逐物理子步分辨微小接触。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [
+     [
+      144.8,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 109.0,
+      "log_step": 141,
+      "text_zh": "观察：视频 106.4–109.0 s 机器人打开夹爪并上撤，工具留在右侧木台；log step 141 对这一中间静置段的任务评估显示 supported=True、released=True、quiet=True，但 distance_to_finish_m=0.00876035，超出 0.005 m 容差。视频 126–138 s 机器人再次接近并夹住木柄，因��中间落台不记为本节点达成。"
+     },
+     {
+      "t": 144.8,
+      "log_step": 176,
+      "text_zh": "观察：视频 137.0–143.4 s，机器人夹着终点工具进行小幅调整，环和木柄保持在木台附近；视频 143.8–144.8667 s，夹爪张开并上撤，工具独立平躺在右侧线脚附近，木柄朝台内，终帧可见夹爪与工具之间的间隙。其终点外观与参考视频 78.4 s 的已释放终态一致，参考仅用于辨认外观。"
+     },
+     {
+      "t": 144.8,
+      "log_step": 176,
+      "text_zh": "推断：以视频 144.8 s 作为释放、离开并完成静置确认的达成时刻；视频最后一秒工具位置和朝向稳定，精确阈值由本次交付文件的回放评估佐证。log step 176 的两次回放均报告 success=True、supported=True、released=True、quiet=True，distance_to_finish_m=0.00006876051863170293、finish_angle_deg=1.3274555793929863；评分器 strict_goal_row=5793、sim_time_s=144.805 且 strict_goal_reached=1，与视频末端一致。没有用 agent 工作耗时换算视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：工具初次落到右侧木台后，机器人在视频 126–138 s 接近并以二指夹爪重新夹住木柄，随后微调工具落点；视频 143.8–144.8667 s 松开并上撤，使工具留在终点支撑面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "value": 6.876051863170293e-05,
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better"
+  },
+  {
+   "name": "finish_angle_deg",
+   "value": 1.3274555793929863,
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "本次评分轨迹完成全部 2/2 个节点：金属环沿左侧上升段、上方横跨段和右侧下降段走到终点，约视频 105.6 s 完成连续遍历。机器人在终点重新抓取并修正初次落台的偏移，随后释放、撤离，约视频 144.8 s 完成终点静置验收；评分器严格成功，两次交付文件回放的终态一致。无状态日志，遍历达成时刻有约半秒不确定性，逐子步金属避碰、毫米级间隙及速度阈值不能单靠画面复测，依据与视频一致的回放评估和最终评分佐证；原生指标数值取自 run.json 中 log step 176 的交付回放评估字段。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "两个节点的时间与值与视频和评分器一致：约 105.6 s 工具在右侧终点附近脱手完成遍历，144.8 s 终态放置达成（strict_goal_row=5793 对应 144.8 s）；初次落台被正确排除，contact_actual 来自实际所见。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": "2026-10-31T00:06:42Z",
+  "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/roboarc-buzz-wire-3d-01-i01-privileged--dr6f5x5/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-privileged--dr6f5x5/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c4bf33788ff748ef4874704c7aa1afc438aed5d72a0b32b728288b9e896a9853",
+   "video_sha256": "c6e0babaa5742f9af4f0337596c1425cb89947659c8bfcf267e9ac43b15e9414",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-privileged--ltdyukb.json

@@ -0,0 +1,216 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-privileged--ltdyukb",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 340,
+   "submitted_actions": 340,
+   "sim_steps": 340,
+   "physics_steps": 3400,
+   "sim_time_s": 8.500000000000053,
+   "final_reward": 0.0,
+   "replay_wall_s": 37.8,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 9,
+   "agent_resets": 45,
+   "agent_sim_steps": 4737,
+   "agent_records": 15,
+   "agent_replays": 7,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.859,
+   "peak_memory_gb": 1.556,
+   "peak_processes": 15,
+   "peak_disk_gb": 1.031,
+   "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/roboarc-buzz-wire-3d-01-i01-privileged--ltdyukb/replay.mp4",
+  "duration_s": 8.6,
+  "fps": 15.0,
+  "frames": 129,
+  "sha256": "4c90168bba6ac8f716f663140411bee0b2c578bb9fc11ef6ed570ad1bc973079"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：本次视频 0.00 s，木柄金属环平放在左侧起点线脚旁；主视角及斜视角均可见工具，机器人夹爪张开且与工具分离。"
+     },
+     {
+      "t": 2.13,
+      "log_step": 274,
+      "text_zh": "观察：本次视频 0.00–2.40 s 的密集帧中，机械臂抬高并向起点上方移动，工具一直留在原位；视频 2.13 s 夹爪仍悬在工具上方。log step 274 的最终生成代码只命令末端到高处、作小幅下降后保持 260 个控制步，没有闭爪、提起或沿线移动命令。"
+     },
+     {
+      "t": 8.53,
+      "log_step": 280,
+      "text_zh": "观察：本次视频 2.5–8.5 s 的逐半秒帧及视频 7.60–8.53 s 的逐帧检查中，工具始终平放左侧起点，未上升到横跨段或到右端；视频 8.53 s 仍如此。log step 280 对最终 340 步文件的两次回放均返回 success false、progress 0.0、episode_end None，与画面一致。交付动作仅到悬空准备姿态，未启动遍历操作，无实际工具接触可标。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 8.53,
+      "log_step": null,
+      "text_zh": "观察：本次视频 0.00 s、2.13 s 和视频 7.60–8.53 s 密集末帧中，工具均在左侧起点，右侧终点线脚周围没有工具，也没有向右侧支撑面搬运或放置的动作。成功参考视频 78.00 s 则清楚显示金属环平躺在右侧线脚附近、木柄向台内且夹爪已张开退出；该参考仅用于确认可见验收外观。本次未尝试此放置步骤。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 2.13,
+   "log_step": 274,
+   "where": "首步前",
+   "evidence_zh": "观察：本次视频 2.13 s 起机器人停在起点工具上方，直到视频 8.53 s 工具仍原地平躺，没有抓取或沿线遍历。log step 274 的最终脚本只保留高处接近、小幅下降及保持；log step 277、280、286 确认交付的是该 340 步 safe_nominal 文件，而非会话中其他推转或抓取测试。推断：评分尝试在第一步的悬空准备阶段即停止，遍历操作从未启动。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 2.13,
+    "log_step": 274,
+    "evidence_zh": "观察：本次视频 0.00–2.40 s 机器人仅空手抬高、移到工具上方；视频 2.5–8.53 s 保持悬空，工具未离开起点。log step 274 的最终代码没有闭爪或遍历指令，log step 286 标明交付文件为 340 步 safe_nominal；本次评分轨迹没有开始抓持金属环沿线行进。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 69,
+    "evidence_zh": "推断：agent 未能形成可执行的取工具控制方案，最终只交付悬空基线。会话测试的日志有独立支持：log step 69 多个下降和闭爪测试均返回 inhand=None、工具原位不动；log step 124 的接近测试记录机器人夹爪压在导线上并停滞；log step 170 的端部夹取测试仍返回 IN HAND: None；log step 284 的环管夹取测试在闭爪后及上提后工具位置相同。log step 235 的生成脚本说明因这些接近、抓取困难保留基线，log step 274、277 将其生成最终交付。这些会话测试只用于解释控制困难，不作为本次视频的碰撞或接触现象。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "本次没有状态日志；两节点未完成可由工具始终留在起点的画面直接判定。根因依据会话测试输出推断，未提供这些测试各自的视频或逐步接触状态，不能进一步区分接触几何、接近轨迹及跟踪控制问题；agent 自述的“所有抓法都不可行”（log step 289）不足以排除合法替代方案，不据此归因任务或环境。评分字段未报告 distance_to_finish_m 和 finish_angle_deg，原生指标保留 null。"
+ },
+ "summary_zh": "本次评分视频中机械臂只抬高并移到起点工具上方悬停，工具从视频 0.00 s 到视频 8.53 s 始终留在左侧起点，遍历与终点放置均未完成（0/2）。断点为金属环走完全线的首步前，现象为没尝试；会话测试反复未能抓起工具，支持控制层面的根因，最终交付的是 340 步悬空基线。缺少状态日志及会话测试的逐步接触证据，控制子类型无法确定。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "两节点均未达成、断点在首步前、没尝试现象与视频一致；交付的是340步悬空基线（log 286、280），根因控制由会话测试推断并已在gap_zh说明，符合FAIL-22处理。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": "2026-10-31T00:07:34Z",
+  "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/roboarc-buzz-wire-3d-01-i01-privileged--ltdyukb/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-privileged--ltdyukb/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7276e4cc1214b1f7e0b7d903a7508594389236193a2f7836a32deba7d70eebb4",
+   "video_sha256": "4c90168bba6ac8f716f663140411bee0b2c578bb9fc11ef6ed570ad1bc973079",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-privileged--sbvj5bu.json

@@ -0,0 +1,222 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-privileged--sbvj5bu",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 1343,
+   "submitted_actions": 1343,
+   "sim_steps": 1343,
+   "physics_steps": 13423,
+   "sim_time_s": 33.557499999999166,
+   "final_reward": 0.0,
+   "replay_wall_s": 140.4,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 6,
+   "agent_resets": 23,
+   "agent_sim_steps": 3962,
+   "agent_records": 171,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.859,
+   "peak_memory_gb": 1.298,
+   "peak_processes": 10,
+   "peak_disk_gb": 1.125,
+   "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/roboarc-buzz-wire-3d-01-i01-privileged--sbvj5bu/replay.mp4",
+  "duration_s": 33.667,
+  "fps": 15.0,
+  "frames": 505,
+  "sha256": "3ab83b41c6e64e57571a2f35e93f63b3091cb563b0512c597bfc881e710e9dc0"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 16.0,
+      "log_step": 316,
+      "text_zh": "观察：视频 0–14 s 机器人接近左侧起点并合拢夹爪；视频 16–30 s 金属环离开起点木台，随夹爪沿左侧竖直导线上升，木柄始终在夹爪中。"
+     },
+     {
+      "t": 33.6,
+      "log_step": 319,
+      "text_zh": "观察：视频 30–33.6 s 机器人在左上弯段抬升并转动工具，末帧金属环仍在该弯段，未经过上方横跨段和右侧下降段。log step 319 保存1343行轨迹，服务返回 episode_end=wire_touched；run.json 评分器 submitted_actions=1343、deterministic=1，与这次保存及视频终态相符。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 8–14 s 单臂二指夹爪合拢夹住工具木柄；视频 16–33.6 s 工具被抬离木台并随夹爪沿左侧导线上行，末帧仍握持。"
+     },
+     {
+      "who": "手持工具",
+      "how": "其他",
+      "note_zh": "观察：视频 33.5–33.6 s 工具金属环在左上弯段与固定导线相接，log step 319 服务报告 wire_touched。候选名称：碰触；这是非预期金属工具碰线，现有抓握、推/扫、压/擦等方式均不能准确描述，不能标为避开。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 33.6,
+      "log_step": 319,
+      "text_zh": "观察：完整视频 0–33.6 s 中，机器人未把工具送到右侧终点；视频 33.6 s 工具仍夹持在左上弯段，右侧线脚附近木台没有工具。对照成功参考视频 78.4 s 的工具平躺在右侧线脚、夹爪张开离开的外观，本次未出现终点放置或释放。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 33.5,
+   "log_step": 319,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 32.6–33.6 s 机器人仍握着工具在左上弯段转动，金属环靠到固定导线上后运动结束；log step 319 将该次轨迹保存为1343行，服务返回 wire_touched，log step 322–323 确认保留该失败轨迹。全线遍历未完成，之后没有恢复或重走。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 33.5,
+    "log_step": 319,
+    "evidence_zh": "观察：视频 32.6–33.6 s 工具环在左上弯段随腕部转动，末段环边与导线贴合，机器人在到达横跨段之前停止；对应保存轨迹的 log step 319 服务返回 episode_end=wire_touched。碰撞对象是固定金属导线，工具仍被握持，没有掉落。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 312 修改环姿态生成的转角符号与横向偏移，log step 315–319 生成并执行逐段位姿路线；视频 16–30 s 工具稳定随夹爪上升，而视频 30–33.6 s 进入弯段转动后碰线，log step 319 仍报告 OBJECT_IN_HAND wand。证据支持持物运动未保持金属环与导线的安全间隙这一控制问题；不能进一步区分指令轨迹几何与实际跟踪误差。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志或逐物理步接触对，无法独立精确量测穿线点、环线间隙及最先碰触的金属部位；碰线由末段画面与对应保存轨迹的服务终止字段共同支持。日志有多次重置试验，本次只使用 log step 313 重置后、log step 319 保存的轨迹判断进度；此前更远的试验不计入。评分器未报告两个原生距离/角度指标，均保留 null；控制根因的细分缺少证据。"
+ },
+ "summary_zh": "机器人夹住木柄，将金属环从左侧起点沿竖直段抬至左上弯段，未走完全线，也未开始终点放置，最终完成0/2个节点。视频 33.5 s 附近持物转弯时发生金属碰线，log step 319 对应保存轨迹返回 wire_touched，断点为全线遍历；归因到控制层，无法细分轨迹几何与跟踪误差。缺少状态日志和逐子步接触记录，微小间隙与首个接触部位无法精确复核。",
+ "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0/2、断点wire_traversed、被挡/碰撞与控制根因均与视频及log step 319(episode_end=wire_touched，sim_steps 1343=33.56 s，与视频时长吻合)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": "2026-10-31T00:04:48Z",
+  "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/roboarc-buzz-wire-3d-01-i01-privileged--sbvj5bu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-privileged--sbvj5bu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8cc3234c4c0629722c2ba9345a48f67ee64e74e0c8aa2074809895ebfd238554",
+   "video_sha256": "3ab83b41c6e64e57571a2f35e93f63b3091cb563b0512c597bfc881e710e9dc0",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-privileged--zcn4zca.json

@@ -0,0 +1,227 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-privileged--zcn4zca",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 909,
+   "submitted_actions": 909,
+   "sim_steps": 909,
+   "physics_steps": 9081,
+   "sim_time_s": 22.70249999999978,
+   "final_reward": 1.0,
+   "replay_wall_s": 107.0,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 909,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 2,
+   "agent_resets": 97,
+   "agent_sim_steps": 16740,
+   "agent_records": 5,
+   "agent_replays": 3,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.856,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.062,
+   "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/roboarc-buzz-wire-3d-01-i01-privileged--zcn4zca/replay.mp4",
+  "duration_s": 22.8,
+  "fps": 15.0,
+  "frames": 342,
+  "sha256": "94697e77311b237250dcb27ff9dc7ec945f6012e70a937cfadd6c8c563a9fb38"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 2,
+  "final": 2,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [
+     [
+      20.8,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 5.667,
+      "log_step": null,
+      "text_zh": "观察：视频 3.3–5.0 s，单臂夹爪的两指位于木柄两侧并闭合；视频 5.667 s，工具已离开起点木台，随夹爪向上移动。"
+     },
+     {
+      "t": 17.6,
+      "log_step": null,
+      "text_zh": "观察：视频 6.4–17.6 s 的主视角、斜视角及腕部视角中，机器人持柄带动金属环沿左侧上升段、上方横跨段和右侧弯曲段连续前进；视频 18.6–21.2 s 继续沿右侧立段下降，未见中途绕到线外或重新起步。"
+     },
+     {
+      "t": 20.8,
+      "log_step": 176,
+      "text_zh": "推断：视频 20.4–21.2 s，环下降至右侧线脚、接近木台，约视频 20.8 s 进入终点最后10毫米，时间误差约0.4秒；随后视频 22.733 s 环已平放在线脚处。评分器 reward 的 success=1、strict_goal_reached=1、deterministic=1，且 log step 176 报告所保存909动作轨迹回放成功，支持完整连续遍历已被接受；逐物理步唯一穿线和毫米阈值不能仅凭画面直接量出。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 3.3–5.0 s 单臂二指夹爪从木柄两侧闭合，随后握住木柄带动整件工具沿线移动。"
+     },
+     {
+      "who": "手持工具",
+      "how": "避开",
+      "note_zh": "推断：视频 5.667–21.2 s 金属环围绕导线随夹爪连续通过全线，未见碰撞；本次严格回放 success=1 支持所持工具金属部分成功避开固定导线，非接触通过。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [
+     [
+      22.7,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 21.733,
+      "log_step": null,
+      "text_zh": "观察：视频 20.933–21.6 s，金属环降至右侧线脚的浅色木台表面，木柄朝台内；视频 21.6–22.733 s 夹爪张开并向上退开，工具留在终点支撑面，环和木柄未随夹爪上升。其终态外观与参考成功视频 78.0 s 的工具落台、夹爪退开相符；参考录像仅用于外观对照。"
+     },
+     {
+      "t": 22.733,
+      "log_step": null,
+      "text_zh": "观察：已检查视频 21.8–22.6 s 每0.2秒的画面及视频 22.733 s 末帧；工具保持平躺在线脚旁，机器人张开的夹爪已悬于工具上方，工具未再移动。"
+     },
+     {
+      "t": 22.7,
+      "log_step": 176,
+      "text_zh": "推断：结合上述释放、落台画面及 reward 中 strict_goal_reached=1、strict_goal_row=909、success_in_hold_only=0、success=1���将含一秒静置确认的整项验收记为约视频 22.7 s 达成并保留至片尾；log step 172 记录从重置初态录制、保存并回放最终轨迹，log step 176 返回909动作且正式回放成功。精确的位姿误差、速度、辅助抓握解除及5毫米机器人间隙由严格成功判定支持，画面不能独立逐项测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 20.4–21.6 s 单臂二指夹爪仍从两侧握住木柄，将工具降到终点木台，之后张开并向上撤离；视频 22.733 s 工具留在支撑面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人夹住木柄，连续带环走完全线，约视频 20.8 s 达到终点；随后落台、松爪退开，约视频 22.7 s 完成静置验收，两个节点最终均成立，本次严格回放成功。没有状态日志，遍历完成时间约有0.4秒误差，逐物理步穿线及终态毫米级阈值以画面和评分器严格成功字段综合判断；reward 未报告两项原生距离与角度指标，故其数值留空。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "成功运行，两个节点均达成，时间（约20.8 s、22.7 s）与帧一致，接触标注合理，无失败层；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": "2026-10-31T00:07:42Z",
+  "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/roboarc-buzz-wire-3d-01-i01-privileged--zcn4zca/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-privileged--zcn4zca/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "858af4566d2bcb5c1a4901c4f1c998b9ca38f907d2e0b2ba1d908c17b6a4a8c5",
+   "video_sha256": "94697e77311b237250dcb27ff9dc7ec945f6012e70a937cfadd6c8c563a9fb38",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-standard--8bywcpg.json

@@ -0,0 +1,234 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-standard--8bywcpg",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 41,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 19856,
+   "live_rows": 19856,
+   "sim_steps": 19856,
+   "physics_steps": 198560,
+   "steps": 8650,
+   "elapsed_s": 10791.9,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-buzz-wire-3d-01-i01-standard--8bywcpg/replay.mp4",
+  "duration_s": 496.467,
+  "fps": 15.0,
+  "frames": 7447,
+  "sha256": "60fe17f69593d10d91b03ce40eba1ee6768fe8f7d2edec1fd837ad72cd640446"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 4.5,
+      "log_step": 240,
+      "text_zh": "观察：视频 4.0–7.0 s，机器人竖直夹爪停在左侧导线上方，指尖仍高于平躺在木台上的木柄；log step 240 的在线安全监测记录夹爪压在导线上，首次 forcing 事件的 control step 为 178。推断：log step 11 的控制频率为 40 Hz，178/40 = 视频约 4.45 s，已与视频 4.5 s 的接近停滞画面核对；整集 19856 个控制步也与约 496.4 s 的录像运动时长一致。这里的碰撞主体是机器人夹爪，不是工具金属环。"
+     },
+     {
+      "t": 37.0,
+      "log_step": 240,
+      "text_zh": "观察：视频 12.0–37.0 s，机器人夹爪沿左侧上升段、横跨段和右侧下降段运动，但金属环及木柄仍躺在左侧线脚旁；视频 37.0 s 夹爪已到右端，工具仍清楚可见于左端。log step 240 虽输出 Gripper opening after grasp: 19.2 mm 和各段执行完毕，画面没有工具跟随，不能据开度或脚本阶段名计入遍历。"
+     },
+     {
+      "t": 290.0,
+      "log_step": 662,
+      "text_zh": "观察：视频 280.0–299.0 s，机器人再次接近并闭爪，随后转动、抬起夹爪；工具仍留在左侧线脚附近，没有随夹爪走上导线。log step 662 的服务计数为 control_steps=11613、sim_time_s=290.325，和视频 290.0 s 左端闭爪姿态相符。"
+     },
+     {
+      "t": 496.4,
+      "log_step": null,
+      "text_zh": "观察：全片每约视频 5 s 的抽帧及各次接近、抬起附近的加密画面均未见金属环走过整条导线；视频 495.4–496.4 s，机器人空夹爪离开左侧区域，金属环和木柄仍在左端木台上。推断：工具没有完成从起点到右端的连续遍历，节点从未成立。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "机器人 Robotiq 夹爪在左端接近、闭爪和后续沿线动作中压住固定导线；视频 4.5 s 与 log step 240 的 pressing=the gripper、on=the wire 相互支持，log step 703 也记录后续夹爪压线。未见工具被可靠夹持后沿线带走，不把闭爪指令或非零开度当作已建立抓握，也不把实际压线标成避开。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 37.0,
+      "log_step": 240,
+      "text_zh": "观察：视频 37.0 s，机器人夹爪在右侧线脚附近下移，工具却仍躺在左侧起点；视频约 42.0–48.0 s，机器人在右侧张爪、撤离，右端支撑面没有留下工具。log step 240 的 Release、Retreat 和 confirmation watch 是脚本执行阶段，不能证明工具被放到终点。"
+     },
+     {
+      "t": 496.4,
+      "log_step": null,
+      "text_zh": "观察：视频 495.4–496.4 s 的末段画面仍显示工具留在左侧线脚旁，右侧终点线脚附近没有工具；与成功参照视频 78.0 s 中金属环平躺在右端、木柄朝台内且夹爪张开的外观不同。推断：本次录像没有达到终点放置及释放静置的验收条件，节点从未成立。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "无接触",
+      "how": "不适用",
+      "note_zh": "视频 37.0–48.0 s，机器人在右端执行空载下移、张爪及撤离，属于已执行的放置尝试；工具始终在左端，这段没有机器人对工具的搬放接触，也没有工具对终点支撑面的接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 4.5,
+   "log_step": 240,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 4.0–7.0 s，机器人为拿起工具而竖直接近左端，夹爪停在导线上方，木柄仍在其指尖下方的木台上；log step 240 的在线安全监测明确记录夹爪压在导线上，事件 control step 178 对应视频约 4.45 s（40 Hz，已核对视频 4.5 s）。视频 12.0–37.0 s 随后的沿线动作没有带走工具，直至视频 496.4 s 工具仍在左端。推断：首次接近受阻发生于走完全线步骤的抓取准备中，后续重试始终没有完成该步骤，因此以此为首个未修复断点。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 4.5,
+    "log_step": 240,
+    "evidence_zh": "观察：视频 4.0–7.0 s，机器人夹爪在左侧导线上方停滞，没有降到木柄夹持高度；log step 240 的服务安全监测记录 pressing=the gripper、on=the wire，事件最大接触力 344.9 N，接近段 max_press_force_n=714.6 N。后续视频 12.0–37.0 s 可见夹爪空载沿线运动而工具留在左端；这是夹爪与固定导线的阻挡接触，不据此声称工具金属碰线或立即判负。"
+   }
+  ],
+  "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：首次抓取选用的竖直接近几何没有给夹爪本体避开上方导线留出空间。log step 210 的脚本设抓取 TCP 为 [0.3600,-0.2979,0.1630] m、朝下，导线平面 y=-0.3060 m、顶部 z=0.3000 m；log step 251 的同一抓取姿态 FK 给 r85_base=[0.3600,-0.2979,0.3069] m，夹爪基体位置与顶部导线非常接近。log step 240 的真实执行监测证实夹爪压线，视频 4.5 s 则显示指尖还未到木柄；该接近设计使需要的夹持无法在计划位置建立。后续 log step 512、703 仍记录夹爪压线，录像中工具持续留在左端。本次尝试没有验证出可成功完成遍历的替代抓取方案。"
+   }
+  ],
+  "gap_zh": "没有状态日志，也没有本次逐物理步的穿线历史或辅助抓握开关记录；后续部分闭爪瞬间木柄被夹爪遮挡，不能确定是否发生过短暂双侧夹持，因此不标已建立抓握或抓持后滑脱。工具未随夹爪走完全线及未到终点则在主、斜视角可判为未达。评分器字段未报告 distance_to_finish_m、finish_angle_deg，故两项原生指标均留 null；不从失败总分补造数值。"
+ },
+ "summary_zh": "本次在线评分轨迹两个节点均未完成：机器人反复接近、闭爪并执行沿线及终点张爪动作，工具却持续留在左侧起点附近，视频 496.4 s 仍未到右端。断点在视频约 4.5 s 的走完全线步骤中：抓取接近几何使夹爪先被固定导线阻挡，在线安全监测也记录夹爪压线，后续尝试没有修复该步骤。没有状态日志，不能细分后续遮挡处是否短暂夹持或复原逐物理步穿线状态，但两个节点未达的画面证据明确。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个节点未达、断点在视频约 4.5 s 的 wire_traversed、被挡/碰撞与抓取和接触设计根因均与视频、日志相符，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": "2026-10-31T00:11:48Z",
+  "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/roboarc-buzz-wire-3d-01-i01-standard--8bywcpg/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-standard--8bywcpg/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "83758d1147bba21c06849f1752044a17b505f7ba3aeb28c743e0fa2e781d2f63",
+   "video_sha256": "60fe17f69593d10d91b03ce40eba1ee6768fe8f7d2edec1fd837ad72cd640446",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-standard--cadldrs.json

@@ -0,0 +1,230 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-standard--cadldrs",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 1,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 4436,
+   "live_rows": 4436,
+   "sim_steps": 4436,
+   "physics_steps": 44359,
+   "steps": 3227,
+   "elapsed_s": 3569.0,
+   "replay_wall_s": 335.9,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.248,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.419,
+   "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/roboarc-buzz-wire-3d-01-i01-standard--cadldrs/replay.mp4",
+  "duration_s": 111.0,
+  "fps": 15.0,
+  "frames": 1665,
+  "sha256": "dc4effbc199826ca262000430035288a7dc9233c66706a6972619dcaec0429e7"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，金属环平放在左侧起点线脚周围，木柄朝台内；工具尚未遍历导线。"
+     },
+     {
+      "t": 13.0,
+      "log_step": 35,
+      "text_zh": "观察：视频 10.0–14.0 s，机器人用收拢的夹爪侧推木柄，工具绕左侧线脚转动，金属环仍在起点木台上；log step 35 的推转脚本及前后测量与此动作一致。"
+     },
+     {
+      "t": 110.9,
+      "log_step": 71,
+      "text_zh": "观察：全片粗查并核对视频 100.27–110.93 s，机器人反复靠近起点木柄，金属环始终未走上横跨段或右侧下降段；视频 110.4–110.93 s 才随夹爪抬离起点木台，仍在左侧立线附近。log step 71 的在线机器人服务状态为 done=true、reason=wire_touched、resumes=0；评分器 live_matches_replay=1。全线遍历未达到。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "观察：视频 10.0–14.0 s，Robotiq夹爪收拢后侧推木柄，使工具绕起点线脚转向；这段是推转，未见夹住抬起。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 110.4–110.93 s，木柄位于二指之间，工具随夹爪上移、离开起点木台；末次夹取确实带起工具，早先的小幅试提未见相同结果。"
+     },
+     {
+      "who": "手持工具",
+      "how": "其他",
+      "note_zh": "观察：视频 110.8–110.93 s，夹爪所持工具的金属环缘靠到左侧立线，log step 71 的机器人服务确认 wire_touched。候选接触方式为“碰触”；现有方式不表达这种非预期金属碰线，不能用“避开”代替。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 110.93,
+      "log_step": 71,
+      "text_zh": "观察：视频 0.0 s 至 110.93 s 的主视角与斜视角中，右侧终点线脚附近一直没有工具；末帧工具仍由夹爪带在左侧起点上方。未发生向终点搬运、终点释放或静置验收，因此该节点未达到且未开始操作。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 110.8,
+   "log_step": 70,
+   "where": "探索",
+   "evidence_zh": "观察：视频 110.4–110.93 s，机器人从起点木柄处上移并带起金属环，环缘向左侧立线靠拢，随后画面结束；log step 70 的 tipcal.py 是夹指高度校准动作，先保持闭爪目标0.79把TCP升至0.24 m，再计划移到台内探测桌面，而非执行沿线轨迹。log step 71 的在线服务确认 wire_touched 并终止。推断：失败发生在遍历步骤的夹取/校准探索阶段，全线遍历未完成，终点放置尚未开始。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 110.8,
+    "log_step": 71,
+    "evidence_zh": "观察：视频 110.4–110.93 s，机器人闭爪抬起工具，金属环离开木台后环缘贴向起点立线；腕部视角可见环缘与导线间隙消失。log step 70 返回 EpisodeOver: wire_touched，log step 71 的机器人服务状态再次确认该金属触线终止原因；不是仅由失败分数推断碰撞。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "推理/规划",
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent在抓取结果尚未可靠确认时，安排了保持闭爪的校准撤离，未先解除可能的持物接触，导致本应独立于工具的探测动作带起工具碰线。依据是log step 66的小幅试提后工具高度统计仍在台面附近，log step 67继续以0.79闭爪目标下降，log step 70的校准脚本没有张爪步骤，直接保持0.79上移并计划横移；视频 110.4–110.93 s实际显示工具被带起。agent在log step 72、75事后也自述误以为未抓住、应先张爪。该归因针对动作顺序和持物前提未核实，不把辅助抓握生效时刻当作已知事实。"
+   }
+  ],
+  "gap_zh": "没有状态日志或逐物理子步接触记录，无法独立确定辅助抓握何时生效、触线的精确子步及首先接触的是环缘还是相邻金属杆；金属触线事实由本次单episode的机器人服务确认，视频支持其发生于最后抬起动作。两个节点均可凭可见位置判为未达，不能把agent此前对试提的判断当作实际未持物的证明。"
+ },
+ "summary_zh": "机器人只在左侧起点推转工具并反复试夹，未完成全线遍历，也未开始终点放置，节点最终为0/2。视频 110.4–110.93 s，机器人在夹指高度校准前保持闭爪上移，实际带起工具碰到起点导线；log step 71的在线服务以wire_touched终止，校准动作前未先解除可能持物的顺序安排是有日志支持的根因。没有状态日志，辅助抓握建立时刻及金属首次接触部位不能精确复核。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个节点未达、final=0、断点 wire_traversed 与视频和 log step 70/71 一致；wire_touched 为服务对本次单 episode 的读数。碰线现象与推理/规划根因均有证据支持，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": "2026-10-31T00:11:17Z",
+  "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/roboarc-buzz-wire-3d-01-i01-standard--cadldrs/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-standard--cadldrs/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "6d8c114a17634966fcfbddcce01b8c567a5c495d6520d5b7d0156f8189476b54",
+   "video_sha256": "dc4effbc199826ca262000430035288a7dc9233c66706a6972619dcaec0429e7",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-standard--mkgyqmv.json

@@ -0,0 +1,235 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-standard--mkgyqmv",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 18,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 9904,
+   "live_rows": 9904,
+   "sim_steps": 9904,
+   "physics_steps": 99035,
+   "steps": 1522,
+   "elapsed_s": 5949.6,
+   "replay_wall_s": 728.7,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.303,
+   "peak_processes": 9,
+   "peak_disk_gb": 1.038,
+   "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/roboarc-buzz-wire-3d-01-i01-standard--mkgyqmv/replay.mp4",
+  "duration_s": 247.667,
+  "fps": 15.0,
+  "frames": 3715,
+  "sha256": "c8a1c47331222bbaacfa727e81e4de77bb4f1142702a5582cad0c06229a874f4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的金属环围在左侧起点线脚，木柄平放木台；视频 0–240 s 的粗查中机器人反复接近、撤回与调整夹爪，工具未走过上方横跨段。"
+     },
+     {
+      "t": 244.0,
+      "log_step": 369,
+      "text_zh": "观察：视频 242.0 s 工具仍在起点附近；视频 244.0–246.4 s 木柄夹在机器人两指之间并随夹爪上升，原来线脚旁的工具已离开台面，环升向左侧竖直段顶部。此前抓取调整后来已修复，不能把早期空夹当作最终失败。"
+     },
+     {
+      "t": 247.6,
+      "log_step": 423,
+      "text_zh": "观察：视频 246.8–247.6 s 机器人在左上弯折处旋转所持工具，末帧仍在该处，未沿横跨段走向右侧、更未下降到终点；log step 423 的在线机器人服务观测为 done=true、reason=wire_touched、control_steps=9904。评分器字段 live_matches_replay=1、resumes=0。"
+     },
+     {
+      "t": 247.6,
+      "log_step": 425,
+      "text_zh": "推断：log step 18 的服务读数 control_steps=161、sim_time_s=4.025 给出 40 Hz；log step 425 的转弯提交结束读数 9901/40=247.525 s，log step 423 的结束读数 9904/40=247.6 s，已与视频 246.6–247.6 s 的左上转弯及最后一帧核对。这是本次在线受评集的控制步映射，不使用 agent 工作时间；全线遍历未达成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 244.0–247.6 s 单臂二指夹爪夹住工具木柄，木柄与金属环随夹爪升起并在左上弯折处转向。"
+     },
+     {
+      "who": "手持工具",
+      "how": "其他",
+      "note_zh": "观察：视频 247.6 s 工具仍由夹爪握持于左上弯折处；log step 423 的机器人服务报告 wire_touched，确认所持工具金属部分触碰固定导线。候选名称“金属触碰”：并非推扫、压擦或成功避开；现有画面不能分清触线的是环缘还是金属杆。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 工具位于左侧起点，右侧终点线脚附近的木台没有工具；此放置条件在开局不成立。"
+     },
+     {
+      "t": 247.6,
+      "log_step": 423,
+      "text_zh": "观察：视频 244.0–247.6 s 工具一直在夹爪中向左上弯折移动，末帧仍由机器人持于左侧，右侧终点木台始终空着；没有向终点放置或松手静置的动作。对照成功参照视频 78.4 s，同一主视角清楚显示工具平躺于右侧线脚、夹爪张开退出；本次未出现该终态。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 247.6,
+   "log_step": 423,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 246.8–247.6 s 机器人握着工具在起点竖直段的左上弯折处转向，尚未走过横跨段；log step 409 提交转弯动作后进行保持观测，log step 425 显示转弯提交结束于控制步 9901，log step 423 的服务最终观测在控制步 9904 报告 done=true、reason=wire_touched。推断：按已核对的 40 Hz 控制步映射，碰线结束约为视频 247.6 s；后续横跨命令在 log step 422 被 EpisodeOver 拒绝，该步骤再未修复。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 247.6,
+    "log_step": 423,
+    "evidence_zh": "观察：视频 246.8–247.6 s 机器人持工具在固定导线左上弯折处转向，金属环附近被手指部分遮挡；本次在线机器人服务在 log step 423 明确报告 wire_touched，表示工具金属部分与固定金属导线发生接触并结束该集。工具未走到横跨段或终点，未见脱手坠落；以这次固定导线碰撞作为单一现象。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：agent 的脚本在 log step 406、408 根据估计的环心相对夹爪偏移生成半径 18 毫米、旋转 90° 的九个转弯目标，log step 409 一次提交并随后保持观察；视频 244.0–246.4 s 已显示机器人能够夹持工具沿竖段上移，但视频 246.8–247.6 s 的持工具转向最终由服务在 log step 423 确认为碰线。该段精细位姿与金属间隙控制未能保持避线，归入控制大类；没有足够证据进一步归为抓取设计、轨迹或逆解、反馈控制或参数调试。"
+   }
+  ],
+  "gap_zh": "没有逐物理步状态或接触对日志；视频 246.8–247.6 s 的环缘及金属杆部分被夹爪遮挡，具体触线部位由画面不能确定，碰线事实依据该次在线服务的 wire_touched 结果。未记录真实工具位姿和穿点轨迹，不能区分环心几何估计偏差、转弯路径偏差与执行跟踪偏差，根因仅标控制大类。评分器未报告 distance_to_finish_m 和 finish_angle_deg，原生指标值留 null。"
+ },
+ "summary_zh": "机器人经多次抓取调整后夹起工具，沿左侧起点竖直段升至上弯折，但未走过横跨段，两个目标节点均未完成（0/2），终点放置未开始。视频约 247.6 s 的持工具转向期间发生金属碰线，在线服务以 wire_touched 结束，断点为全线遍历的操作中，根因仅能归到精细位姿与间隙控制。末段金属部分被夹爪遮挡且没有物理步接触日志，不能确定触线部位或更细的控制失效来源。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个节点均未达成、断点 wire_traversed 操作中、被挡/碰撞 + 控制的标注与视频末段及 log step 409/423/425/426 的服务读数一致，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": "2026-10-31T00:13:10Z",
+  "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/roboarc-buzz-wire-3d-01-i01-standard--mkgyqmv/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-standard--mkgyqmv/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "204359feb67f83ac3c5c0f451313a68f61844226ca73ddfceaafb8bbba6b215b",
+   "video_sha256": "c8a1c47331222bbaacfa727e81e4de77bb4f1142702a5582cad0c06229a874f4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-standard--p3u7k6p.json

@@ -0,0 +1,221 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-standard--p3u7k6p",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 28,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 43669,
+   "live_rows": 43669,
+   "sim_steps": 43669,
+   "physics_steps": 436690,
+   "steps": 7782,
+   "elapsed_s": 8428.1,
+   "replay_wall_s": 3681.1,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.385,
+   "peak_processes": 10,
+   "peak_disk_gb": 1.19,
+   "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/roboarc-buzz-wire-3d-01-i01-standard--p3u7k6p/replay.mp4",
+  "duration_s": 1091.8,
+  "fps": 15.0,
+  "frames": 16377,
+  "sha256": "0b1213d3de6ec448f288f87abde386a394e5d526f86005d06dba482fa16840c5"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，工具的金属环平放在左侧起点线脚周围，木柄放在木台上；机器人夹爪空着。"
+     },
+     {
+      "t": 262.0,
+      "log_step": null,
+      "text_zh": "观察：视频 260.0–265.0 s，机器人夹爪向起点木柄靠近、闭合后回撤，腕部视角中的金属环和木柄仍留在木台上，没有随夹爪提起。视频 330.0–369.5 s 的密集画面再次显示指端接近、抵住木柄后离开，工具仍在起点。"
+     },
+     {
+      "t": 1091.7,
+      "log_step": 366,
+      "text_zh": "观察：视频全程每约 10 s 的粗采样中未见金属环沿左侧上升段、上方横跨段及右侧下降段通过；视频 1090.8–1091.7 s 的末秒画面中，环仍平放在左侧线脚，机器人空爪离开工具。log step 366 的在线服务状态为 done=false、reason=null；评分器 success=0、live_matches_replay=1。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频 350.0–355.0 s，单臂夹爪的指端抵在起点工具木柄上方并闭合，随后回撤；木柄未进入可承载的对向夹持，工具没有随爪提起。实际为指端对木柄的抵压、擦碰。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1091.7,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 与视频 1091.7 s 的主视角、斜视角均显示工具在左侧起点，右侧终点线脚附近没有工具；全片粗采样中未见机器人持工具移向终点或在终点释放。对照成功参考视频 78.0 s，终点放置应可见环平躺在右侧线脚周围、木柄朝台内，该外观在本次末帧中明确不存在。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better"
+  },
+  {
+   "name": "finish_angle_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 262.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 260.0–265.0 s，机器人在起点尝试夹取木柄，闭爪后空爪回撤，环没有被带起。此后仍反复处理起点抓取，视频 354.0–357.0 s 再次出现同样结果；直到视频 1091.7 s 工具仍在起点，沿线遍历未完成。断点时间取一次清楚可见的空夹回撤作导航锚点。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 262.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 261.0–265.0 s 及视频 350.0–357.0 s，机器人指端靠到木柄、闭合再上移，金属环和木柄均留在木台上，没有出现工具随夹爪离台的阶段。所需的可提起工具的夹持接触未建立，因而无法开始沿线移动。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "evidence_zh": "推断：视频 350.0–355.0 s 的腕部近景和视频 354.0 s 的主、斜视角显示，机器人指端抵在木柄上方，闭合时未从木柄两侧形成承载夹持，随后回撤而工具留台。log step 216 中 agent 脚本命令闭爪并上抬，服务反馈夹爪开度为 0；log step 314 又逐渐闭合至开度 0 后上抬，视频全程仍未见工具被提起。接近目标后所用指端与木柄的接触几何没有建立有效抓握，是沿线动作无法启动的直接原因。",
+    "source_zh": "视频 / 日志"
+   }
+  ],
+  "gap_zh": "没有状态日志、逐物理步接触对或辅助抓握激活记录，无法复核每次尝试的指垫接触细节；抓取接触设计根因由可见几何、空爪回撤与日志中的闭爪命令联合推断。评分器未报告 distance_to_finish_m 和 finish_angle_deg，原生指标值保留 null。"
+ },
+ "summary_zh": "机器人反复在起点接近、抵压木柄并闭爪，但未把工具从木台提起；金属环走完全线和工具放到终点两项均未达到，最终为 0/2。断点在沿线遍历的取工具阶段，现象为未建立有效夹持，证据指向抓取接触几何不适配；视频 1091.7 s 工具仍在左侧起点。没有状态日志及逐步接触对，接触设计根因属于结合视频与命令反馈的推断。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两节点均未达成、工具全程留在左侧起点，与视频一致；断点、没建立接触、抓取和接触设计根因均有帧与日志支撑，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": "2026-10-31T00:02:48Z",
+  "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/roboarc-buzz-wire-3d-01-i01-standard--p3u7k6p/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-standard--p3u7k6p/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "ae65f557824e7c5ee53fdbeba5c3c3bc8b867b48970a47e2bbe4d90a0bf34d08",
+   "video_sha256": "0b1213d3de6ec448f288f87abde386a394e5d526f86005d06dba482fa16840c5",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-standard--wbkz9rs.json

@@ -0,0 +1,231 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-standard--wbkz9rs",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 34,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 20087,
+   "live_rows": 20087,
+   "sim_steps": 20087,
+   "physics_steps": 200870,
+   "steps": 4130,
+   "elapsed_s": 9553.7,
+   "replay_wall_s": 1322.6,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-buzz-wire-3d-01-i01-standard--wbkz9rs/replay.mp4",
+  "duration_s": 502.267,
+  "fps": 15.0,
+  "frames": 7534,
+  "sha256": "2065cb1a375fe3a73e51ec231761565d0f835a6989ff20975d62023dc8b608f4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的主视角和斜视角中，工具平躺在左侧线脚旁，金属环套在起点，木柄朝台内；右侧终点没有工具。"
+     },
+     {
+      "t": 8.9,
+      "log_step": 67,
+      "text_zh": "观察：视频 8.9 s 夹爪在导线上方闭合，工具仍躺在左侧木台，木柄未进入指间；视频 5.0–22.0 s 的密集帧也没有工具随手抬起。log step 66 的抓取命令返回 TCP 高度约 0.314 m，而该脚本的目标高度为 0.1615 m；log step 67 的服务读数为 control_steps=358。推断：控制频率由 log step 34 的服务规格给出为 40 Hz，358/40=8.95 s，与视频 8.9 s 的空夹姿态吻合；该换算仅用于在线控制步与视频核对，不使用 agent 工作时间。"
+     },
+     {
+      "t": 212.0,
+      "log_step": null,
+      "text_zh": "观察：视频 40.0–75.0 s、210.0–234.0 s 的密集帧及全片约每隔 10 s 的检查中，机器人反复改变夹爪位置和姿态，工具的环与木柄一直留在左侧线脚旁；视频 212.0 s 腕相机可清楚看见工具仍在木台上，未被夹持。未出现沿左侧上升段、上横段、右侧下降段的工具连续移动。"
+     },
+     {
+      "t": 501.8,
+      "log_step": 266,
+      "text_zh": "观察：视频 495.0–502.0 s 的密集检查及视频 501.8 s 的斜视角中，工具仍在左侧起点，机器人在其上方转腕，右侧线脚为空；log step 266 的服务最终状态为 control_steps=20087、done=false、reason=null、resumes=0，评分字段 live_matches_replay=1。推断：这是同一个未重置在线 episode；虽有保存的求解脚本，实际评分动作未完成遍历。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "其他",
+      "note_zh": "观察：候选方式“空夹／未夹持的抓取尝试”。机器人 Robotiq 二指夹爪在工具木柄上方或侧前方接近、闭合并撤离（视频 8.9 s、50.0 s、212.0 s），没有看见双指夹住木柄或工具随夹爪运动；因此不标抓握。现有方式没有空夹这一项；瞬时轻触是否发生无法由局部遮挡的画面确定，也不把计划中的手持工具避线当作已发生。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 工具在左侧起点，右侧终点支撑面为空，初态不满足终点放置。"
+     },
+     {
+      "t": 501.8,
+      "log_step": null,
+      "text_zh": "观察：全片粗采样和视频 495.0–502.0 s 的密集帧均没有机器人将工具送到右侧；视频 501.8 s 斜视角仍能看见左侧线脚旁的工具，右侧终点没有工具。参考成功录像视频 78.3 s 可��工具平躺于右侧线脚旁且夹爪张开离开，此可见结果在本次录像中从未出现。推断：本次未开始终点放置，也未达到释放静置验收。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 8.9,
+   "log_step": 67,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 8.9 s 机器人执行首次抓取后，夹爪闭合在木柄上方，工具仍留在左侧起点；此后视频 50.0 s、212.0 s 及最后视频 501.8 s 都未见工具被拿起或沿线推进。log step 66–69 的接近与闭合命令未把 TCP 降到抓取目标，后续 log step 264–265 又在抓取前的转腕中返回 FLIP FAIL／FLIP2 FAIL。推断：最早未修复的失败属于金属环走完全线步骤中的工具抓取准备，而不是尚未尝试的终点放置。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 8.9,
+    "log_step": 67,
+    "evidence_zh": "观察：视频 8.9 s 夹爪闭合时木柄仍平躺在下方台面，视频 50.0–63.0 s 夹爪再次接近和撤走时工具没有随手抬起；视频 210.0–234.0 s 腕相机仍看见完整工具留在起点，视频 501.8 s 斜视角也未见其被拿走。推断：遍历所需的夹爪与木柄的有效夹持始终未建立；不能把夹爪闭合读数当成抓住工具，也没有已抓住后丢失的证据。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 256,
+    "source_zh": "日志",
+    "evidence_zh": "观察：log step 66 的抓取目标 TCP 高度为 0.1615 m，服务返回实际约 0.314 m，log step 69 再次下降后仍为 0.2901 m、位置误差 129.9 mm；log step 100 的实际与指令关节位置不一致且多个关节力矩达到限幅。log step 255 的脚本数值检查还显示姿态插值在比例 0.05 时转过 118.79°，log step 256 修正 Rodrigues 公式括号后相同比例变为 4.05°；修正后 log step 264–265 的转腕仍未执行到抓取朝向，返回 FLIP FAIL／FLIP2 FAIL。推断：接近和转腕的控制实现及执行未能把夹爪送到有效抓取位姿，解释了视频 8.9 s、50.0 s、212.0 s 中一直未夹住工具的现象；根因停在控制，未把 agent 对具体碰撞体或关节限位的猜测当作已验证事实。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 中有完整运行记录、评分录像和成功参考录像，但没有本次状态日志或逐对接触记录；部分接近时指尖与木柄接触边界被夹爪遮挡，无法排除瞬时轻触，也无法确认每次停滞的具体碰撞体。推断：这些缺口不影响工具一直留在起点、两个节点均未达成的判断，但不足以区分接触布局、碰撞、关节限位和跟踪误差在各次停滞中的贡献，故不增加碰撞现象或更细的控制根因；评分器未报告两项原生终点误差指标，值保留 null。"
+ },
+ "summary_zh": "机器人反复接近左侧木柄、闭合夹爪和调整转腕，但评分录像始终没有工具随手抬起或沿线遍历，终点放置也未开始，最终完成 0/2 个节点。断点为金属环走完全线步骤的抓取准备：有效夹持未建立，日志显示下降不到位、姿态插值代码错误及后续转腕执行失败；保存的脚本不等于在线 episode 已执行其遍历与放置段。没有状态与逐对接触日志，无法进一步核实各次停滞的具体阻挡体。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "视频抽帧确认工具全程留在左侧起点，两节点未达成，断点、现象与根因均有日志和画面支持，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": "2026-10-31T00:02:45Z",
+  "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/roboarc-buzz-wire-3d-01-i01-standard--wbkz9rs/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-standard--wbkz9rs/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0e4083c332f9c3faaa07fc27557333fc01549daced9a8c72129e5654cfb26a82",
+   "video_sha256": "2065cb1a375fe3a73e51ec231761565d0f835a6989ff20975d62023dc8b608f4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_01/trajectories/roboarc-buzz-wire-3d-01-i01-standard--xmq4htk.json

@@ -0,0 +1,222 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-01-i01-standard--xmq4htk",
+ "taskKey": "roboarc/buzz_wire_3d_01",
+ "dag_sha256": "d23bc0e98b6a63fe1839e0af463b5396eda975513d4942a86d72433aeb161750",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 1483,
+   "live_rows": 1483,
+   "sim_steps": 1483,
+   "physics_steps": 14823,
+   "steps": 166,
+   "elapsed_s": 822.4,
+   "replay_wall_s": 93.3,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-buzz-wire-3d-01-i01-standard--xmq4htk/replay.mp4",
+  "duration_s": 37.133,
+  "fps": 15.0,
+  "frames": 557,
+  "sha256": "08f3b83a1c9f8cbc07e4732154544de55b032d1170cbc00a62582e5836111015"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "wire_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 22.0,
+      "log_step": null,
+      "text_zh": "观察：视频 20.0–28.0 s，机器人二指夹爪夹住木柄，工具随夹爪上升，金属环由左侧线脚附近沿左侧直线段向上移动。"
+     },
+     {
+      "t": 37.067,
+      "log_step": 44,
+      "text_zh": "观察：视频 32.0–37.067 s，机器人持工具在左上第一个弯段转动，结束时金属环仍在该弯段附近；没有沿上方横跨段到达右侧下降段和终点。log step 44 中在线机器人服务返回 done=True、reason=wire_touched，表示金属工具触线导致本集终止；运行评分字段 live_matches_replay=1、resumes=0 与本次录像一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 22.0–37.067 s，机器人单个二指夹爪夹持木柄，木柄和金属环一起随夹爪抬升、转动。"
+     },
+     {
+      "who": "手持工具",
+      "how": "其他",
+      "note_zh": "观察：视频 36.0–37.067 s，夹爪所持工具的金属环在固定导线左上弯段与导线贴近；log step 44 的机器人服务报告 wire_touched。候选“非预期碰触”：现有方式没有碰撞接触项，终止时实际发生金属工具碰线，不能记成避开。"
+     }
+    ]
+   },
+   "tool_on_finish_support": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 37.067,
+      "log_step": 44,
+      "text_zh": "观察：视频 0.0–37.067 s，右侧终点线脚附近木台上始终没有工具；视频 37.067 s 工具仍由机器人夹持在左上弯段，未进行终点放置、松爪撤离或静止确认。成功参考录像视频 78.467 s 可见工具平躺在右侧终点线脚旁且夹爪张开，本次录像没有出现该外观。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10.0,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wire_traversed",
+   "t": 37.0,
+   "log_step": 44,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 36.0–37.067 s，机器人仍夹持木柄，在左上第一个弯段继续旋转金属环，尚未走到上方横跨段；log step 44 的在线机器人服务在 first_bend_exit0 返回 done=True、reason=wire_touched，随后本集停止。该次碰线发生于沿线遍历操作中，后续没有恢复或完成遍历。"
+  },
+  "unmet": {
+   "wire_traversed": false,
+   "tool_on_finish_support": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 37.0,
+    "log_step": 44,
+    "evidence_zh": "观察：视频 36.0–37.067 s，机器人将夹持的金属环转向左上导线弯段，环与导线贴近；log step 44 中机器人服务明确报告 reason=wire_touched 并结束本集，证实工具金属部分与固定金属导线发生接触。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 37.0,
+    "log_step": 42,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 34、39、42 中 agent 的脚本把环心沿拟合导线路线采样，并将环的朝向随弯段旋转，以逆解生成关节位置序列；log step 44 的服务反馈却在该过弯动作中报告 wire_touched，视频 36.0–37.067 s 也显示工具仍停留在首个弯段。持工具过弯时的位置、朝向与净空控制未能维持无碰线条件，归于控制；没有足够证据进一步区分标定误差、接触几何或执行跟踪子类。"
+   }
+  ],
+  "gap_zh": "没有状态日志或逐物理步接触对，视频分辨率不足以辨明终止时是环缘哪一点或金属杆触线；碰线事实由单次在线 episode 的服务 reason=wire_touched 确认。现有记录不足以进一步拆分控制根因，也未报告 distance_to_finish_m、finish_angle_deg 的数值。"
+ },
+ "summary_zh": "机器人夹持木柄，带金属环从左侧线脚沿直线段上升，在左上第一个弯段碰线终止；全线遍历与终点释放放置均未完成，最终为 0/2 节点。视频 36.0–37.067 s 与 log step 44 的服务 wire_touched 反馈共同支持断点，根因仅归于持工具过弯的控制失败。缺少状态日志及接触对，不能进一步确定触线部位或控制误差的具体来源。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标签与视频和日志一致：工具被夹持上升并在左上第一个弯段碰线终止（log step 44 wire_touched），两节点均未达成，断点、现象和停在“控制”的根因均有证据支持。",
+    "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": "2026-10-31T00:11:07Z",
+  "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/roboarc-buzz-wire-3d-01-i01-standard--xmq4htk/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-01-i01-standard--xmq4htk/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "455b93e4a325b07d2e45ea73d3826a2dc275f8429b9e403f8bac97dcfbfa1770",
+   "video_sha256": "08f3b83a1c9f8cbc07e4732154544de55b032d1170cbc00a62582e5836111015",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-privileged--bxvpklc.json

@@ -0,0 +1,263 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-privileged--bxvpklc",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 9023,
+   "submitted_actions": 9023,
+   "sim_steps": 9023,
+   "physics_steps": 90230,
+   "sim_time_s": 225.57500000003415,
+   "final_reward": 0.0,
+   "replay_wall_s": 1252.8,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 1,
+   "agent_resets": 21,
+   "agent_sim_steps": 0,
+   "agent_records": 0,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.864,
+   "peak_memory_gb": 4.111,
+   "peak_processes": 25,
+   "peak_disk_gb": 0.296,
+   "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/roboarc-buzz-wire-3d-10-i01-privileged--bxvpklc/replay.mp4",
+  "duration_s": 225.667,
+  "fps": 15.0,
+  "frames": 3385,
+  "sha256": "1cdfcfc4325474b79b0e6b5cd3ef3afa9bb0c69c9bc4932ad2ecabc0e60c8038"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 13.0,
+      "log_step": null,
+      "text_zh": "观察：视频约9–19 s，单臂夹爪夹住木柄，工具离开起端台面并沿起始立柱向上移动；随后抽查全程各视角，工具经过弯线和绝缘套，但没有到达另一端下降立柱的底端。"
+     },
+     {
+      "t": 219.0,
+      "log_step": 141,
+      "text_zh": "观察：视频210–225.6 s的密集帧显示，机器人在螺旋段附近降低工具、张开夹爪并后撤，工具挂在空中的黑色绝缘套上摆动后停下；视频225.6 s仍未到终端立柱，完整通过未成立。log step 141明确执行降低、松手、后撤及等待，log step 144保存9023行，log step 148保存的仍为9023行部分轨迹，与视频末态相符。"
+     },
+     {
+      "t": 225.6,
+      "log_step": 159,
+      "text_zh": "观察：log step 159中agent自述交付只走到约1.16 m、全线约1.488 m；这只是与视频所见中途悬挂一致的会话说明，不据此赋予通过节点完成时间。评分字段n_actions=9023、deterministic=1、strict_goal_reached=0。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约9–19 s夹爪两指夹住木柄并带动工具上升；视频约59–73 s及约137–161 s可见松手挂线、调整机械臂再夹住木柄，随后继续带着环沿线移动；视频约218–225.6 s最终松手，工具留在中途绝缘套上。工具本身的路径是受评对象，按直接抓握记一次接触链。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 225.6,
+      "log_step": null,
+      "text_zh": "观察：视频全程粗采样及末段210–225.6 s密集帧中，工具没有被带到终端支撑台平放；最后一帧木柄与金属环仍悬在线的中途，终点台面没有工具。故放置条件从未成立，也未见机器人开始该终点放置动作。"
+     },
+     {
+      "t": 225.6,
+      "log_step": 148,
+      "text_zh": "观察：成功参考inputs/oracle/replay.mp4约255 s可见工具沿终端下降，视频343.9 s工具已平躺在支撑台、环位于终端脚边；本段视频225.6 s缺少这些可见关系。log step 148交付说明亦明确末段与终点放置未解出；参考仅用于识别完成线索，不作为本次完成的证据。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 219.0,
+   "log_step": 141,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约217–221 s机器人在螺旋段附近将工具挂到绝缘套并松开后退，直至视频225.6 s仍停在该处，没有继续到终端。log step 141记录这次有意挂线与撤离，log step 144、148保存的是到此为止的9023行轨迹；尚在处理沿线通过步骤，终点放置尚未启动。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 145,
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent未解出可从当前挂线姿态继续的抓取与机械臂运动。log step 142–143对该姿态搜索侧向抓取，一侧无解、另一侧候选关节4约为-3.07 rad且限位余量接近零；log step 145的多组离线接续轨迹在沿线位置约1.17–1.20 m出现超出脚本阈值的位置误差，多组关节4同样约为-3.07 rad。log step 148保存部分轨迹，log step 159明确报告末段未解出，与视频225.6 s中途悬挂相符。这些是会话求解失败证据，不能据此声称交付视频实际发生关节顶限或碰撞。",
+    "note_zh": "未验证当前挂线姿态下的可行抓取与连续运动，证据不足以区分接触几何设计与轨迹/逆解算法，保留控制大类。"
+   }
+  ],
+  "gap_zh": "观察：inputs/没有逐帧状态日志；视频不能核实毫米级通过误差、子步金属接触或最终静止指标，但工具未到终端及未放到台面清楚可见。交付在有意挂线后结束；log step 142–147、151–156中的碰撞、限位和逆解失败属于会话规划测试，没有进入这份交付轨迹，故现象层留空，不把有意松手写成没保持住或把测试写成视频中的够不着。根因由会话求解记录推断，尚不能区分抓取设计与求解器不足，也未排除其他合法方案。"
+ },
+ "summary_zh": "机器人抓住木柄，沿起始立柱、弯线及螺旋段移动并在绝缘套处换握，最后将工具挂在中途后撤离；两个节点均未达成，终态0/2。断点为金属环走完全线，日志显示后续抓取与连续运动求解未成功，因此只谨慎归入控制；交付未执行后续失败测试，现象层留空。没有状态日志，无法细核子步接触和毫米级指标，但视频末秒清楚显示工具仍悬挂、终点台面未放置。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "两节点均未达成、断点 loop_traversed、末态悬挂与视频一致；现象层留空并给出 gap_zh、根因保守归入控制，均有日志与视频支持，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": "2026-10-30T23:45:36Z",
+  "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/roboarc-buzz-wire-3d-10-i01-privileged--bxvpklc/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-privileged--bxvpklc/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "66a81a336f3e8997d4f3d2788a3c48551eaff9da85d4946f542f34be3a858223",
+   "video_sha256": "1cdfcfc4325474b79b0e6b5cd3ef3afa9bb0c69c9bc4932ad2ecabc0e60c8038",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-privileged--pnwtcaa.json

@@ -0,0 +1,262 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-privileged--pnwtcaa",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 942,
+   "submitted_actions": 942,
+   "sim_steps": 942,
+   "physics_steps": 9420,
+   "sim_time_s": 23.54999999999973,
+   "final_reward": 0.0,
+   "replay_wall_s": 106.1,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 1,
+   "agent_resets": 11,
+   "agent_sim_steps": 0,
+   "agent_records": 3,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-buzz-wire-3d-10-i01-privileged--pnwtcaa/replay.mp4",
+  "duration_s": 23.6,
+  "fps": 15.0,
+  "frames": 354,
+  "sha256": "c50d647fcf3b971b3796b307af3ff4236928d7a50de5264da91b8beea6603ce1"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，金属环位于画面左前方起端立柱脚边，木柄平放在木质台面上，机器人夹爪尚在右侧上方。全程粗查并在视频 17.0–23.53 s 密查，机器人只接近并夹合木柄，环未沿起端立柱上升，更未经过上方弯线、螺旋段和终端下降段。"
+     },
+     {
+      "t": 21.07,
+      "log_step": 176,
+      "text_zh": "观察：视频 20.0 s 夹爪在柄两侧，视频 21.07 s 两指已向内夹住手柄，环仍在起端脚边；视频 22.6–23.53 s 保持该位置。log step 176 的同一组 942 动作执行记录报告 in_hand=wand、sim_steps=942；log step 177–178 将该抓取动作数组保存为交付文件，log step 231 最终仍检查到 942 动作。"
+     },
+     {
+      "t": 23.53,
+      "log_step": 234,
+      "text_zh": "推断：视频 23.53 s 环仍在起端，结合 log step 234 中 agent 明确自述交付文件仅有抓取阶段，完整连续通过未达成；会话中的离线沿线航点搜索没有作为视频进度计入。成功参考视频 236.0 s 已将环带至另一端黑色绝缘下降柱，提供终端位置的视觉对照。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约 20.0–21.07 s，单臂 Robotiq 二指夹爪从木柄两侧合拢并夹住工具手柄，至视频 23.53 s 保持夹合；金属环仍留在起端。该接触属于沿线移动的抓握准备。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 工具在左前方起端脚边，终端为画面中部另一根黑色下降立柱；工具并未处于终点休止位置。"
+     },
+     {
+      "t": 23.53,
+      "log_step": null,
+      "text_zh": "观察：视频全程至视频 23.53 s，工具均未被搬到终端立柱脚边，末段机器人仍在起端夹着木柄，没有终点放置、松手和撤离动作。成功参考视频 344.4 s 可见环和柄已在终端黑色立柱脚边，夹爪打开；本次没有出现该终点放置线索。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 21.07,
+   "log_step": 178,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 21.07 s 机器人已夹住起端木柄，随后至视频 23.53 s 环始终停在起端脚边，没有启动沿立柱向上的通过运动。log step 177–178 保存的是抓取阶段动作，log step 231 确认最终交付仍为该 942 动作文件；log step 234 中 agent 自述通过和放置均未执行。此处以抓握准备结束后的停留作为导航锚点，断在金属环走完全线这一步。"
+  },
+  "phenomena": [],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent 的离线轨迹生成未产出可执行的完整沿线动作，导致评分轨迹止于抓握准备。log step 195、203–204 的本地求解输出多次给出 why=ik，log step 206、209–213 又在连续航点求解和避障检查间反复失败；log step 226 与 232 的后续输出仍只有截至 k=125 的部分航点。log step 231 检查最终交付仍为 942 动作，log step 234 中 agent 明确自述螺旋和终段未完成、保存文件仅覆盖抓取。上述是会话中轨迹生成能力的证据，支持控制类归因；未将本地检查中的碰撞预测当成评分视频中的碰撞。",
+    "note_zh": "完整握姿方案的沿线可行性未验证，且日志同时包含几何近似、候选搜索和逆解问题，因此保留控制总类，不强分子类。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 提供本次视频和完整会话记录，但没有本次状态日志；评分器 reward 未报告 DAG 所列八项原生终态指标，故其值均为 null。视频只包含沿线移动前的接近与抓握准备，没有实际沿线运动中的脱线、碰撞、滑落或到位失败，失败现象层留空（CASES.md FAIL-23）；不能把会话离线求解失败当成评分轨迹的运动现象。日志支持控制类归因，但不能验证完整接触方案是否可行或进一步确定控制子类。",
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  }
+ },
+ "summary_zh": "评分轨迹只完成了起端工具的接近与抓握准备，金属环始终停在起端，走完全线与终点放置均未达成，最终为 0/2。断点在金属环走完全线的操作中：会话中的离线轨迹求解和避障搜索未生成完整可执行动作，最终交付仍是抓取阶段文件。视频没有展示沿线运动中的具体物理失误，现象层留空；无状态日志及原生终态指标，控制子类亦不能确定。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达、断点（loop_traversed，操作中）、现象层留空（FAIL-23）与控制类归因均有视频和日志支持；视频只含接近与抓握，环始终在起端，与标注一致。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": "2026-10-30T23:45:04Z",
+  "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/roboarc-buzz-wire-3d-10-i01-privileged--pnwtcaa/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-privileged--pnwtcaa/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "ac3f44aec77eba7e840949773ed9014b3c1b2b0c1f8d3c54e3b3986715ab36e1",
+   "video_sha256": "c50d647fcf3b971b3796b307af3ff4236928d7a50de5264da91b8beea6603ce1",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-privileged--qzf74sy.json

@@ -0,0 +1,264 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-privileged--qzf74sy",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 4000,
+   "submitted_actions": 4000,
+   "sim_steps": 4000,
+   "physics_steps": 40000,
+   "sim_time_s": 100.00000000000561,
+   "final_reward": 0.0,
+   "replay_wall_s": 448.1,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 1,
+   "agent_resets": 16,
+   "agent_sim_steps": 0,
+   "agent_records": 59,
+   "agent_replays": 6,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.547,
+   "peak_memory_gb": 1.355,
+   "peak_processes": 8,
+   "peak_disk_gb": 0.893,
+   "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/roboarc-buzz-wire-3d-10-i01-privileged--qzf74sy/replay.mp4",
+  "duration_s": 100.067,
+  "fps": 15.0,
+  "frames": 1501,
+  "sha256": "55934b3f805c361246386d2d7d90028ac4e5a5ed145965bead738a51b939f3d4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：本次视频起始工具平躺在靠前的起端立柱脚边；全程每隔约4秒的外部及腕部画面显示机器人主要在工具上方调姿，工具没有沿立体弯线前进。"
+     },
+     {
+      "t": 99.0,
+      "log_step": 354,
+      "text_zh": "观察：视频85–98.5 s机器人下降接近木柄，视频99–100 s逐0.1秒检查及视频100 s末帧可见二指夹住木柄外端，金属环仍在原起端脚边，未上行至首个弯折，更未到另一端；log step 354保存的GRASP_CHECKPOINT为4000行。"
+     },
+     {
+      "t": 100.0,
+      "log_step": 386,
+      "text_zh": "观察：log step 386将最终交付重新保存为4000行prefix_success，log step 389–390明确自述仅在起点握住工具、尚未遍历；评分器报告n_actions=4000、sim_time_s约100、deterministic=1，与视频停在起端相符。后续会话试验的wire_touched没有包含在该交付中。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频98.5–100 s单臂二指夹爪在起端闭合夹住工具木柄外端；工具留在起端，只有抓取准备，没有沿线移动。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 100.0,
+      "log_step": null,
+      "text_zh": "观察：本次视频0–100 s工具始终在原起端区域，末帧仍在夹爪中，机器人没有将其送到另一端立柱脚边，也没有松手撤离。成功参考视频344.4 s的工具位于另一端脚边且张开的夹爪已抬离，说明终点放置在这些视角中有可辨认的线索；本次未出现该状态。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": null,
+   "log_step": 386,
+   "where": "首步前",
+   "evidence_zh": "观察：视频末段98.5–100 s机器人仅完成起端木柄夹持，环仍留在起端；全片没有开始沿线通过的动作。log step 386最终选回仅含抓取准备的4000行检查点，log step 390亦自述before traversal。因此断在金属环走完全线的沿线动作启动前，无法定位单一物理失误时刻。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": null,
+    "log_step": 386,
+    "evidence_zh": "观察：视频0–100 s显示接近、调姿和最终闭爪准备；视频99–100 s工具仍在起端，未开始沿铜色线向上通过。log step 386及389–390确认交付截止于抓取检查点。这里的没尝试仅指沿线遍历动作未启动，机器人已做抓取准备。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "轨迹或逆解",
+    "source_zh": "日志",
+    "evidence_zh": "推断：会话中未获得可安全执行的连续沿线轨迹，解释了最终交付为何停在起端。log step 363与366的逆解返回失败，修改环绕线方向后log step 372仍有逆解失败；log step 379–380得到的76个关节目标间有最高约4.76弧度的跳变，log step 384还显示靠前目标间约1.3弧度的关节变化。log step 382–383执行其中前段的独立会话试验以wire_touched结束，随后log step 386回存抓取检查点。这些是会话控制证据，不是本次评分视频中的碰撞现象。"
+   }
+  ],
+  "gap_zh": "无逐帧状态日志，普通画面不能精确验证毫米公差或金属子步接触；但工具停在起端、未完成遍历及终点放置可直接辨认。交付没有沿线动作，故没有可定位的沿线物理失误时刻；轨迹或逆解根因由会话诊断与回存记录推断，不能进一步确定试验碰线的具体接触几何。DAG列出的八项原生终态指标均未在评分器reward字段报告，value保留null。"
+ },
+ "summary_zh": "评分轨迹只完成起端工具的夹持准备，金属环未开始沿线遍历，也没有终点放置，最终0/2节点。断点为金属环走完全线的首步前，现象为没尝试；会话中的逆解失败、不连续关节目标及沿线试验碰线支持控制层的轨迹或逆解问题，最终交付已回存为抓取检查点。无状态日志，无法细查子步接触与毫米级指标，但不影响两个未达节点的判断。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两节点均未达成与视频一致：视频99–100 s起端二指夹住木柄、环留在起端，全程无沿线移动；交付为log step 354/386保存的4000行抓取检查点，断点loop_traversed首步前、没尝试、轨迹或逆解（标为推断）可接受。",
+    "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": "2026-10-30T23:46:42Z",
+  "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/roboarc-buzz-wire-3d-10-i01-privileged--qzf74sy/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-privileged--qzf74sy/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "77cd31daedf4c1acbd0f395ed1a8591418c35bdc82f7bd42c7f983c074df9c2f",
+   "video_sha256": "55934b3f805c361246386d2d7d90028ac4e5a5ed145965bead738a51b939f3d4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-privileged--rpshzmh.json

@@ -0,0 +1,271 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-privileged--rpshzmh",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 13777,
+   "submitted_actions": 13777,
+   "sim_steps": 13777,
+   "physics_steps": 137768,
+   "sim_time_s": 344.4199999999606,
+   "final_reward": 1.0,
+   "replay_wall_s": 1591.3,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 13777,
+   "strict_replay_required": 0,
+   "agent_sessions": 3,
+   "agent_resets": 8,
+   "agent_sim_steps": 44508,
+   "agent_records": 3,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-buzz-wire-3d-10-i01-privileged--rpshzmh/replay.mp4",
+  "duration_s": 344.533,
+  "fps": 15.0,
+  "frames": 5168,
+  "sha256": "062d5789d6c4a9ee02e8b7eca422217db6556e57f4850be4acea4f5b4a540719"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 2,
+  "final": 2,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [
+     [
+      238.5,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 14,
+      "log_step": null,
+      "text_zh": "观察：视频 0 s 工具在起端台面，视频 6–14 s 单臂二指夹爪合拢包住木柄并将工具提起；视频 20–230 s 的外部和腕部视角显示金属环随夹爪沿立体线依次经过深弯、上段及螺旋，随后进入绝缘终端下降段。"
+     },
+     {
+      "t": 238.5,
+      "log_step": 141,
+      "text_zh": "观察：视频 228–238.6 s 机械臂带环沿终端立柱下降至脚边，视频 238.6–242 s 开爪暂放；log step 141 的模拟器原生历史字段 traversal_complete=True、coverage_complete=True，log step 143 记录完整通过前缀有9543个控制动作。"
+     },
+     {
+      "t": 238.5,
+      "log_step": 143,
+      "text_zh": "推断：log step 143 的交付前缀按40 Hz运行，9543/40=238.575 s；已与视频 238.4–238.8 s 终端下降停止并开始开爪的画面核对，首次完整通过约标在视频 238.5 s（约半秒不确定性）。log step 175、193 保留该前缀后续重放，log step 286、287 最终13777动作重放成功且终态哈希一致，评分字段 deterministic=1、all_success=1，支持整条交付轨迹通过原生连续性与金属避线验收；随后脱线不撤销该历史节点。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 6–14 s 单臂Robotiq二指夹爪从两侧夹住工具木柄，随后持柄控制金属环沿线运动至终端下降段。"
+     },
+     {
+      "who": "手持工具",
+      "how": "避开",
+      "note_zh": "推断：视频中的夹爪所持金属环沿铜线及绝缘段通过，结合 log step 141 的有效完整通过与最终严格重放成功，工具金属部分完成了对裸露金属线的避开；普通画面不能独立排除逐物理子步的微小接触。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [
+     [
+      344.4,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 261,
+      "log_step": 170,
+      "text_zh": "观察：视频 240–267 s 机械臂暂放并再夹住工具调整，视频 270–300 s 工具留在终端脚边而机械臂撤离、重新接近；log step 170 原生验收字段虽为 supported=True、released=True，但 finish_angle_deg=164.54、stable_time_s=0，暂放尚未达到本节点。"
+     },
+     {
+      "t": 330,
+      "log_step": 256,
+      "text_zh": "观察：视频 300–330 s 夹爪重新夹持工具，将柄转向台内并降低至终点附近；视频 330–342.8 s 夹爪仍包围木柄，因此不在到达目标位姿时提前记为完成。"
+     },
+     {
+      "t": 344.467,
+      "log_step": 264,
+      "text_zh": "观察：视频 343.0–344.467 s 夹爪张开并抬起撤离，工具留在终端立柱脚边木台上，最后一秒逐帧可见工具未随夹爪抬起。log step 264 的原生验收器 task.result.metrics 报告 traversal_complete、supported、released、quiet 均为True，assist_released=True、robot_clear=True，位置误差0.0025548856 m、角度误差4.464494度、净距0.0207472182 m，stable_time_s=0.9999999999999897，模拟器已判成功；八项 native_metrics 均取此处原生验收输出。"
+     },
+     {
+      "t": 344.4,
+      "log_step": 287,
+      "text_zh": "推断：log step 262、265 保存最终动作，log step 286 新进程从初态重放13777动作成功，log step 287 验证与原运行的终态哈希一致，且评分器 strict_goal_reached=1、strict_goal_row=13777。13777/40≈344.425 s 与视频最后的撤离画面一致，故将完整一秒联合确认后的节点到达时刻标为约视频 344.4 s；毫米级距离与速度漂移依赖原生验收，非画面直接测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 240–267 s 及视频 300–342.8 s，单臂二指夹爪先后包住木柄，暂放、换抓并旋转调整至终点休止方向；视频 343.0–344.467 s 松开并撤离，工具留由终点木质支撑台承托。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00255488557441961
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": 4.464494086522143
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": 0.020747218152017518
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": 4.970946739017081e-14
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": 1.5648109352756198e-13
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": 1.904440938856148e-12
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": 5.431723650462988e-12
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": 0.9999999999999897
+  }
+ ],
+ "failure": null,
+ "summary_zh": "观察：机器人在视频约238.5 s 完成金属环沿全线通过，随后经过终点暂放、换抓和转柄，于视频约344.4 s 松手撤离并通过一秒联合静置验收，两个节点最终均完成；评分器 success=1、deterministic=1，log step 286、287 也确认最终交付重放成功且终态哈希一致。推断：成功来自连续沿线操控及终点姿态、释放和净距均获原生验收，早期会话碰线与受阻试验已被替换，不是交付轨迹的失败。观察：无状态日志，普通视频不能独立核验毫米级公差、逐子步接触及速度漂移，这些依据 log step 264 的原生验收输出与严格重放结果；本运行本身是oracle，没有独立参考录像。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个节点的到达时刻（238.5 s、344.4 s）与 log step 143/264/286/287 及评分字段一致，接触标注合理，成功运行无失败层，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": "2026-10-30T23:52:54Z",
+  "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/roboarc-buzz-wire-3d-10-i01-privileged--rpshzmh/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-privileged--rpshzmh/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "5c1ba2641d1c6bd69ef9c9b0e9becbf83c59e4a031001300e0d66249e1505893",
+   "video_sha256": "062d5789d6c4a9ee02e8b7eca422217db6556e57f4850be4acea4f5b4a540719",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-privileged--xscu9qf.json

@@ -0,0 +1,246 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-privileged--xscu9qf",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 10,
+   "submitted_actions": 10,
+   "sim_steps": 10,
+   "physics_steps": 100,
+   "sim_time_s": 0.24999999999999997,
+   "final_reward": 0.0,
+   "replay_wall_s": 12.7,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 207,
+   "agent_resets": 115,
+   "agent_sim_steps": 37519,
+   "agent_records": 1,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-buzz-wire-3d-10-i01-privileged--xscu9qf/replay.mp4",
+  "duration_s": 0.333,
+  "fps": 15.0,
+  "frames": 5,
+  "sha256": "9e2c12fdb9f1c63a613031332fcfaa2970b22e60552e60adee195c1240c45099"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 42,
+      "text_zh": "观察：已检查评分视频全部5帧（视频0.000、0.067、0.133、0.200、0.267 s），机械臂夹爪始终张开并停在木台右侧上方，未靠近或夹持环柄工具；工具一直位于起端立柱脚边，没有沿铜色弯线移动。log step 42 的脚本明确录制 hold(10) 并保存为交付文件；log step 46 重放该占位文件，返回 n_actions=10、progress=0.0，与评分器字段及本视频一致。"
+     },
+     {
+      "t": 0.267,
+      "log_step": null,
+      "text_zh": "观察：评分视频0.267 s的末帧仍是初始布局；成功参考视频235 s则可见夹爪已带着工具来到另一端下降立柱附近。推断：本评分轨迹没有发生完整沿线通过，不能把会话中其他测试的运动计为本节点进度。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.267,
+      "log_step": null,
+      "text_zh": "观察：评分视频0.000—0.267 s的全部5帧中，工具保持在画面左前方起端立柱脚边，终端下降立柱脚边没有出现放置工具的动作或工具；夹爪始终远离工具。成功参考视频344 s在终端脚边可见工具和夹爪近景，而本轨迹没有相应终点放置变化。推断：工具从未到终点支撑，放置及松手静置验收均未成立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 0.0,
+   "log_step": 42,
+   "where": "首步前",
+   "evidence_zh": "观察：评分视频0.000—0.267 s全部5���均显示夹爪停在工具之外，工具留在起端，机器人没有启动抓取或沿线运动。log step 42 保存的是仅含 hold(10) 的占位轨迹，log step 46 已重放为10个动作、progress=0.0；最终评分器同样报告 n_actions=10、strict_goal_reached=0、deterministic=1。推断：实际交付停在金属环沿线移动这一步开始之前。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.0,
+    "log_step": 42,
+    "evidence_zh": "观察：评分视频0.000、0.067、0.133、0.200、0.267 s均未见夹爪接近工具或工具沿线运动；log step 42 的录制段只执行 hold(10)，并由 log step 46 重放确认。推断：本评分轨迹根本没有启动首步，后续会话测试不属于这份交付。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "观察：本输入没有状态日志；评分器未报告DAG所列8项原生几何、净距及静止指标，故这些指标值为null。log step 898 的完整执行脚本仅在成功时保存替换交付，log step 939 显示另一次会话测试因 wire_touched 失败，最后 log step 963 仍在等待新的测试，没有最终完成报告。推断：可以确定最终保留了早期占位轨迹，但这些会话记录不足以将未形成有效交付唯一归因于某一控制、规划或保存操作缺陷，根因留空；会话中的碰线不作为评分视频的失败现象。"
+ },
+ "summary_zh": "评分轨迹只有10步保持动作，全部5帧中夹爪远离工具，工具留在起端，沿线通过和终点放置均未完成，最终0/2。断点为金属环走完全线的首步前，现象是没尝试；log step 42、46确认交付内容是早期占位录制。后续会话测试未体现在评分视频中，无法可靠确定占位轨迹未被有效替换的具体根因，且没有状态日志或原生公差指标。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "评分视频仅5帧（约0.27 s），夹爪始终张开远离工具，首末帧布局不变；交付为 log step 42 保存、step 46 重放的 hold(10) 占位轨迹，与评分器 n_actions=10、progress=0 一致。两节点均未达成，断点为首步前+没尝试符合03规则（交付基线从未启动该步；会话超时由智能体负责，非外因），根因留空并有 gap_zh，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": "2026-10-30T23:46:30Z",
+  "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/roboarc-buzz-wire-3d-10-i01-privileged--xscu9qf/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-privileged--xscu9qf/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "93f0589069799472266b5b1207df536cc6fc9846c5517a1ff71f76db9f99b4cb",
+   "video_sha256": "9e2c12fdb9f1c63a613031332fcfaa2970b22e60552e60adee195c1240c45099",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--7attetu.json

@@ -0,0 +1,277 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--7attetu",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 4,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 4915,
+   "live_rows": 4915,
+   "sim_steps": 4915,
+   "physics_steps": 49141,
+   "steps": 1126,
+   "elapsed_s": 1763.2,
+   "replay_wall_s": 378.2,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.33,
+   "peak_processes": 10,
+   "peak_disk_gb": 1.157,
+   "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/roboarc-buzz-wire-3d-10-i01-standard--7attetu/replay.mp4",
+  "duration_s": 122.933,
+  "fps": 15.0,
+  "frames": 1844,
+  "sha256": "8dc20bf7097d14b3734dfde488ecbee1e1775f5ebf0497062138f479e103dfe6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "观察：视频 0 s 的工具平放在起始立柱脚边，机器人尚未夹住木柄；实时服务在 log step 8 返回 control_steps=0、done=false。"
+     },
+     {
+      "t": 95,
+      "log_step": 58,
+      "text_zh": "观察：视频 82–95 s，机器人夹爪握着木柄上提，环离开起始台面，沿第一根竖直线段上升；视频 100–120 s，工具继续绕过第一个顶部弯折，尚未走过后面的 U 形段、绕柱段与终端下降段。"
+     },
+     {
+      "t": 122.867,
+      "log_step": 84,
+      "text_zh": "观察：视频 121.8–122.867 s，夹爪仍握柄在第一个顶部弯折之后转动工具，环仍在高处，未到终端立柱脚边；log step 84 的实时服务执行记录 clock_handle_for_u 返回 done=true、reason=wire_touched。评分字段 live_matches_replay=1、deterministic=1、resumes=0，与视频的单次连续尝试一致。"
+     },
+     {
+      "t": 122.867,
+      "log_step": null,
+      "text_zh": "推断：全段视频仅出现起段上升及首个顶弯附近的移动，未出现参考成功视频 230–240 s 所示的沿终端绝缘立柱下降；因此完整通过事件从未成立，而非仅由总分为零判断。参考录像仅用于辨认终点的可见形态。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 82–95 s，Robotiq 二指夹爪夹住工具木柄，带动金属环离开起端并上升；视频 100–122.867 s 持续握柄带环过首个顶弯并转柄，未见工具脱手。"
+     },
+     {
+      "who": "手持工具",
+      "how": "其他",
+      "note_zh": "观察：log step 84 的实时服务返回 wire_touched，视频 121.8–122.867 s 对应夹爪持柄转动工具的末段。候选名称“意外碰触”：所持工具的金属部分触及裸露金属线；现有方式没有一般碰触项，且没有依据认定这是推/扫或压/擦。精确接触部件和物理子步不可见。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0 s，工具在起端铜色竖线脚边，终端黑色绝缘立柱脚边没有工具，终点放置初始不成立。"
+     },
+     {
+      "t": 122.867,
+      "log_step": 84,
+      "text_zh": "观察：全段视频 0–122.867 s 未见机器人把工具放到终点支撑处；最后一秒夹爪仍在首个顶部弯折后握柄，工具悬在高处，终端黑色立柱脚边保持空置。对照参考成功视频 320–344.4 s，终点放置时环在该黑柱脚边、柄平躺于支撑台，本次没有出现该关系，亦未开始终点松手放置动作。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 122.8,
+   "log_step": 84,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 121.8–122.867 s，机器人持柄在第一顶部弯折之后调整工具朝向，尚未进入后续 U 形路线；log step 83 下发 clock_handle_for_u 转柄动作，log step 84 的实时服务返回 done=true、reason=wire_touched，随后没有继续运动或修复。推断：碰线发生于这一末段转柄动作，视频约 122.8 s 为导航锚点；失败步骤是仍在进行的全线通过，不是尚未开始的终点放置。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 122.8,
+    "log_step": 84,
+    "evidence_zh": "观察：视频 121.8–122.867 s，夹爪持有的环柄工具在顶部线段附近转动，录像结束时仍在该处；log step 84 的实时服务记录明确返回 wire_touched、done=true，即工具金属部分与固定裸线发生碰触并触发停止。本次是一次无重置实时 episode，评分字段 live_matches_replay=1；此接触依据服务读数，而非 agent 的失败自述。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 77–82 的 agent 脚本计算多个转柄角度的逆解，log step 83 将约 60 度的转柄调整作为进入 U 形段的准备并以小关节增量执行；视频 121.8–122.867 s 可见机器人实际持柄转动，但 log step 84 的服务随即报告 wire_touched。该操作没有保持工具金属部分与裸线的间隙，故归于控制；缺少实际工具位姿与逐子步接触，不能进一步断定为轨迹/逆解、反馈跟踪或工具变换估计中的哪一种误差。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 提供了完整视频和运行记录，但没有状态日志。末段小环及线在画面中细小、局部被夹爪遮挡；实时服务确认了 wire_touched，却没有提供接触对、逐子步时刻或末段实测环位姿，因而碰触锚点仅定位到视频约 122.8 s，根因保留在“控制”层。评分字段未报告 DAG 所列的八个原生终态指标，其值均填 null。"
+ },
+ "summary_zh": "机器人夹住环柄工具，沿起始直线抬升并通过第一个顶部弯折，但没有走完后续路线，两个节点最终均未达成（0/2）。视频约 122.8 s，机器人在为 U 形段转柄时发生金属工具碰裸线，实时服务在 log step 84 以 wire_touched 结束本次尝试，断点为全线通过操作中的碰撞，根因归于控制。无状态日志和精细接触对，无法进一步确定碰触部件及具体控制误差。",
+ "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（0/2）、断点 loop_traversed 在首个顶弯后转柄、碰撞现象（log step 84 wire_touched，服务读数）及停在“控制”的根因均与视频和日志一致；时钟均为视频时间，未见换算的 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": "2026-10-30T23:45:48Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--7attetu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--7attetu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0364b22529f14c1973d3573282607b8ade23a0471592d3fbeb460fbfe15ce2e2",
+   "video_sha256": "8dc20bf7097d14b3734dfde488ecbee1e1775f5ebf0497062138f479e103dfe6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--a3mz6ca.json

@@ -0,0 +1,269 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--a3mz6ca",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 32,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 10472,
+   "live_rows": 10472,
+   "sim_steps": 10472,
+   "physics_steps": 104720,
+   "steps": 1685,
+   "elapsed_s": 10807.9,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-buzz-wire-3d-10-i01-standard--a3mz6ca/replay.mp4",
+  "duration_s": 261.867,
+  "fps": 15.0,
+  "frames": 3928,
+  "sha256": "7a419274bd70b322637c456dc8622ab09c4e0fe9634325624880884483d2c685"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，金属环套在近侧起端立线脚部，木柄平放在台面，机械臂尚未抓持工具。"
+     },
+     {
+      "t": 45.0,
+      "log_step": 79,
+      "text_zh": "观察：视频 44.5–47.0 s，机器人在木柄上方偏离的位置合爪并抬升，金属环和木柄仍留在起端台面，没有随夹爪升起。推断：首次抓取未建立携带工具所需的夹持；log step 79 的服务 control_steps=1783、log step 81 的1825，按 log step 4 的40 Hz对应视频44.575 s、45.625 s，已在这段合爪抬升画面核对；未把agent工作时间换算为视频时间。"
+     },
+     {
+      "t": 198.0,
+      "log_step": 158,
+      "text_zh": "观察：视频 182.5–200.0 s，机器人多次把指端降到木柄附近、合爪和抬升；视频 200.0 s 夹爪升起后，金属环和木柄仍在原来的起端台面。log step 158 的agent脚本打印WELD False；log step 159 随后仅凭爪间距自述几乎肯定已抓住，画面不支持工具被带走。"
+     },
+     {
+      "t": 261.8,
+      "log_step": 160,
+      "text_zh": "观察：视频 206.0、220.0、230.0、240.0、250.0、258.0及260.87–261.8 s，机器人末端在空间中移动，而金属环仍在近侧起端立线脚部、柄仍平放台面；未出现工具沿弯线到终端的连续运动。推断：完整通过事件从未成立，不能把log step 160开始执行的沿线脚本当作环已通过。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频 182.5–198.0 s，单臂Robotiq夹爪的指端反复下探并压到起端工具的木柄上缘附近，随后合爪、抬起，工具未跟随。推断：实际可辨接触为指端压触，未见两侧指垫可靠夹住并携带木柄；不把空夹或爪间距直接记作抓握。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 200.0,
+      "log_step": null,
+      "text_zh": "观察：视频 200.0 s，机器人抬起后，工具仍平放在近侧起端，黑色绝缘终端立柱脚边没有放好的环柄工具。"
+     },
+     {
+      "t": 261.8,
+      "log_step": 160,
+      "text_zh": "观察：视频 260.87、261.2、261.6、261.8 s，工具一直在起端，机器人末端仍在上方移动，未把工具搬到终端支撑处。推断：终点放置未成立，也未进入实际携物放置阶段；log step 159所写的后续放置代码不等于已执行该阶段。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 45.0,
+   "log_step": 79,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 44.5–47.0 s，机器人首次合爪后抬升，环柄未离开起端台面；后续视频 182.5–200.0 s 重抓仍未带起工具，视频 261.8 s 工具仍留原处。推断：沿线步骤在获取工具控制的阶段失败且始终未修复，因此断点为金属环走完全线，而非尚未实际开始的终点放置。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 45.0,
+    "log_step": 79,
+    "evidence_zh": "观察：视频 44.5–47.0 s，机器人合爪的位置在木柄上方，抬升时环和柄仍停在台面；视频 183.4、187.3、198.0、200.0 s的后续合爪和抬升也没有带走工具。推断：未建立能携带环柄的有效夹持；后续局部指端触碰不等于建立该夹持，未见先持有工具再掉落的过程。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "观察：log step 109 的机器人服务逐行返回IK无解及关节分支跳变拒绝；log step 153 的agent脚本记录末端下降、合爪并抬升后，视频 182.5–183.4 s环仍在台面。log step 157记录合爪后爪距为0并中止，log step 158增加停留再合爪，仍打印WELD False；视频 198.0–200.0 s再次显示末端抬起而工具不动。推断：抓取动作未形成可用夹持，反复调整接近轨迹和合爪时机仍未解决，归为控制；现有证据不足以唯一细分接触几何、逆解与跟踪因素。"
+   }
+  ],
+  "gap_zh": "没有状态日志，也没有服务原生逐时刻的抓持、通过及终点验收字段；局部指垫接触受遮挡，不能精确核验夹持触点、金属碰线子步或毫米级终态指标。节点未达由工具始终留在可见起端的画面判断，不由零分反推；控制根因保留大类，不能从agent多次变化的诊断中确定唯一子类。评分器未报告DAG所列八项原生指标，均填null。"
+ },
+ "summary_zh": "本次实时尝试两个节点均未完成（终态0/2）：机械臂反复在起端工具附近下探、合爪、抬升，但金属环未被携带沿线移动，视频261.8 s仍留在起端台面。断点为金属环走完全线，现象是未建立有效夹持，日志和画面支持控制失败；末段空手执行沿线动作未修复抓取。无状态日志与原生接触、验收指标，无法进一步核验细小接触或唯一控制子类。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "两节点均未达成、断点为首次抓取失败（视频约45 s）且始终未修复，与帧一致；时钟换算有control_steps核对，无需修改。root cause 取大类控制，证据尚可。",
+    "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": "2026-10-31T00:15:05Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--a3mz6ca/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--a3mz6ca/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "1387a79fe1f90589bdea019dd6898f76863afffb5558cf8ffb587bc4e30aaf39",
+   "video_sha256": "7a419274bd70b322637c456dc8622ab09c4e0fe9634325624880884483d2c685",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--bctuxfd.json

@@ -0,0 +1,272 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--bctuxfd",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 9,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 6246,
+   "live_rows": 6246,
+   "sim_steps": 6246,
+   "physics_steps": 62455,
+   "steps": 1205,
+   "elapsed_s": 6881.3,
+   "replay_wall_s": 415.3,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.33,
+   "peak_processes": 9,
+   "peak_disk_gb": 1.121,
+   "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/roboarc-buzz-wire-3d-10-i01-standard--bctuxfd/replay.mp4",
+  "duration_s": 156.2,
+  "fps": 15.0,
+  "frames": 2343,
+  "sha256": "e10d2e51e28e6317db46c6b836f5afb7af52ecbba3232bebe559623738e1424e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0 s，环柄工具平放在起端支撑台上，机械臂尚未握住工具；完整走线尚未发生。"
+     },
+     {
+      "t": 149,
+      "log_step": 325,
+      "text_zh": "观察：视频 147–154 s，夹爪带着环柄工具离开原先的起端支撑，沿起始竖直线段向上移动；环和木柄随夹爪上升，并未走过中间弯线或到达另一端下降立柱。"
+     },
+     {
+      "t": 156.13,
+      "log_step": 350,
+      "text_zh": "观察：视频 155–156.13 s，工具停在起始竖直段的上端弯折附近，后方中间线段和终端仍未经过。log step 350 的机器人服务返回 EpisodeOver: wire_touched；评分字段 live_matches_replay=1、resumes=0，支持这是该次无重置评分尝试的终止，而非另一次测试。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 147–156.13 s，单臂二指夹爪闭合夹住环柄工具的木柄，工具随夹爪升离起端并沿竖直段上移；此前多次接近、闭爪和抬升未将工具带离原位。"
+     },
+     {
+      "who": "手持工具",
+      "how": "其他",
+      "note_zh": "观察：log step 350 的机器人服务返回 wire_touched，视频 156.13 s 工具位于初始上端弯折附近。候选名称“金属碰触”：所持工具的金属部分碰到裸线；现有推/扫、压/擦均要求这里未证实的受力方式，避开不描述这次已发生的接触；具体是环还是金属杆接触无法分辨。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0 s，工具在起端，终端黑色绝缘下降立柱脚边没有工具。"
+     },
+     {
+      "t": 156.13,
+      "log_step": null,
+      "text_zh": "观察：视频全程粗查及视频 140–156.13 s 密集帧中，工具只从起端沿第一段竖直线升起，末帧仍由夹爪持在高处；机器人没有将其移至终端支撑、放下或撤离。终点放置条件从未成立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 156.1,
+   "log_step": 350,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 147–156.13 s，机器人持环柄工具沿第一段竖直线升至上端弯折附近，未走完全线；log step 350 的机器人服务返回 EpisodeOver: wire_touched，log step 351 终末读数为 control_steps=6246、sim_time_s=156.1375。推断：触线使当前走线步骤不可继续，断点取视频约 156.1 s；6246 个控制步约对应 156.15 s，与视频末端位置及停止相符，未使用 agent 工作时间换算。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 156.1,
+    "log_step": 350,
+    "evidence_zh": "观察：视频 155–156.13 s，机器人仍夹持工具在起始竖直段上端附近；log step 350 的机器人服务明确返回 wire_touched，表示金属工具与裸露金属线发生接触并终止该局，未能继续过弯。该接触由实时单局服务报告，并非仅据 agent 的自述认定。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：log step 325、327、329、335 的控制脚本保持工具朝向、按竖直方向逐段上升；视频 147–156.13 s 也显示工具随夹爪上升至首个上端弯折，随后服务在 log step 350 报告 wire_touched。机器人未能维持工具相对裸线的无接触运动，归于控制；现有证据不足以细分为轨迹、反馈或接触布局问题。log step 335 的上升后末端位姿与 log step 351 终末位姿相同，不能据 agent 在 log step 351 的自述断言倾斜运动已经实际执行。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态或金属接触对日志，普通视频无法分辨具体接触的是环还是金属杆，也不能核验毫米级余隙。log step 351 的输出截断，log step 335 未打印 done/reason，且其末端位姿与终末读数相同；因此不能确定触线发生于末次上升末尾还是后续倾斜请求刚开始，不能细分控制根因。原生终点距离、角度、净距、速度、漂移及静置时长均未在评分字段报告，指标保留 null；这些缺口不影响两个节点均未完成的判断。"
+ },
+ "summary_zh": "机器人反复接近并尝试抓取起端木柄，随后确实夹持工具沿第一段竖直线升至上端弯折附近，两个目标节点均未完成。视频约 156.1 s 对应的终止位置仍在起始线段，实时服务报告 wire_touched，断点为“金属环走完全线”的操作中，现象为碰撞，根因仅能归到控制。缺少逐帧接触对和完整终末状态，无法确认碰触部件或更细的控制失效机制，也不能确认后续倾斜命令实际执行。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两节点均未完成、断点在 loop_traversed、wire_touched 终止与视频末帧（工具仍在起始竖直段上端）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": "2026-10-30T23:47:29Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--bctuxfd/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--bctuxfd/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "42e35713bde09bfb30ab70485c3f04a6a9728a8f521ec00cd0d2b39f75a773d4",
+   "video_sha256": "e10d2e51e28e6317db46c6b836f5afb7af52ecbba3232bebe559623738e1424e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--hxsgxhk.json

@@ -0,0 +1,269 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--hxsgxhk",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 11,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 20532,
+   "live_rows": 20532,
+   "sim_steps": 20532,
+   "physics_steps": 205314,
+   "steps": 681,
+   "elapsed_s": 7105.5,
+   "replay_wall_s": 1564.0,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.305,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.415,
+   "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/roboarc-buzz-wire-3d-10-i01-standard--hxsgxhk/replay.mp4",
+  "duration_s": 513.4,
+  "fps": 15.0,
+  "frames": 7701,
+  "sha256": "9498925305f359b3e9870b755202996988653f69b7792bb281865b0fce2d90e0"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 9,
+      "text_zh": "观察：视频 0.0 s，环柄工具平放在起点铜色竖线脚边，机器人夹爪尚未夹住木柄。"
+     },
+     {
+      "t": 492.9,
+      "log_step": 305,
+      "text_zh": "观察：视频 480.0、490.0、492.9、500.0、508.0 s，机器人二指夹爪夹住木柄，工具离开起点台面，金属环随夹爪沿起点竖线向上移动；环没有走到上方弯折，更没有经过中段弯线和终端黑色立柱。"
+     },
+     {
+      "t": 513.3,
+      "log_step": 323,
+      "text_zh": "观察：视频 512.0、512.5、513.0、513.3 s，机器人仍持环停在起点竖线中上段并小幅转向；log step 322 的实时机器人服务返回 EpisodeOver: wire_touched，log step 323 返回 done=True、reason=wire_touched、control_steps=20532、resumes=0。完整通过未发生。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 490.0–513.3 s，单臂二指夹爪从木柄两侧夹住工具，木柄与金属环随夹爪一起离台上移。"
+     },
+     {
+      "who": "手持工具",
+      "how": "其他",
+      "note_zh": "观察：视频 513.3 s 工具仍由夹爪持有，log step 322–323 的实时服务记录 wire_touched。候选接触方式为“金属碰线”：所持工具的金属部分碰到裸露金属线，现有方式没有对应的碰触值，不能用“避开”替代这次碰撞；此前持环沿起点线段上移。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 513.3,
+      "log_step": 323,
+      "text_zh": "观察：视频 0.0–513.3 s 的全程粗采样及视频 480.0–513.3 s 的末段近查，工具一直位于起点一侧，末帧仍在夹爪中、悬在线段上，终端黑色立柱脚边没有放下的工具。参考成功录像在其视频 239.0 s 可见环到达终端立柱、其视频 344.4 s 可见工具留在终点台面且夹爪打开；本次录像没有出现这些终点位置变化，未开始终点放置。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 513.3,
+   "log_step": 322,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 510.0–513.3 s，机器人夹住工具在起点竖直线段上调整朝向，尚未沿后续弯线通过；log step 322 的转向指令段收到实时服务 EpisodeOver: wire_touched，log step 323 确认终局，后续指令在 log step 324 被拒绝。推断：log step 5 给出控制频率 40 Hz，log step 305 的 control_steps=19717 对应视频 492.9 s，已与该处持环上移后停住的画面核对；终局 control_steps=20532 对应视频 513.3 s，与末段转向吻合，故以视频 513.3 s 为碰线断点锚点。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 513.3,
+    "log_step": 323,
+    "evidence_zh": "观察：视频 512.0–513.3 s，机器人保持夹持并小幅调整工具朝向，金属环仍围着起点裸露竖线；同一实时局的服务在 log step 322–323 报告 wire_touched 并结束。服务的任务文本在 log step 6 将该终止条件定义为金属工具接触裸露金属线；碰线发生在沿线移动这一步，工具未掉落，也未松手。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 513.3,
+    "log_step": 321,
+    "evidence_zh": "推断：log step 321 的控制脚本用固定的工具坐标偏移估计环心，并通过补偿末端平移来执行五段朝向插值，指令执行后才检查关节力矩；log step 322 显示平移对线完成、力矩近零，随后在该朝向调整段触发 wire_touched。结合视频 500.0–510.0 s 的持环上移与视频 512.0–513.3 s 的末端转向，失败指向持物转向时未维持工具与裸线间隙的控制问题；现有证据不足以细分为估计偏移误差、逆解轨迹偏差或反馈控制问题。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或金属部件接触对；录像不能分辨首先碰线的是环还是金属杆，也不能核验毫米级间隙。log step 324、326 中 agent 关于手腕或肘部扫线、逆解分支跳变的解释只是其自述，未被接触记录验证，未据此细分根因。原生终点距离、角度、速度、漂移和确认时长未在评分字段中报告，均记为 null。",
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  }
+ },
+ "summary_zh": "机器人最终夹住木柄，带着环沿起点竖直线段上移，但未走完全线，也未开始终点放置，两个节点均未达成。视频 513.3 s 附近的持物转向触发实时服务 wire_touched 终止，断点为“金属环走完全线”，证据支持转向控制未维持避线间隙。无逐帧状态及接触对，无法确认首先碰线的金属部件或进一步确定控制子类型。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个节点均未达成、断点为 loop_traversed 的判断与视频和日志一致；视频时间由 control_steps/40 Hz 换算并在 492.9 s 与 513.3 s 的画面处核对，根因停在“控制”并写明证据不足以细分，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": "2026-10-30T23:49:07Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--hxsgxhk/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--hxsgxhk/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b18325819e1d2bf9aa619c83ed53ea5f90eaebf418035af0a9d0fcd8b47a2eb0",
+   "video_sha256": "9498925305f359b3e9870b755202996988653f69b7792bb281865b0fce2d90e0",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9.json

@@ -0,0 +1,343 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--j8elhf9",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 10,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 39858,
+   "live_rows": 39858,
+   "sim_steps": 39858,
+   "physics_steps": 398580,
+   "steps": 8967,
+   "elapsed_s": 9211.5,
+   "replay_wall_s": 3196.7,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.511,
+   "peak_processes": 25,
+   "peak_disk_gb": 0.594,
+   "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/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9/replay.mp4",
+  "duration_s": 996.533,
+  "fps": 15.0,
+  "frames": 14948,
+  "sha256": "d66fa4572d332efafd128923478d3e9815418af3d09f84ed2b46ecd5a51abeb3"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [
+     [
+      574.2,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 574.2,
+      "log_step": 204,
+      "text_zh": "观察：已通看全片粗采样；机器人夹持木柄，带环沿起端直立段、上方弯线、U形段及后方立体回弯依次移动，中途把环挂在黑色绝缘套上再换握。视频 573.7–574.2 s 的逐帧腕部画面可见环沿终端黑色立柱下滑，最后套在柱脚，工具落到木台；log step 204 的实时服务仍返回 done=False、reason=None，未因金属碰裸线终止。"
+     },
+     {
+      "t": 574.2,
+      "log_step": 203,
+      "text_zh": "推断：结合全程可见的沿线顺序、末端环仍套着终端立柱及未发生碰线终止，按视觉证据将完整通过记在视频约 574.2 s，此后保留。成功参考在 oracle 视频 230–241 s 也由同一末端下降至柱脚，用于辨认终端位置；本次没有原生 traversal_complete 字段，不能逐物理子步核验唯一交点和连续游标。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "单臂二指夹爪夹住环柄工具的木柄，带环沿线移动；数次松开让环搭挂绝缘套，再夹住垂下的木柄继续。"
+     },
+     {
+      "who": "手持工具",
+      "how": "避开",
+      "note_zh": "所持工具的金属环围绕裸露铜线通过，裸线没有触发终止；环在黑色绝缘套处搭挂并沿绝缘终端滑下。金属工具与裸线的避开由画面及实时服务未触发碰线终止共同推断，非子步接触日志核验。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 574.2,
+      "log_step": 206,
+      "text_zh": "观察：视频 574.2 s 工具已落在终点柱脚的木台上，但柄指向柱的左侧；机器人随后重新抓住柄绕柱转动，再以指尖推柄。log step 206 是 agent 自行从相机估计落台位姿的输出，并非评分器的终点验收字段。"
+     },
+     {
+      "t": 817.5,
+      "log_step": 235,
+      "text_zh": "观察：视频 800 s 夹爪贴近台面推柄；视频 817.5 s 机器人已撤开，工具平躺在柱脚，柄在固定外部视图中指向右上。视频 988 s、995.7 s、996.1 s 及末秒视频 996.466 s 仍保持该朝向。成功参考的 oracle 视频 344.0–344.466 s 中夹爪仍靠近工具两侧，外部视图遮挡柱脚，腕部视图的柄影模糊；没有可靠可比的离手后休止姿态。"
+     },
+     {
+      "t": 995.7,
+      "log_step": 256,
+      "text_zh": "推断：结合终点反复调整后仍未验收，保留工具放到终点未达成的判断，不以“工具在终点台面上、机器人已离开”替代全部验收条件。log step 256 的实时服务仍为 done=False、reason=None，评分字段 live_success=0、replay_success=0、live_matches_replay=1 与未验收相符。工具与成功参考的休止朝向可能不同，但参考末帧受夹爪遮挡，只能粗略判断；缺少原生目标位姿、角度和距离数值，不能确认是否超出 10 度容差，也不能确定唯一未满足条件。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "单臂二指夹爪在终点木台上重新夹住工具木柄，绕终端立柱转动后松开。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "随后夹爪以指尖沿台面拨推木柄来调整朝向；视频 800 s 可见夹爪贴近柄，视频 817.5 s 及末段撤离后的画面中柄仍朝右上；log step 256 的实时服务仍未报告成功。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "wand_on_finish",
+   "t": 817.5,
+   "log_step": 235,
+   "where": "操作中",
+   "evidence_zh": "观察：机器人完成末端下降后，在终点台上重新抓握、绕柱转柄和推柄；视频 800 s 推动后，视频 817.5 s 撤手留下工具平躺在柱脚、柄朝右上，视频 995.7 s 仍保持该朝向。log step 235 的推动及撤离没有取得服务成功，log step 256 仍返回 done=False；断点保留为终点放置，无法仅凭这些画面确定是哪项终态公差未满足。"
+  },
+  "unmet": {
+   "loop_traversed": true,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 817.5,
+    "log_step": 235,
+    "evidence_zh": "推断：视频 817.5 s、995.7 s 工具平躺在终点柱脚，柄在固定外部视图中朝右上，机器人已撤开；结合多次转柄和拨推后仍未验收，将位姿不准保留为可能的现象。与成功参考的休止朝向可能不同，但 oracle 视频 344.0–344.466 s 夹爪仍靠近工具两侧且遮挡柄，参考朝向只能粗略判断，不能把“参考柄朝右下”当作可靠观察。inputs/run.json 的 reward.live_success=0、replay_success=0 只确认未成功，不证明角度超出 10 度或位姿是唯一未满足条件；不补造毫米或角度测量。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "推理/规划",
+    "t": 817.5,
+    "log_step": 257,
+    "evidence_zh": "推断：可能存在将起端朝向沿用为终点目标的规划偏差。log step 49 的 agent 脚本估计起端柄角，log step 252 保存的流程说明写“push the handle to the start yaw”，log step 257 自述将柄推成“inward like the start pose”并以起端角度比较终态；这些记录显示 agent 以起端朝向为目标，但该目标是否正确未由评分字段验证。oracle 视频末段缺少可可靠辨认的离手后休止姿态，因此不再以视频与成功参考朝向不同断言目标选择有误，此根因仅保留为低置信度推断。",
+    "source_zh": "agent 自述 / 日志"
+   }
+  ],
+  "gap_zh": "本次没有状态日志，也没有原生 traversal_complete、逐子步穿环交点、终点目标位姿及 distance_to_finish_m、finish_angle_deg、稳定确认字段。完整通过按可见沿线路径和终端下滑推断，不能排除画面未分辨的短暂交点歧义；oracle 视频 344.0–344.466 s 的夹爪仍靠近工具两侧并遮挡柄，不能可靠提取成功参考的离手后休止朝向。本次柄朝右上可见，但与参考是否不同、是否超出精确公差、位姿是否为唯一未满足条件以及起端朝向是否为错误目标均未验证。agent 在 log step 257 提到“终端可能双交点”和尺寸不符，均未获原生字段证实，不据此另标通过失败、任务设计问题或环境根因。"
+ },
+ "summary_zh": "机器人夹持环柄沿线移动，多次在绝缘套处换握，视频约 574.2 s 环沿终端立柱落到木台，按视觉路径判为完成全线通过���最终进度为 1/2。断在工具放到终点：视频 817.5–995.7 s 工具平躺在柱脚、柄朝右上且机器人已撤开，服务及最终评分仍未验收；位姿不准及沿用起端朝向的规划偏差均保留为推断。成功参考末段夹爪遮挡使休止朝向比较不可靠，且缺少原生通过历史、目标位姿和终点误差字段，故 conf 从 6 降至 5，无法确认具体超差或唯一失败原因。",
+ "conf": 5,
+ "review": {
+  "status": "agreed",
+  "rounds": 3,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "loop_traversed 约 574 s 达成、wand_on_finish 未达成、断点在终点放置，均与画面一致；但位姿不准/推理规划的证据依赖“oracle 344 s 休止姿态柄指向右下”，该帧实际仍是夹爪握着工具、并无休止姿态，该观察不成立。",
+    "n_issues": 2,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "loop_traversed≈574 s 达成、wand_on_finish 未达成、断点在终点放置，与画面和 live_success=0 一致；但“成功参考 344 s 柄指向右下”被写成观察，实际 oracle 视频末帧（344.4 s）夹爪仍跨在工具两侧、视频在此结束，柄朝向被遮挡/只有一个模糊暗影，不能当作可靠的参考休止姿态；位姿不准和推理/规划的视频支撑应降为推断。",
+    "n_issues": 2,
+    "n_major": 0
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 3,
+    "kind": "confirm",
+    "verdict": "agree",
+    "summary_zh": "I1、I2 两项接受均已在 prev/trajectory.json 中完整落实：参考朝向比较降为推断，位姿不准与规划根因标为低置信度推断，conf 降为 5，其余数值未变。",
+    "n_issues": 0,
+    "n_checks": 2
+   }
+  ],
+  "open_zh": [],
+  "confirm": {
+   "verdict": "confirmed",
+   "summary_zh": "I1、I2 两项接受均已在 prev/trajectory.json 中完整落实：参考朝向比较降为推断，位姿不准与规划根因标为低置信度推断，conf 降为 5，其余数值未变。",
+   "checks_ok": "2/2",
+   "n_factual_errors": 0
+  }
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:18:40Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9/04_review_r2"
+   },
+   {
+    "stage": "reconcile",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9/05_reconcile_r2"
+   },
+   {
+    "stage": "confirm",
+    "round": 3,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--j8elhf9/06_confirm_r3"
+   }
+  ],
+  "confirm_prompt": "prompts/06_confirm.md",
+  "confirm_prompt_sha256": "b9f56c095f7ea5e609e58bb21e076f565766c370333d71c7a3512cffb4d78267",
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "cb0b4306109919e18252d1388f522c7db9d6b9f829a2e06709be936181c92a66",
+   "video_sha256": "d66fa4572d332efafd128923478d3e9815418af3d09f84ed2b46ecd5a51abeb3",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--lawubs6.json

@@ -0,0 +1,277 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--lawubs6",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 104,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 41415,
+   "live_rows": 41415,
+   "sim_steps": 41415,
+   "physics_steps": 414150,
+   "steps": 12228,
+   "elapsed_s": 6714.9,
+   "replay_wall_s": 2845.0,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-buzz-wire-3d-10-i01-standard--lawubs6/replay.mp4",
+  "duration_s": 1035.467,
+  "fps": 15.0,
+  "frames": 15532,
+  "sha256": "f95530ceb49cbf895617fe6576924b58a7944df2f7a1406633ea1619780748cf"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，金属环套在外部主视角左前方起端立线的底部，木柄平躺在台面；机器人夹爪尚在工具右上方。"
+     },
+     {
+      "t": 168.0,
+      "log_step": null,
+      "text_zh": "观察：已按约每10秒抽帧检查完整视频，并细查视频 145–194 s。视频 150–170 s，机器人在木柄上方反复张开、闭合并抬爪；视频 168、170、184 s 木柄仍平躺原处、环仍套在起端底部，没有随夹爪升上立线。"
+     },
+     {
+      "t": 1035.4,
+      "log_step": 418,
+      "text_zh": "观察：视频 1010–1032 s 再次下降闭爪，视频 1033 s 机器人空手撤走；视频 1034.4–1035.4 s 逐帧检查中工具仍在起端。log step 418 的实时服务读数也显示末次闭爪间距降为0，log step 422 的episode状态为done=False、reason=None；最终评分live_success=0、replay_success=0且live_matches_replay=1。"
+     },
+     {
+      "t": 1035.4,
+      "log_step": null,
+      "text_zh": "推断：工具始终未被带离起端，未出现沿整根弯线连续移动至黑色绝缘终端的过程，本事件从未成立。参考成功录像的参考视频 20、100、235 s 可见工具随夹爪升起、沿高位线段移动并下降至黑色终端；它们仅用于辨认视觉线索，不作为本次运行证据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "单臂二指夹爪在起端木柄旁反复下降和闭合，指端拨碰木柄，使工具略绕起端立线转动；未观察到两侧夹稳木柄并带起工具。依据视频 150–170 s、1026–1033 s 的近景及初末木柄朝向变化，实际是拨碰而非成功抓握。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，工具位于铜色起端立线底部；黑色绝缘终端立柱脚边没有工具。"
+     },
+     {
+      "t": 1035.4,
+      "log_step": null,
+      "text_zh": "观察：全程抽帧及视频 1034.4–1035.4 s 的末秒帧中，工具仍在起端，终端黑色立柱脚边未出现工具；机器人没有把工具搬运到终点支撑处。"
+     },
+     {
+      "t": 1035.4,
+      "log_step": null,
+      "text_zh": "推断：工具未到终点，终点放置验收从未成立；这一判断来自可见位置，不依赖零评分，也不需要以毫米级静置公差推断。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 150.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 150 s 是起端抓取失败的导航锚点；机器人夹爪下降后重新张开、闭合并抬起，但视频 168–184 s 工具仍躺在起端，未建立能带工具沿线运动的夹持。视频 1010–1035.4 s 的末次重试和撤离后仍是同一结果，第一步始终没有完成；终点放置没有开始。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 168.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 166–170 s，机器人两指合拢后向上抬，木柄却一直平躺在其下方，金属环也仍套在起端底部；视频 184 s 夹爪已抬高而工具留在原处。视频 1026–1033 s 再次闭爪、空手撤离也未带起工具。指端偶有拨碰，但沿线移动所需的有效抓握没有建立。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：机器人在已选定的木柄抓取区域反复逼近和闭爪，却没有实现能夹住木柄的末端位姿与接触。log step 254 的实时服务读数显示逆解返回ik=True，但上移指令后TCP几乎未升高，joint2实际位置0.669 rad、命令0.5261 rad、驱动力矩为−50 N·m；log step 396 更换逆解种子后仍有明显末端和关节跟踪误差。log step 418 末次目标(0.434,−0.472,0.164) m实际为(0.4402,−0.4674,0.1693) m，闭爪间距随后为0；视频 1026–1033 s 对应的实际结果是工具留在台面。证据支持抓取执行控制失败，尚不足以唯一划分为接触设计、逆解或反馈控制子类。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志，视频不能精确核验子步金属接触、辅助抓取状态及毫米级终态速度、漂移公差；评分字段也没有报告DAG所列八项原生指标，故这些值留null。手指局部遮挡木柄接触区，但抬爪前后工具始终在起端，足以判定未完成抓取和沿线通过。关节力矩饱和及跟踪误差的具体物理来源未被独立确定，无法区分碰撞阻挡与当前臂姿、接触方案的影响；agent在log step 423将其解释为硬力矩限制，这不证明所有合法抓法都不可行。"
+ },
+ "summary_zh": "机械臂反复在起端下降、闭爪并测试抬起，但工具始终留在起端，金属环走完全线与工具放到终点均未完成（0/2）。断点为第一步操作中的抓取失败，现象是未建立有效抓握，日志中的末端跟踪误差与力矩饱和支持控制层归因。无状态日志，不能精确核验细粒度接触或唯一确定控制失败的物理来源。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个节点均未达成，视频全程工具停在起端，与标注一致；断点、没建立接触、控制根因（保持在未分子类的 控制）均有视频和服务读数支持，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": "2026-10-30T23:47:23Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--lawubs6/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--lawubs6/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0616b12ad98e6ce2430687878357e88ebeae4a80a9105b25c79c043dd92c53a8",
+   "video_sha256": "f95530ceb49cbf895617fe6576924b58a7944df2f7a1406633ea1619780748cf",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--u5xr6ty.json

@@ -0,0 +1,250 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--u5xr6ty",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 8,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 5388,
+   "live_rows": 5388,
+   "sim_steps": 5388,
+   "physics_steps": 53880,
+   "steps": 1475,
+   "elapsed_s": 10789.4,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.589,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.319,
+   "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/roboarc-buzz-wire-3d-10-i01-standard--u5xr6ty/replay.mp4",
+  "duration_s": 134.8,
+  "fps": 15.0,
+  "frames": 2022,
+  "sha256": "ac6a4ef0987bdee9472b4f9b357cb7a8f5f60e87e5ae229c00025a7b19cbb5b8"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，环柄工具平放在起端立柱脚边；全片粗查及视频 118.0–134.73 s 的密集帧显示，机器人此前反复靠近、闭爪和退回，末段才握住木柄并带环上移。"
+     },
+     {
+      "t": 134.73,
+      "log_step": 1039,
+      "text_zh": "观察：视频 130.0–134.73 s，夹爪带环沿起端竖直段升至上方弯线附近；最后一秒各视角仍显示工具在路径起始区域，未沿后续弯线到终端下降立柱。log step 1039 的最终执行脚本只生成中心线索引 6–96 的路径，log step 1040 启动该脚本；这些索引是脚本路径点，不是视频时钟。完整通过的可见必要条件未满足。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约 127.0–134.73 s，单臂 Robotiq 二指夹爪合拢围住木柄，随后木柄与金属环随夹爪一起离开起端台面并沿线升高；此前还有多次重新接近和闭爪尝试。"
+     },
+     {
+      "who": "手持工具",
+      "how": "避开",
+      "note_zh": "观察：视频 131.0–134.73 s，所持金属环沿起端线段穿行，未见工具撞线或掉落；这里只记录已观看的局部避线运动，普通视频不能核验全部物理子步的微小接触。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 134.73,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–134.73 s 的各视角未出现工具到终端立柱脚边放置、松手撤离的过程；视频 134.73 s，工具仍由夹爪握持在起端上方。作为视觉对照，成功参考视频 325.0 s 与 344.4 s 显示工具已在另一端立柱底部的台面上、夹爪张开；本次始终没有出现该位置关系，故本节点未成立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 134.0,
+   "log_step": 1039,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 133.0–134.73 s，机器人持环仅走到起端上方的部分线段，仍未进入终端下降立柱；log step 1039 的最终执行脚本把路径限定在索引 6–96，而完整中心线终点为索引 499（log step 870）。推断：本次断在完整沿线通过这一步；视频 134.0 s 是末段未完成运动的导航锚点，画面没有显示一个突发碰撞或脱手时刻。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "轨迹或逆解",
+    "source_zh": "日志",
+    "evidence_zh": "观察：agent 的离线求解程序在 log step 491 返回中心线索引 105 处失败；log step 870 的动态规划输出“DP FAILED: No continuous path from start to finish”，只能连接到索引 108；log step 918 的另一轮候选连通检查在索引 355 后无可连接路径。log step 1039 最后改为只生成索引 6–96 的局部路径，log step 1040 执行，视频 130.0–134.73 s 也只显示起端上移与局部弯线运动。推断：agent 未求得并执行贯穿全线的连续关节轨迹，使第一节点停留在局部通过；这些求解失败不证明任务本身不可达。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 没有状态日志，评分器 reward 也未报告 DAG 所列的距离、角度、净距及静止窗口原生指标，因此这些 native_metrics 均为 null，不能核验毫米级公差或逐子步金属接触。视频末段显示的是一段尚未完成的沿线运动，未显示可确定的碰撞、失握、卡阻或瞄准偏差；早先抓取困难后来已有持物上移，故不将其当作未修复的失败现象。现象层留空，根因仅由日志中的连续轨迹求解失败和最终局部执行作有限归因；log step 1048 关于零金属接触等内容是 agent 自述，不能代替服务读数。"
+ },
+ "summary_zh": "机器人最终握住环柄工具，沿起端立柱上移至上方局部弯线，两个节点均未完成，终点放置尚未开始。断点为金属环走完全线；日志多次报告连续轨迹求解失败，最后执行脚本仅覆盖起始部分。无状态日志及原生精细指标，视频也没有显示可明确分类的未修复碰撞或失握，故失败现象层留空。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "视频 127–134.73 s 各视角与标注一致：夹爪握木柄带环沿起端立柱上移，未到终端；两节点均未成立，断点 loop_traversed 合理；log step 1039/1040 的局部路径（索引 6–96）与根因证据一致，现象层留空有 gap_zh 说明。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": "2026-10-30T23:48:58Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--u5xr6ty/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--u5xr6ty/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "62b18056da2394df971022ffa6ba39cae8c0eb4d1a13ede63f0fd43f1fbeeab2",
+   "video_sha256": "ac6a4ef0987bdee9472b4f9b357cb7a8f5f60e87e5ae229c00025a7b19cbb5b8",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/buzz_wire_3d_10/trajectories/roboarc-buzz-wire-3d-10-i01-standard--ysxytkr.json

@@ -0,0 +1,262 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-buzz-wire-3d-10-i01-standard--ysxytkr",
+ "taskKey": "roboarc/buzz_wire_3d_10",
+ "dag_sha256": "983d09ae6790a176236d3b486160ebb8032b6397017c34a91416aee319de567b",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 1,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 361,
+   "live_rows": 361,
+   "sim_steps": 361,
+   "physics_steps": 3610,
+   "steps": 297,
+   "elapsed_s": 8141.4,
+   "replay_wall_s": 34.9,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-buzz-wire-3d-10-i01-standard--ysxytkr/replay.mp4",
+  "duration_s": 9.133,
+  "fps": 15.0,
+  "frames": 137,
+  "sha256": "a523640cd306e2fb780c1bb1c75d0b23a3ec0e0569b71bc8d6b265cd9b336706"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "loop_traversed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0 s工具的木柄平放在前侧起端，金属环套在起端线柱底部；视频0.0–5.5 s机械臂先抬起并改变腕部观察视角，工具未离开原位。"
+     },
+     {
+      "t": 6.4,
+      "log_step": 219,
+      "text_zh": "观察：视频5.8–6.8 s机械臂向起端工具接近，随后停在工具上方；视频7.0–9.07 s夹爪在空处闭合，前侧起端的木柄及金属环仍在原位，未沿立体弯线移动。log step 219实际提交的是接近并闭合夹爪的动作段，尚未提交沿线运动。"
+     },
+     {
+      "t": 6.9,
+      "log_step": 232,
+      "text_zh": "观察：log step 232读取本次实时动作段的服务安全监测，记录link5压在线上，最大压载225.5 N、最长强推2.67 s。推断：log step 211所示控制频率40 Hz，监测事件control step 275对应约视频275/40=6.88 s；已与视频6.8–7.0 s臂段停滞画面核对，事件时刻不等于首次接触时刻。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频6.4–9.07 s接近工具的单臂停滞，夹爪未夹住仍在起端的木柄；log step 232服务监测确认实际承压接触为机械臂link5与弯线。此处按单臂归入手指/夹爪，作用端实际是link5臂段，非夹爪抓握；未发生手持金属环沿线避让。"
+     }
+    ]
+   },
+   "wand_on_finish": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 9.07,
+      "log_step": null,
+      "text_zh": "观察：视频0.0–9.07 s工具始终留在前侧起端，结尾木柄仍横放在起端线柱脚边；没有到另一端下降立柱脚边的放置、松手和撤离。成功参考视频约240 s可见工具随夹爪到终端，参考视频344.4 s环位于另一端柱脚；本次所有视角均未出现该转移。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "distance_to_finish_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "finish_angle_deg",
+   "unit": "deg",
+   "threshold": 10,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "clearance_lower_bound_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "maximum_speed_m_s",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_drift_m",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_angular_speed_rad_s",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "maximum_rotation_drift_rad",
+   "unit": "rad",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "stable_time_s",
+   "unit": "s",
+   "threshold": 1,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "loop_traversed",
+   "t": 6.4,
+   "log_step": 232,
+   "where": "操作中",
+   "evidence_zh": "观察：视频6.2–6.8 s机器人为获取起端工具而下降，随后在工具上方停滞；视频8.2–9.07 s夹爪闭合但工具仍在起端。log step 232的本次实时服务监测明确记录link5压在线上，log step 234显示实际TCP高度约0.370 m而接近目标在log step 236降至约0.159 m。推断：沿线步骤在接近抓取阶段被臂段碰线阻断，之后未修复。"
+  },
+  "unmet": {
+   "loop_traversed": false,
+   "wand_on_finish": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 6.4,
+    "log_step": 232,
+    "evidence_zh": "观察：视频6.4–9.07 s机械臂停在线的上部，夹爪始终悬在起端工具上方，工具没有被带动；log step 232服务监测记录pressing=\"the robot's link5\"、on=\"the wire\"，最大压载225.5 N。推断：接近路径被link5与线的接触阻断，未能到达木柄抓取位姿。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：接近动作的几何避碰检查不足，使臂段碰线后关节目标仍继续下发。log step 216构造并保存接近轨迹，其避碰检查仅用arm_tcp、hand、link7、link6的点及半径；log step 219整段提交接近、保持和闭爪动作，未依据接触反馈修改；log step 232服务确认link5压线，log step 234实际关节与指令明显偏离且TCP停在约0.370 m，未到log step 236所列约0.159 m的抓取目标。上述证据支持控制问题，但不足以进一步区分碰撞形状建模与在线反馈控制各自的贡献。"
+   }
+  ],
+  "gap_zh": "无逐帧状态日志；link5的具体碰线位置部分被机器人遮挡，接触对象与压力采用本次单一实时episode的服务监测记录，首次接触时间仅作视频约6.4 s的粗定位。毫米级通过和终点静置指标未在评分器最终字段中报告。log step 254虽出现模型过滤中断，但此前本次评分轨迹已经发生且始终未修复的物理阻塞，按此前物理失败标注；不把过滤中断当作机器人现象或根因。"
+ },
+ "summary_zh": "本次完成0/2节点：机器人先移动腕部相机观察，再接近起端工具，但link5压在线上而停滞，夹爪未抓起木柄，金属环未沿线移动，也未尝试终点放置。断点为“金属环走完全线”的操作中，服务接触监测与视频停滞相互印证，日志支持接近轨迹避碰及接触反馈方面的控制问题；无逐帧状态日志，首触时间及更细控制归因仍不确定。末尾模型过滤发生在上述未修复物理失败之后。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两个节点均未达成、断点为loop_traversed（操作中）、被挡/碰撞（link5压线，服务监测读数）及控制根因均有视频与log支持；未发现改变结论的错误。",
+    "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": "2026-10-30T23:54:40Z",
+  "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/roboarc-buzz-wire-3d-10-i01-standard--ysxytkr/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-buzz-wire-3d-10-i01-standard--ysxytkr/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "83b32e4ffc529720cbd879db0c84bf6367a0118da3c1996e1e3c610b6a72d2a7",
+   "video_sha256": "a523640cd306e2fb780c1bb1c75d0b23a3ec0e0569b71bc8d6b265cd9b336706",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-privileged--g4mc3kt.json

@@ -0,0 +1,288 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-privileged--g4mc3kt",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 1130,
+   "submitted_actions": 1130,
+   "sim_steps": 1130,
+   "physics_steps": 11300,
+   "sim_time_s": 28.249999999999464,
+   "final_reward": 0.0,
+   "replay_wall_s": 98.4,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 1,
+   "agent_resets": 2,
+   "agent_sim_steps": 0,
+   "agent_records": 2,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.825,
+   "peak_memory_gb": 1.463,
+   "peak_processes": 14,
+   "peak_disk_gb": 1.006,
+   "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/roboarc-cast-keyhole-03-i01-privileged--g4mc3kt/replay.mp4",
+  "duration_s": 28.333,
+  "fps": 15.0,
+  "frames": 425,
+  "sha256": "9921ad3e68aaaf89d84c3cc733e0efa7270427918bf787aa8eb5c44ef7843d51"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 0,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      0.1
+     ]
+    ],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 29,
+      "text_zh": "推断：视频 0.0 s 金色件斜立，两夹爪均未接触拼图；log step 29 的冻结初态金色件 bounds 的 y 上界为 0.08076 m。按任务给出的初态余留组件最靠目标点再外移 5 mm 的固定分隔面定义，开局该关系成立。"
+     },
+     {
+      "t": 0.1,
+      "log_step": 246,
+      "text_zh": "推断：视频 0.000、0.067、0.133、0.200 s 显示金色件在两夹爪尚远离时自行倾转沉降，随后粗查视频全程及视频 27.3–28.27 s 末段均未见金色件回移。log step 246 的最后录制输出金色件末态 y 上界 0.0994 m，动作数 1130 与 log step 248 的提交文件及评分器 n_actions=1130 一致，物件平躺姿态也与视频相符；按 log step 29 初态和任务定义，固定分隔面 y≈0.08576 m，末态超出约 13.6 mm。bounds 随倾转变化正是整件顶点判据需要计入的变化，不能仅以主视图上沿几乎不变排除越面；故保留 final=false，视频约 0.1 s 只是联合证据推断的丢失锚点，并非直接量测。成功参考视频 0.0–1.0 s 也有类似无接触沉降，仅用于识别该开局现象，不证明本次节点值。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.6,
+      "log_step": 246,
+      "text_zh": "观察：视频 2.5–4.3 s 的密集帧显示右夹爪靠近银色件右端并闭合，银色件仍留在桌面原拼图处；视频 10–24 s 右夹爪转动、抬起时银色件没有随行。log step 246 的交付录制输出 grasp 的 in_hand、got 均为 None，路径结束仍为 None，与视频未抓起的结果一致。"
+     },
+     {
+      "t": 28.27,
+      "log_step": 248,
+      "text_zh": "观察：视频 25–28.27 s 右夹爪空手移到绿框上方，银色件整个长框仍在绿框近侧桌面，与金色件轮廓重叠，绿色目标区内没有银色件。log step 248 核对的提交文件为 1130×14 动作，与 log step 246 的最后录制一致；评分器 strict_goal_reached=0。成功参考视频 675.0 s 的银色件已与金色件分开并由右夹爪持于绿框上方，仅用于辨识应有的视觉形态。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频 3.3–4.4 s 机器人右夹爪在银色件右端闭合，局部被夹爪壳体遮挡，俯视、侧视和腕视均不能确认夹指是否实际触及或仅从旁闭合；视频 20–28.27 s 银色件仍留原处，不能记为抓握。左夹爪未参与操作。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 3.6,
+   "log_step": 246,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 2.5–4.4 s 机器人右夹爪靠近银色件右端并闭合；视频 10–24 s 夹爪运动而银色件留在原处，视频 27.3–28.27 s 夹爪已空手离开，银色件仍在绿色目标区外。推断：机器人实际着手且未修复的是银色件入区步骤的抓取环节，以视频约 3.6 s 为失败锚点（docs/03_trajectory.md Step 3.1；DECISIONS.md S14）。开局金色件无接触沉降不是该机器人操作失败。"
+  },
+  "unmet": {
+   "gold_on_system_side": false,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 3.6,
+    "log_step": 246,
+    "evidence_zh": "观察：视频 3.3–4.4 s 机器人右夹爪在银色件右端闭合，视频 10–24 s 夹爪转动、抬起时银色件没有随行，视频 28.27 s 仍在原拼图处；log step 246 中脚本调用 object_in_hand 的输出 in_hand、got 均为 None，路径结束仍为 None，仅作辅助。推断：所需的持物接触从未建立，标接触·没建立接触；不表示已排除短暂单指触碰，也没有先持住再丢失的可见证据。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 246,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 244 的脚本及 log step 246 的重录修改显示机器人按抓银色件、解脱、入区的目标顺序执行；没有证据证明该目标或顺序选错。视频 3.3–4.4 s 闭爪未形成持物、随后空手执行，与 log step 246 的 in_hand=None 一致；同一步工具输出还记录右臂 right/forearm_roll 的 IK 解距当前关节 2.42 rad、将摆臂的警告。因此归入抓取执行的控制问题。局部接触被遮挡，尚不能证实具体接触几何不可行，也不能证明接触方案可行而仅逆解失败；依 docs/03_trajectory.md Step 3.3 判别次序停在控制，不强选抓取和接触设计或轨迹或逆解。agent 在 log step 249 将开局倾倒归咎于 IK 的自述与视频不符，未用作因果证据。"
+   }
+  ],
+  "gap_zh": "没有状态日志，固定分隔面不可见；金色件 final=false 来自冻结初态、任务分隔面定义和与交付一致的末态 bounds，视频约 0.1 s 的关系丢失时间仍是推断，主视图上沿几乎不变不能单独量测越面。任务设计问题：金色件在夹爪远离时即从斜立初态自行沉降，按上述边界推断会失去开局保持关系，不能把此沉降标成机器人位姿误差；成功参考视频 0.0–1.0 s 也见类似沉降，说明不能据此宣称任务无合法解。银色件闭爪处遮挡，无法确认短暂单侧指接触，contact_actual 保持未知；所需持物接触未建立和未入区可见。抓取几何不适配与 IK 跳支对抓取失败的具体贡献无法区分，可行替代抓法也未经本次证据验证，根因停在控制并将 conf 降至 4。七项原生数值指标均未由评分器单独报告，保留 null。"
+ },
+ "summary_zh": "本次右夹爪尝试抓取银色件却未形成持物，随后空手执行路径，视频 28.27 s 银色件仍在目标区外，断点为银色件入区的抓取环节，现象为没建立接触、根因归控制。任务设计问题：金色件开局在两夹爪远离时自行沉降，结合初态与交付末态 bounds 推断其越过固定分隔面，保留最终两节点均不成立，但不把自然沉降归为机器人位姿误差或规划失败。无状态日志、分隔面不可见，金色件丢失时间及抓取控制子因未能直接判定，conf 降至 4。",
+ "conf": 4,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点值与时间��体可接受，但断点与根因选错：金色件在视频约0.1 s 的倾转是开局物理沉降（成功的 oracle 视频 0.0–1.0 s 同样出现），不是机器人行为；机器人实际着手并失败的是银色件抓取（视频 2.5–4.4 s 右夹爪闭合未抓起），应作断点，并补 没建立接触 现象。把‘位姿不准’‘推理/规划’归于一个无机器人接触的沉降事件缺乏证据。",
+    "n_issues": 4,
+    "n_major": 3
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "断点（银色件入区，视频约 3.6 s）、没建立接触现象、控制根因及两节点 final=false 与帧和 log 相符；金色件越面时间为已声明的推断，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": "2026-10-31T00:15:51Z",
+  "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/roboarc-cast-keyhole-03-i01-privileged--g4mc3kt/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-privileged--g4mc3kt/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-privileged--g4mc3kt/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-privileged--g4mc3kt/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "54904a674b8357746fa70d2648252474b867f52d57e71de05f22877d7fc15396",
+   "video_sha256": "9921ad3e68aaaf89d84c3cc733e0efa7270427918bf787aa8eb5c44ef7843d51",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-privileged--i4yrn3k.json

@@ -0,0 +1,269 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-privileged--i4yrn3k",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 620,
+   "submitted_actions": 620,
+   "sim_steps": 620,
+   "physics_steps": 6200,
+   "sim_time_s": 15.500000000000153,
+   "final_reward": 0.0,
+   "replay_wall_s": 61.7,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 2,
+   "agent_resets": 10,
+   "agent_sim_steps": 1890,
+   "agent_records": 2,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-cast-keyhole-03-i01-privileged--i4yrn3k/replay.mp4",
+  "duration_s": 15.6,
+  "fps": 15.0,
+  "frames": 234,
+  "sha256": "e7f3682ffef8f813820437ac65d0ef6676e5ace7ec0018c4f9c29664c814ed03"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色件位于绿框近侧，与银色件套接；视频 0.0–1.0 s 有初始落稳，随后直至视频 15.53 s，金色件一直留在同一桌面位置，两臂均未触及它。"
+     },
+     {
+      "t": 15.53,
+      "log_step": null,
+      "text_zh": "推断：DAG确认初态已在系统侧，视频 0.0–15.53 s 未见金色件向目标方向移送，故将该初始关系记为全程维持；不是仅凭金色件未进入绿框判定。隐形分隔面的毫米级余量无状态数据复核。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.4,
+      "log_step": null,
+      "text_zh": "观察：视频 1.4–3.4 s 右臂向银色件方向转动、降低，右夹爪在银色件右侧仍有明显间隔处合拢；银色件继续留在原处，未随夹爪运动。视频 3.4–14.6 s 可见夹爪重复开合，但未将银色件夹在两指之间。"
+     },
+     {
+      "t": 15.53,
+      "log_step": null,
+      "text_zh": "观察：视频 14.6–15.53 s 各视图仍显示银色件与金色件相互套接、留在绿框近侧，整个银色轮廓未进入绿框，目标区为空；全程未见分离、提起或搬入目标区。"
+     },
+     {
+      "t": null,
+      "log_step": 134,
+      "text_zh": "观察：log step 132 的脚本输出为抓取失败并保存 partial (620, 14)；log step 134 读回交付文件，动作数为620且元数据注明 grasp failed，与评分器 n_actions=620 及视频中空夹后未搬动物体一致。此前 log step 127 的1202行动作属于另一会话尝试，不计入本次进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "其他",
+      "note_zh": "观察：视频 3.4–14.6 s 右臂二指夹爪反复开合，目标为银色钥匙孔件，但手指一直在物体右侧的空处，未夹住或推动银色件；左臂未参与。候选名称：空夹（未接触的抓取尝试）；抓握要求实际夹住，推/扫、压/擦等也没有发生，故用其他记录这次实际尝试。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 3.4,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 3.4 s，右臂在银色件右侧闭合夹爪时，银色件没有位于两指之间，仍与金色件套接在原处；之后到视频 15.53 s 末帧都未抓取、解脱或入区。此处为银色件入区步骤的抓取阶段失败，后续开合未修复。"
+  },
+  "unmet": {
+   "gold_on_system_side": true,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 3.4,
+    "log_step": null,
+    "evidence_zh": "观察：视频 3.2–3.8 s 右夹爪在银色件右侧开合，指尖与银色件之间留有可见间隔，物体没有进入两指间；视频 5.6–15.53 s 的后续开合也没有使银色件离开原位或随手移动，所需抓握从未建立。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "轨迹或逆解",
+    "log_step": 132,
+    "evidence_zh": "推断：log step 132 对最终保存620行的尝试输出 right arm IK failed，位置残差33.9毫米及48.6毫米，随后报告 in_hand: None、no_bilateral_candidate 并保存 grasp failed 轨迹；log step 134 确认提交的正是620行失败文件。结合视频 3.4–15.53 s 夹爪在银色件旁空合，说明本次抓取运动未求解并执行到需要的末端位姿，使夹指未到达物体；这不证明整项任务或所有抓姿不可达。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "无状态日志或接触对记录，不能数值复核隐形分隔面、网格边距和速度；金色件维持系统侧由DAG给定初态及视频中未被操作、未明显向目标移动推断。评分器未报告DAG列出的七项原生指标，均记为null。日志的会话调试不逐条映射视频时间；解脱路径未在该交付轨迹中执行，不能据会话中的搜索结果证明其物理可行性。"
+ },
+ "summary_zh": "评分轨迹维持了初始的金色件留系统侧关系，最终为1/2节点；银色件始终与金色件套接在绿框外。右臂在视频约 3.4 s 开始空夹，未建立抓握，日志中最终620行尝试的IK残差与此一致，银色件解脱和入区均未发生。无状态日志，金色件相对隐形分隔面的精确余量及七项原生指标无法数值复核。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点值、断点、现象与根因均与视频和日志一致：金色件初态在系统侧、银色件始终未脱离，右夹爪在其右侧空合；log step 132/134 的IK残差和620行失败文件与评分器n_actions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": "2026-10-30T23:49:45Z",
+  "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/roboarc-cast-keyhole-03-i01-privileged--i4yrn3k/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-privileged--i4yrn3k/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b48aedc467fcb8cac535b3ad869ceb0472f5632bfc65a97cb852f1aa7fa16b3b",
+   "video_sha256": "e7f3682ffef8f813820437ac65d0ef6676e5ace7ec0018c4f9c29664c814ed03",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-privileged--mwr6abo.json

@@ -0,0 +1,264 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-privileged--mwr6abo",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 1,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 27172,
+   "submitted_actions": 27172,
+   "sim_steps": 27172,
+   "physics_steps": 271711,
+   "sim_time_s": 679.2774999996561,
+   "final_reward": 1.0,
+   "replay_wall_s": 1879.8,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 1,
+   "strict_goal_row": 27172,
+   "strict_replay_required": 0,
+   "agent_sessions": 3,
+   "agent_resets": 8,
+   "agent_sim_steps": 119144,
+   "agent_records": 53,
+   "agent_replays": 5,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-cast-keyhole-03-i01-privileged--mwr6abo/replay.mp4",
+  "duration_s": 679.4,
+  "fps": 15.0,
+  "frames": 10191,
+  "sha256": "f78d3e8890079696b2f33799b530736e7c9c8c779bda12835c135c3bd1c17571"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 2,
+  "final": 2,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，两件在绿框近侧相互套接，金色件位于原拼图一侧，符合本节点的初态系统侧关系。"
+     },
+     {
+      "t": 15,
+      "log_step": null,
+      "text_zh": "观察：视频 13.0–15.0 s，左夹爪夹住金色件并抬起，银色件随互锁结构离开桌面；视频 50.0–55.0 s，右夹爪接住银色件、左夹爪松开金色件，视频 80.0–100.0 s 左夹爪重新夹住金色件的长端。"
+     },
+     {
+      "t": 317.5,
+      "log_step": 182,
+      "text_zh": "推断：视频全程每隔约 10 s 的三视图及视频 35.0–105.0 s 的换抓帧中，金色件虽有抬高、转动和平移，未见向目标侧越过分隔面的明确迹象；log step 182 的已录制检查点 12706 显示金色件水平 y 上界为 -0.02719 m，仍在目标近侧。分隔面本身不显示，连续保持按这些观察推断，未把抬高导致的画面投影上移当作越界。"
+     },
+     {
+      "t": 679.33,
+      "log_step": 274,
+      "text_zh": "观察：视频 629.0–679.33 s，左夹爪持金色件留在画面左下方，右夹爪仅将已分开的银色件送向绿框；视频 678.267–679.333 s 的逐帧检查中金色件保持该处。log step 274 的交付生成器重放结果为 success=True、episode_end=success；评分器 strict_goal_reached=1、success=1，支持金色件末态满足系统侧判据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 13.0–15.0 s 左臂二指夹爪夹持并抬起金色件；视频 50.0–55.0 s 松开换抓，视频 80.0–100.0 s 重新夹住金色件长端，后续持住并调整其位姿，末态仍夹持。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [
+     [
+      679.3,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 银色件在绿框外并与金色件套接；视频 13.0–15.0 s 随左爪持起的金色件升离原桌面位置。"
+     },
+     {
+      "t": 100,
+      "log_step": null,
+      "text_zh": "观察：视频 50.0–55.0 s 右夹爪从新方向夹住银色件、抬持互锁组件；视频 80.0–100.0 s 左夹爪重抓金色件后，两臂各持一件。视频 140.0–600.0 s 两件经历缓慢的相对平移、翻转和过槽，尚未入目标区。"
+     },
+     {
+      "t": 620,
+      "log_step": 212,
+      "text_zh": "观察：视频 606.0–620.0 s，左爪的金色件向左下方退开，右爪的银色长框与之拉开明显间隙；视频 621.0–632.0 s 两件保持分开，银色件随后开始向绿框移动。log step 212 的录制检查点 25170 也记录两夹爪分别持有金色件和银色件。"
+     },
+     {
+      "t": 679.3,
+      "log_step": 274,
+      "text_zh": "推断：视频 650.0–679.33 s 右爪把银色件悬持于绿框上方，末秒逐帧显示其位姿趋于稳定，金色件仍远离绿框。悬空件在斜视图中的轮廓不能直接当作桌面垂直投影，严格边距、高度、速度与共同确认采用本次交付重放的评分结果：success=1、strict_goal_reached=1、strict_goal_row=27172；log step 274 的生成器重放也在 sim_steps=27172 成功结束。按该交付的 40 Hz 控制行计数，27172/40=679.3 s（与视频末帧相合），并以 log step 212 的检查点 25170/40=629.25 s 对应视频 629.0–632.0 s 分离后开始入区运动作可见事件核对，故完整验收时间记为约视频 679.3 s；不是将 agent 工作时间换算成视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 13.0–15.0 s 左爪先持金色件带起互锁的银色件；视频 50.0–55.0 s 右爪换抓银色件并支承组件、左爪暂时松开；视频 80.0–100.0 s 左爪重抓金色件。随后左爪夹金色件、右爪夹银色件，协同翻转解脱；视频 629.0–679.33 s 左爪留持金色件、右爪将银色件移入目标体积，末态两爪均未释放。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "双臂通过换抓及相对平移、翻转将两件解脱，左爪将金色件留在系统侧，右爪悬持银色件送入目标体积，约视频 679.3 s 通过共同验收，最终完成 2/2 个节点，无失败断点。金色件从视频 0.0 s 起计为成立，途中未见明确失去系统侧关系的证据。无状态日志，隐形分隔面的逐点边界和银色件的毫米边距、精确速度无法从斜视录像独立测量，连续区间含视觉推断、严格末态依评分器成功字段确认；评分器未逐项报告 DAG 原生指标，七项数值均留空。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "两节点终局成立、无断点，与评分器 success=1/strict_goal_reached=1 及视频末帧（金色件在左下、银色件悬于绿框上方）一致；时钟换算有事件核对，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": "2026-10-30T23:48:37Z",
+  "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/roboarc-cast-keyhole-03-i01-privileged--mwr6abo/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-privileged--mwr6abo/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c8d410d3297d7fe5ba4974ef9e7a560995efe23278f677e34411b7d4b2f1d441",
+   "video_sha256": "f78d3e8890079696b2f33799b530736e7c9c8c779bda12835c135c3bd1c17571",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-privileged--wcrfat2.json

@@ -0,0 +1,297 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-privileged--wcrfat2",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 33,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 13336,
+   "submitted_actions": 13336,
+   "sim_steps": 13336,
+   "physics_steps": 133360,
+   "sim_time_s": 333.39999999997065,
+   "final_reward": 0.0,
+   "replay_wall_s": 1325.3,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 72,
+   "agent_resets": 105,
+   "agent_sim_steps": 153329,
+   "agent_records": 19,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-cast-keyhole-03-i01-privileged--wcrfat2/replay.mp4",
+  "duration_s": 333.467,
+  "fps": 15.0,
+  "frames": 5002,
+  "sha256": "7cf95589f76c1e2745f7d04d96fe449be1f024865cb19e67a8ee63d56ecc9ce9"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 0,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      178.0
+     ]
+    ],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 4,
+      "text_zh": "观察：视频 0.0 s，金色长框在绿色目标框近侧，与银色件套接；log step 4 的初态金色件 bounds 给出最高 y=0.08076 m，按任务固定分隔面向目标外移 5 mm 的定义，初态满足系统侧条件。"
+     },
+     {
+      "t": 174.0,
+      "log_step": null,
+      "text_zh": "观察：视频 10–174 s，金色件先随右臂操作转动、向画面近侧移动，随后右夹爪直接推拨金色件；视频 174 s 金色件仍横放在目标框近侧，明显远于初态的目标近边。"
+     },
+     {
+      "t": 178.0,
+      "log_step": 210,
+      "text_zh": "推断：视频 176–180 s，双臂接触推压使两件大幅转向，金色件圆头转到朝向目标的一端；结合后续持续向目标侧移动，估计约视频 178 s 失去系统侧关系，边界时刻不确定约数秒。log step 210 对最终保存的 run 12 输出金色件 ymax=0.114–0.158 m，已超过该脚本使用的固定分隔面 y=0.08576 m；这些迭代没有视频控制步计数，未把其日志时钟换算成视频时钟。"
+     },
+     {
+      "t": 333.4,
+      "log_step": 211,
+      "text_zh": "推断：视频 308、326、332.5、332.8、333.4 s，机器人末段推拨后两件仍靠在目标近侧偏左，金色件圆头没有明确退回初态系统侧范围；log step 211 的 run 12 结束输出 GOLD OK: False，log step 212 确认保存 13336 行，log step 215 明确提交 run 12，与评分器 n_actions=13336、deterministic=1 相符，支持终态为假。末帧本身不能量出隐形分隔面的毫米距离。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "推/扫",
+      "note_zh": "观察：视频约 96–138 s 右夹爪指端直接推拨金色件；视频约 140–326 s 的双臂协同推压中，一侧指端推拨两件接合处，另一侧接触银色件，金色件随之移动、转向；末段再次尝试把金色件拨回系统侧。未见左右夹爪各自抓稳一件的双臂抓握。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 4.0,
+      "log_step": null,
+      "text_zh": "观察：视频约 4–8 s 右二指夹爪夹住银色圆头；视频约 10 s 两件转向后落回桌上，右臂随后重新靠近并夹持银色件。视频 16–60 s 可见右臂反复提拉、转腕，金色件在银色件另一端随动；这段两件一直在绿框近侧。"
+     },
+     {
+      "t": 62.25,
+      "log_step": null,
+      "text_zh": "观察：视频 61.25–62.25 s，右臂持银色件转腕时金色件一起摆动；视频 62.5 s 两件落回桌面、银色长框斜���金色件，视频 66–90 s 右夹爪离开并向目标上方空手移动，银色件仍留在原处，没有随夹爪进入目标区。"
+     },
+     {
+      "t": 180.0,
+      "log_step": 207,
+      "text_zh": "观察：视频约 140–326 s，左右夹爪先后及同时从近侧推压，银色件翻转、向左偏移并带动金色件；银色件的整个轮廓从未落在绿框以内。log step 207 将最终运行的控制脚本改为左臂保持推压、右臂加入推压，符合这段实际双臂动作。"
+     },
+     {
+      "t": 333.4,
+      "log_step": 211,
+      "text_zh": "观察：视频 332.5、332.8、333.4 s，两臂撤开后银色件仍斜靠金色件，完整留在绿框近侧偏左，绿框内没有银色长框；最终保存的 run 12 在 log step 211 输出 SILVER IN SQUARE: False，评分器 strict_goal_reached=0。对照成功示例 oracle 视频 675、679.3 s，银色件由右臂单独持于目标上方、金色件由左臂留在近侧，形态与本次末段明显不同；示例只用于辨认形态。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "推/扫",
+      "note_zh": "观察：视频约 140–326 s 左右夹爪指端协同直接推压银色件及两件接合处，造成滑移与转动，未形成双臂各持一件的抓握。此前视频约 4–62 s 为右侧单只二指夹爪抓握银色圆头并提拉、转腕，金色件借桌面或与银色件的接触承托；视频约 62 s 后该抓握丢失。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 62.25,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 61.25–62.25 s 右夹爪转腕后两件落回桌上，视频 62.5–90 s 银色件留在桌面而右臂空手向目标移动，所需的持件输送没有发生；视频约 140–326 s 的推移重试也未把银色件完整送入绿框，视频 333.4 s 仍在框外。金色件系统侧关系的后续丢失未修复，但银色件入区操作更早已失败且从未完成。"
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 62.25,
+    "log_step": null,
+    "evidence_zh": "观察：视频 60.75–61.75 s 右夹爪夹持银色圆头，转腕时银色件和另一端的金色件一起摆动；视频 62.25–62.5 s 两件回落并斜靠在桌上，视频 66–90 s 夹爪已空而银色件留在原处。不是在目标内按计划放置后失稳，持件输送在框外就已丢失。"
+   },
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 176.5,
+    "log_step": 207,
+    "evidence_zh": "观察：视频 174–180 s 左右夹爪在两件相接处推压，银色件翻转、金色件被一并转动和带向目标侧；视频 184–198 s 及 220–326 s 两件持续贴靠，推压反复牵动金色件而未使银色件单独进入绿框。log step 207 的脚本明确让左臂保持推压、右臂接着推并周期性前冲，视频中可见的碰动还破坏了金色件原有的位置关系。"
+   }
+  ],
+  "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：视频约 60.75–62.5 s，右夹爪只夹银色圆头，转腕时另一端金色件的随动载荷使两件摆动，银色件最终未留在夹爪内；后续视频约 140–326 s 又用同侧指端推压相互贴靠的两件，没有把金色件与银色件的运动隔开。log step 207 可读到最终脚本只按银色件 bounds 选两个近侧推点，左臂保持推压后右臂继续前推，并加入前冲；log step 210 的同一 run 12 输出银色件 x 下界约 -0.400 m、金色件 ymax 超过分隔面，log step 211 同时报告两项目标为假。这支持接触位置及承载布局不适配当前两件状态，不能产生所需的独立入区运动；没有足够证据确定具体摩擦、夹持力或辅助抓���断开阈值。成功示例使用双臂各持一件，但其镜像版本不同，未据此确认当前版本的替代接触方案。"
+   }
+  ],
+  "gap_zh": "无状态日志，固定分隔面不可见，金色件越界约视频 178 s 为结合转向前后画面与最终保存 run 12 的对象边界输出作出的估计，误差约数秒，不能重建毫米级临界抖动；其终态为假有同一交付运行的 GOLD OK: False 支持。银色件始终明显在框外，无需猜测速度或 0.5 s 保持是否达标；七项原生连续指标均未由评分器报告，故 value 为 null。日志中“完全分离”的叙述涉及多次会话测试，未替代本次视频中的持件、移动证据；部分生成脚本输出被截断，没有逐帧接触力或辅助抓取状态，根因只推断到接触设计。",
+  "unmet": {
+   "gold_on_system_side": false,
+   "silver_in_goal": false
+  }
+ },
+ "summary_zh": "金色件留系统侧在开局成立，约视频 178 s 的双臂推压转向后丢失；银色件全程未完整入区，终态为 0/2。断点为银色件入目标区操作：约视频 62 s 持件接触丢失，后续同侧推压又使两件共同转动、偏移，接触布局未能实现独立输送。无状态日志，金色件越界时刻只能粗估，接触力及辅助抓取失效机制无法进一步核实。",
+ "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（银色件入区，约视频62 s持件丢失）、现象与根因与可见画面和最终run 12输出一致；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": "2026-10-30T23:55:24Z",
+  "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/roboarc-cast-keyhole-03-i01-privileged--wcrfat2/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-privileged--wcrfat2/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "bce0f7c0d7d3811b9d651843491b06e52927c57ea6c8624cc355025ea3565d2d",
+   "video_sha256": "7cf95589f76c1e2745f7d04d96fe449be1f024865cb19e67a8ee63d56ecc9ce9",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-standard--7swrrxf.json

@@ -0,0 +1,281 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-standard--7swrrxf",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 19158,
+   "live_rows": 19158,
+   "sim_steps": 19158,
+   "physics_steps": 191580,
+   "steps": 2481,
+   "elapsed_s": 10818.5,
+   "replay_wall_s": 998.3,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-cast-keyhole-03-i01-standard--7swrrxf/replay.mp4",
+  "duration_s": 479.0,
+  "fps": 15.0,
+  "frames": 7185,
+  "sha256": "8b212c1e95e0515069b83e053df9ac570210dc4168f4de23e129ab87524ac89f"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      62.5
+     ],
+     [
+      128.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色件与银色件套接于绿框近侧；金色件处于任务规定的初始系统侧位置。"
+     },
+     {
+      "t": 62.5,
+      "log_step": null,
+      "text_zh": "推断：视频 62.5–63.0 s 左夹爪失去金色件，两件落回桌面并转斜；结合视频 64.0 s、80.0–118.0 s 的桌面及腕部视图，金色弧端向目标方向超出初态最靠近目标的边缘，位移明显大于分隔面相对该边缘的5毫米，故系统侧关系失去。越界时刻约为视频 62.5 s。"
+     },
+     {
+      "t": 128.0,
+      "log_step": 70,
+      "text_zh": "推断：视频 120.0–128.0 s 机器人重新双手夹持、抬起并将组合件向近侧移开，约视频 128.0 s 金色件退回系统侧；视频 136.0–472.0 s 各段操作及视频 478.9 s 的主视图、正面视图显示金色件此后位于目标近侧的操作空间。log step 70 的脚本确实命令双手将组合件沿远离目标方向平移；此处时间取自画面，恢复边界估计误差约2秒。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 16.0–60.0 s 左二指夹爪夹持金色件并抬起；视频 62.5 s 金色件脱离夹爪，之后重新夹持弧端、杆部并多次换抓。金色件由左夹爪固定和转动，右夹爪操作银色件的协作记在银色件节点。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 26.0,
+      "log_step": 27,
+      "text_zh": "观察：视频 20.0–27.0 s 机器人双夹爪夹着两件尝试相对抽拉，但银色齿形框仍穿在金色框内；视频 38.0–47.0 s 右夹爪离开后银色件仍挂在金色件上，未解脱。log step 27 的脚本命令右手继续平移20毫米。"
+     },
+     {
+      "t": 232.0,
+      "log_step": null,
+      "text_zh": "观察：视频 128.0–232.0 s 两臂反复改变两件夹角并抽拉；视频 240.0–304.0 s 右手夹住银色件，金色件曾悬挂其上，左手再夹金色件。两件一直存在可见的套接，没有形成独立的银色件入区状态。"
+     },
+     {
+      "t": 478.9,
+      "log_step": 375,
+      "text_zh": "观察：视频 400.0–454.0 s 两件齿口交接处仍相互卡挂；视频 475.75–478.75 s 的逐帧检查及视频 478.9 s 末帧显示银色件仍套在金色件上，且主体伸在绿框右侧，整个轮廓没有进入目标体积。评分器 live_success=0、replay_success=0 且 live_matches_replay=1 与画面一致；log step 375 的在线服务仍报告 done=false。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 16.0–27.0 s、128.0–224.0 s 及后续多段中，左二指夹爪夹金色件、右二指夹爪夹银色件，协同抬升、旋转和抽拉；各手分别夹住一件。视频 232.0–264.0 s 及约视频 450.0 s 的恢复片段中，左手空出，右手单独夹银色件并承受仍套接的金色件，随后左手重新抓取金色件。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 26.0,
+   "log_step": 27,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 24.0–27.0 s 右手继续抽拉时，银色齿形件仍卡在金色件内；后续虽反复重抓和调整角度，至视频 478.9 s 仍未解脱入区。log step 27 执行右手20毫米抽拉，log step 28 的在线安全监测记录该段最大受压接触力194.2牛；这是机器人服务测得的接触载荷，不是 agent 的文字判断。金色件中途的系统侧关系后来恢复，未修复的步骤是银色件解脱入区。"
+  },
+  "unmet": {
+   "gold_on_system_side": true,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 26.0,
+    "log_step": 28,
+    "evidence_zh": "观察：视频 24.0–27.0 s 抽拉没有使两件脱开；视频 38.0–47.0 s 银色件挂在金色框中，视频 128.0–224.0 s 及400.0–448.0 s 的腕部视图反复显示银色齿部与金色环框相互卡住，改变夹角后也未通过，末帧视频 478.9 s 仍套接。log step 28 的机器人服务测得最大受压接触力194.2牛，与有阻挡的抽拉一致。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：实际双手相对位姿和抓持未稳定维持，导致细齿通道的执行反复受阻。log step 317 中 agent 脚本用服务实测TCP减去规划TCP，打印左手y/z误差约+12/-12毫米、右手y/z误差约-10/-12毫米，说明两手相对位置明显偏离计划；log step 325、357 的服务实测左夹爪间隙均降为0，log step 357 同时测得受压接触力92.9牛及约39毫米跟踪误差。视频 400.0–454.0 s 可见两件继续卡挂、左手撤开后金色件挂在银色件上而非保持既定双手姿态。上述证据支持控制问题，但尚不足以把根因唯一细分为抓取设计或反馈控制。"
+   }
+  ],
+  "gap_zh": "没有状态日志或金色件全网格相对固定分隔面的数值读数；金色件的越界与退回按多视图、初态边缘及明显平移推断，视频 62.5 s 和128.0 s 的分界仅为近似，不能精确核验5毫米边界或排除边界附近更短的变化。未提供DAG所列七项原生度量的评分数值，均记null。agent 的相机几何重建不是场景真值，现有证据不能验证整条规划路线的实际可行性，亦不能进一步区分规划误差、抓持内滑转和反馈跟踪各自的贡献。"
+ },
+ "summary_zh": "机器人建立了双手抓持并多次抬升、转动和抽拉两件；金色件留系统侧的关系开局成立，约视频 62.5 s 失去、约视频 128.0 s 恢复，最终为1/2节点成立。银色件始终未解脱入区，视频 478.9 s 仍明显套接且伸出绿框，断点是解脱入区操作中的卡阻，日志还显示相对位姿跟踪偏差和抓持丢失。无状态日志，金色件隐形分隔面的跨越时间及控制问题的细分归因有不确定性。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "已对照视频核查：银色件始终套在金色件上、末帧仍伸出绿框，silver_in_goal 未达成与断点（银色件解脱入区，约视频 26 s）成立；金色件系统侧在视频 63–100 s 明显被带离初态位置，视频 128 s 之后回到近侧，与标注一致。未发现需要改动内容的错误（金色件后段的系统侧恢复为近似推断，文件已在 gap_zh 说明）。",
+    "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": "2026-10-31T00:00:24Z",
+  "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/roboarc-cast-keyhole-03-i01-standard--7swrrxf/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--7swrrxf/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "dc6d592f07b800a2d0949977bfcc218f88599692b1d01dfb46a799e1ee410ada",
+   "video_sha256": "8b212c1e95e0515069b83e053df9ac570210dc4168f4de23e129ab87524ac89f",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-standard--am7ruxx.json

@@ -0,0 +1,279 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-standard--am7ruxx",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 2978,
+   "live_rows": 2978,
+   "sim_steps": 2978,
+   "physics_steps": 29780,
+   "steps": 368,
+   "elapsed_s": 10819.4,
+   "replay_wall_s": 154.1,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-cast-keyhole-03-i01-standard--am7ruxx/replay.mp4",
+  "duration_s": 74.533,
+  "fps": 15.0,
+  "frames": 1118,
+  "sha256": "84e3c2d9772d3797c3e7bea4ce3568a6befc4054bd3b62af83e2936812f64763"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色长框在绿框近侧，与银色件套接，初态满足DAG所述系统侧关系。"
+     },
+     {
+      "t": 60,
+      "log_step": null,
+      "text_zh": "推断：对比视频 0.0、18.0、30.0、52.0、60.0 s 和视频 73.4–74.47 s，金色件主要向画面右侧及近侧滑移，未见向目标远侧推进；结合其初态系统侧位置，按视觉判断该关系持续成立。此判断依据初态轮廓和运动方向，非仅依据未进入绿框；未取得隐形分隔面的逐顶点距离。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频约 56–66 s 左臂二指夹爪夹住金色件近侧横杆，右臂夹爪同时夹银色件作相对拉动；金色件仍在桌面上，有小幅随动。此前左夹爪曾从侧面和上方多次试夹、重新定位。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 30,
+      "log_step": 179,
+      "text_zh": "观察：视频 28.5–30.3 s 右夹爪夹着银色件作短距离向右拉动，金色件随动，两件仍套接且在绿框近侧。log step 179 在线机器人服务返回 control step 1200；按任务40 Hz控制率，1200/40=视频 30.0 s，与所查拉动帧相符。"
+     },
+     {
+      "t": 65,
+      "log_step": 332,
+      "text_zh": "观察：视频 56.0–66.0 s 左夹爪夹金色件、右夹爪反复拉银色件，两件仍近乎平放并保持套接；视频 65 s 附近右夹爪向上退开时，银色件仍在桌面原互锁处，未被携离。"
+     },
+     {
+      "t": 74.47,
+      "log_step": 386,
+      "text_zh": "观察：视频 73.4–74.47 s 的主视、侧视及腕视中，银色长框仍在绿色目标框近侧外面并与金色件重叠套接，整个绿框未承接银色件；后续复夹也未使它解脱或入区。评分器 live_success=0、replay_success=0、live_matches_replay=1，与画面一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：前段右臂二指夹爪独自多次闭合试夹银色框边，并有拉动拼图的实际接触；视频约 56–66 s 左夹爪夹金色横杆、右夹爪夹银色件，双臂配合拉动。夹爪确有包夹接触，但未观察到银色件被可靠携离互锁位置，不能由闭合间隙认定辅助抓取已建立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better"
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "value": null,
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better"
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better"
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "value": null,
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better"
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "value": null,
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better"
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "value": null,
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better"
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "value": null,
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 29.5,
+   "log_step": 179,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 28.5–30.3 s 机器人右夹爪沿银色件长轴拉动时，金色件被一起带动，互锁未解除；视频 56–66 s 改为双臂夹持后仍未完成相对脱出，视频 74.47 s 银色件仍在目标区外。这一步一直未完成，后续复夹未修复。"
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "带动本体",
+    "t": 29.5,
+    "log_step": 190,
+    "evidence_zh": "观察：视频 28.5–30.3 s 右夹爪向右拉银色件，金色长框也向右滑移，两件没有拉开；视频 10–13 s 的较早试拉同样使两件整体右移。agent 在 log step 190 自述“拉动了包括金色件在内的整个组件”，与画面相符。"
+   },
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 60,
+    "log_step": 328,
+    "evidence_zh": "观察：视频 56–65.3 s 左夹爪夹金色横杆、右夹爪向外拉银色框，金色弯端仍扣在银色框内，银色件没有沿所拉方向通过并脱出。log step 328 的在线服务接触监测返回本次读取窗口最大压接触力175.6 N；它是基准机器人服务的传感读数，并非agent估算。该步 control step 2595/40=视频约64.88 s，已在视频64.5–65.0 s的拉动处核对。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "推理/规划",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：agent在log step 170 明确选择沿银色长轴拉出；log steps 269、287、291、294、298 的命令继续增加右末端横向位置，试图以金色件固定加银色件直拉来解除互锁。视频 28.5–30.3 s 及56–66 s显示两件保持近乎共面的互扣形态，直拉使组件随动或卡住；期间主要改变夹爪抓位，未形成让两件相对转位并通过槽口的解锁路径。由明确的动作策略及其几何后果，判断解锁路径规划不足，而非仅从失败或预算耗尽反推。"
+   }
+  ],
+  "gap_zh": "没有状态日志，缺少逐顶点分隔面距离、银色件速度和辅助抓取建立/解除标志；金色件系统侧全程成立为依据初态及画面运动的推断，不能精确核验5毫米隐形分隔面。若干中段log工具返回文本为空或截断，无法确认每次试夹的实际承载状态，故不进一步断言具体辅助抓取故障或失握阈值。",
+  "unmet": {
+   "gold_on_system_side": true,
+   "silver_in_goal": false
+  }
+ },
+ "summary_zh": "评分尝试最终保持金色件留系统侧，按视觉推断为1/2个节点；银色件始终未解脱并进入目标区。机器人先单臂拉动、后双臂夹持对拉，出现带动金色件和互锁阻挡，日志中的轴向直拉策略支持解锁路径规划不足的判断。无状态日志，隐形分隔面的毫米级判定及每次辅助抓取状态仍不能精确核验。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "标注与视频一致：金色件全程留在系统侧，银色件始终未解脱入区（视频0.0/29/66/74.4 s帧），control step 2978/40=74.45 s与视频时长吻合；断点、现象与推理/规划根因有证据支持，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": "2026-10-30T23:52:19Z",
+  "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/roboarc-cast-keyhole-03-i01-standard--am7ruxx/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--am7ruxx/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b5ff3e982164424be1db3636388bfcbcd513c6c426c39c4d7764a97564635263",
+   "video_sha256": "84e3c2d9772d3797c3e7bea4ce3568a6befc4054bd3b62af83e2936812f64763",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-standard--aqktokm.json

@@ -0,0 +1,299 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-standard--aqktokm",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "runtime_binding_ok": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "n_actions": 4633,
+   "live_rows": 4633,
+   "sim_steps": 4633,
+   "physics_steps": 46330,
+   "steps": 1106,
+   "elapsed_s": 10778.6,
+   "video_rendered": 1,
+   "video_pending": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-cast-keyhole-03-i01-standard--aqktokm/replay.mp4",
+  "duration_s": 115.933,
+  "fps": 15.0,
+  "frames": 1739,
+  "sha256": "7569c20aafa89242bf9712fc34e3d493c8d3943c05b29f0c2fd880dffe5b53a1"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，两件呈支起的套接姿态，金色件位于绿框近侧；task/task_dag.json 的 gold_on_system_side.criterion_zh 明确初态满足固定分隔面的系统侧条件。"
+     },
+     {
+      "t": 0.2,
+      "log_step": 69,
+      "text_zh": "推断：对比视频 0.0、0.067、0.133、0.2 s 的主视放大图及右上侧视，拼图姿态变低，但金色件朝目标的远侧边基本留在原处，未见足以支持越界的向目标展开；结合 DAG 给定的初态成立，保留系统侧关系。log step 69 的 gold y[-0.029,0.094]、z[0,0.113] 是 agent 自行分割、反投影得到的可见点范围，含可疑离群点，不是服务的网格边界读数；log step 73 关于越界的自述及 log step 31、79 的点云估计均不用于判定越界。"
+     },
+     {
+      "t": 115.867,
+      "log_step": null,
+      "text_zh": "推断：已检查全程粗抽帧、视频 28.0–46.4 s 的试夹段、视频 106.0–115.6 s 的末次操作及视频 114.933–115.867 s 的逐帧末段。金色件大部分时间保持原处；视频约 109.2–111.6 s 随仍套接的银色件向画面左下转动、滑移，朝目标的边缘未见越过其原有远侧位置，末段仍留在目标近侧。结合初态满足及这些相对位移，记持续满足、最终 true；这是视觉推断，未由状态日志或原生分隔面距离复核，不能仅以未进入绿框代替严格判据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频 6.4–8.5 s，左臂闭合夹爪沿金色件上方和近侧边缘下探、扫过，执行 log step 63 的推倒拼图尝试；接近处被夹爪遮住，金色件没有明确随动，无法确认夹指实际碰到金色件还是从其上方掠过。未见包夹抓握。视频 108.8–111.6 s 的金色件随动来自仍套接的银色件，不能据此虚构右夹爪直接抓金色件。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，银色件与金色件套接，整个组件位于绿色虚线框近侧，银色件右端伸在框外。"
+     },
+     {
+      "t": 40.8,
+      "log_step": 129,
+      "text_zh": "观察：视频 28.0–40.8 s，右臂夹爪反复在银色长条右端及开槽处下探、闭合或张开，随后抬起时银色件仍留在原桌面位置；视频 42.4–46.0 s 又在槽附近尝试，仍未解脱、搬入目标区。log step 104、121、123、125、129 的工具回传也均未给出成功。"
+     },
+     {
+      "t": 111.6,
+      "log_step": 169,
+      "text_zh": "观察：视频 70.0–78.0 s、89.2–96.0 s 及 106.0–114.0 s 的细帧中，右夹爪再次下降到银色件右段，末次接触使整套拼图在桌面转动、滑移，但夹爪抬起后银色件未随之升起。视频 114.933–115.867 s，银色轮廓仍在绿框近侧，与金色件套接，绿框内没有完整银色件；log step 169 回传 done=False。"
+     },
+     {
+      "t": 115.867,
+      "log_step": null,
+      "text_zh": "观察：成功参考视频 675.0 s 中，银色件由右夹爪持有、与左夹爪持有的金色件明显分开，银色轮廓位于绿框上方；本次视频 115.867 s 清楚可见银色件仍在桌面框外。因此未达成不是目标被遮挡造成的未知；参考仅用于辨认形态，不作为本次达成证据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频 33.6 s、38.4–40.0 s 以及 43.2 s，右臂夹指在银色件右端或槽边下压、擦触；后续空手抬起，未见形成能带起银色件的对向夹持，故不记抓握。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "观察：视频 108.8–111.6 s，右臂夹爪在银色件右段接触后使银色件沿桌面转动、滑移，并带动仍套接的金色件；随后夹爪抬起而两件留在桌面。此段只有右夹爪操作，未见两臂同时抓住两件协同解脱。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 28.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 28.0 s，右臂夹爪开始在银色件右端下探；视频 28.0–40.8 s 多次下压、闭合或张开后空手抬起，银色件仍留在桌面，视频 42.4–46.4 s 的槽附近试夹也未带起银色件。视频 108.8–111.6 s 末次操作仅使套接组件沿桌面滑转，视频 114.933–115.867 s 银色件仍在绿框外并与金色件套接。因此首个已证实反复失败且未修复的步骤是银色件解脱入区；视频约 28.0 s 是首次下探的导航锚点。"
+  },
+  "unmet": {
+   "gold_on_system_side": true,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 33.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频 33.6–40.8 s，右臂夹指在银色件右端或槽边下压、擦触后抬起，银色件未随夹爪升起；视频 43.2–46.4 s 再次试夹也空手撤离。视频 108.8–111.6 s 夹爪虽推动整套拼图滑转，抬起后两件仍留在桌面。未建立带起银色件所需的对向夹持，故记没建立接触；并非完全没有擦触，也没有先抓住后掉落的证据。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "没有状态日志、冻结网格或逐件分隔面距离读数。金色件持续留在系统侧是由 DAG 给定的初态及视频中远侧边和相对位移推断，严格的毫米级边界仍不能直接复核；log step 31、69、79、169 的 agent 点云估计及 log step 73 的越界自述不作原生判据，整体 success=0 也不能决定金色件的单独真值。银色件未被夹起可由视频直接确认，但现有画面不能区分夹指深度、槽内间隙、定位误差或互锁阻碍各自的因果作用；log step 168、170、172 对失败原因的 agent 总结未经独立验证，故根因继续留空。原生七项几何、速度及共同确认时间均未在 reward 中报告，值保留 null。"
+ },
+ "summary_zh": "金色件初态满足系统侧条件，开头姿态变低及末段滑转均未显示朝目标越过原有远侧边，按视觉推断持续满足，最终进度为 1/2。右夹爪从视频约 28.0 s 起反复试夹银色件，未建立能带起它的夹持，视频 115.867 s 仍套接在桌面框外，断点为银色件入目标区。金色件的严格分隔面边界及银色件抓取失败的具体根因缺少独立读数，因此保持置信度 6，根因留空。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "金色件在视频 0.0 s 与 0.2 s 的俯视中远侧边基本没动（仅姿态由支起变平），“视频约 0.1 s 越过分隔面并失去”只靠 agent 的点云估计和叙述（log step 73、79），而这些点云含离群点（gold z 到 0.113 m、y 下限 −0.029 m），不是对毕业局的可信观测；断点因此应落在银色件步骤，而不是金色件。",
+    "n_issues": 4,
+    "n_major": 3
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "金色件初态满足、视频中未见越过分隔面，银色件始终套接在框外，断点为银色件入区、现象为没建立接触，根因因缺乏独立证据留空并写明 gap_zh，均与帧一致。",
+    "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": "2026-10-31T00:17:27Z",
+  "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/roboarc-cast-keyhole-03-i01-standard--aqktokm/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--aqktokm/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--aqktokm/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--aqktokm/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "e040841e314e6dd88a9f71f54dd2accef476e94afe23020657a3a4c43462e1d7",
+   "video_sha256": "7569c20aafa89242bf9712fc34e3d493c8d3943c05b29f0c2fd880dffe5b53a1",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-standard--egdgok4.json

@@ -0,0 +1,287 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-standard--egdgok4",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 36880,
+   "live_rows": 36880,
+   "sim_steps": 36880,
+   "physics_steps": 368800,
+   "steps": 9068,
+   "elapsed_s": 6947.2,
+   "replay_wall_s": 1774.4,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-cast-keyhole-03-i01-standard--egdgok4/replay.mp4",
+  "duration_s": 922.067,
+  "fps": 15.0,
+  "frames": 13831,
+  "sha256": "9bab5cf289b812eab306d75cc042fbcd9f5626cdaa747820d2fa1f1eb336135e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色长框位于绿框近侧的初始拼图位置；本节点按任务初态从视频 0.0 s 成立。"
+     },
+     {
+      "t": 150,
+      "log_step": null,
+      "text_zh": "推断：对照视频 0.0 s 的初始位置与视频 35–63 s、80–145 s、215–269 s、340–378 s、490–538 s 的主视图及侧视，机器人主要在拼图近侧抬升、翻转金色件，没有辨认出金色件向目标远侧跨过初态分隔位置的运动；抬高后主视图与绿框重叠属于透视投影，不能单据此记越界。按多视图空间关系保留系统侧区间，隐形分隔面的毫米级余量未复算。"
+     },
+     {
+      "t": 921.2,
+      "log_step": null,
+      "text_zh": "推断：视频 920.8–922.0 s 金色件在左夹爪下方朝机器人近侧斜悬，较初态更靠近机器人，整个可见轮廓远离绿框近边，结合侧视判断终态仍在系统侧；此判断不是由 success=0 或“未进绿框”单独推出。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 35–63 s 左臂二指夹爪夹住金色框并将其抬起；视频 130–150 s 再夹金色件，其后多次开爪、重抓。右臂对银色件的协同抓持记在银色件节点。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 银色件与金色件相套，均在绿框近侧，银色件没有完整进入目标区。"
+     },
+     {
+      "t": 150,
+      "log_step": null,
+      "text_zh": "观察：视频 58–63 s 左爪抬金色件时银色件随互锁处悬起；视频 90–95 s 右爪已夹住银色件，左爪离开后金色件仍挂在银色件上。视频 145–168 s 左右夹爪分别夹金、银两件，腕视图可见银色横档仍穿在金色槽口，未解脱。"
+     },
+     {
+      "t": 700,
+      "log_step": null,
+      "text_zh": "观察：视频 0–920 s 的全程每隔约10秒抽帧及关键转动、换抓处的加密帧中，两件始终相连；视频 700–920 s 双臂主要在原拼图近侧小幅调整，没有把完整银色轮廓送进绿框。"
+     },
+     {
+      "t": 922,
+      "log_step": null,
+      "text_zh": "观察：视频 920.8、921.0、921.27、921.47、921.73、922.0 s 银色框仍被右爪夹持并与金色件互锁，银色件大部在绿框近侧之外，故入区节点终态为假。成功参考录像的参考视频 675.0 s 则显示银色整件已独立悬于绿框内、金色件在左下方，仅用来核对视觉形态。"
+     },
+     {
+      "t": null,
+      "log_step": 247,
+      "text_zh": "观察：agent 在 log step 247 自述两件尚未分离并决定停止；评分器报告 live_success=0、replay_success=0、deterministic=1、live_matches_replay=1，且无重置或恢复。节点为假依据本次视频中未解脱及未完整入区的画面，评分结果作为旁证。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 90–95 s 右夹爪夹住银色件，视频 145–168 s 及后续解脱尝试中左夹爪夹住金色件、右夹爪夹住银色件，两臂各自通过二指相向夹持配合平移和转动。此前左夹爪单独抓金色件抬起互锁组件，后续也多次松开左爪换抓。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better"
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "value": null,
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better"
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "value": null,
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better"
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "value": null,
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better"
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "value": null,
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better"
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "value": null,
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better"
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "value": null,
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 156,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 148–168 s 机器人双臂分别夹住金色件和银色件，尝试相对移动，但银色横档一直卡在金色槽内；之后虽反复改姿与重抓，视频 920.8–922.0 s 仍未解脱或入区。本次从未修复的是银色件入目标区步骤中的解脱过程。"
+  },
+  "unmet": {
+   "gold_on_system_side": true,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 156,
+    "log_step": null,
+    "evidence_zh": "观察：视频 148、156、164 s 腕视图显示银色横档横穿金色框的槽口，机器人改变右爪位置时横档仍被槽壁限制；视频 215–269 s 与340–378 s 改姿、松爪后两件仍相套。视频 920.8–922.0 s 互锁也未消失，阻断了银色件独立移入目标区。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：机器人执行相对位姿调整时，实际运动与脚本设定的精细位移不一致，未能稳定控制两件沿槽通过。log step 152–154 中脚本将相对坐标 s 设为40毫米而其估计值仍约45毫米，依据机器人本体反馈计算的右TCP目标与实测偏差在金色件坐标系为[-7.4,17.6,-1.1]毫米；log step 237 再次输出STUCK，log step 238 的视觉估计与TCP推算也不一致。视频 148–168 s 及末段相套不脱与这些执行问题相符；证据不足以把控制进一步唯一归为抓取设计、反馈控制或参数问题。"
+   }
+  ],
+  "gap_zh": "没有状态日志，也没有金色件全部网格顶点相对固定分隔面的逐帧距离；金色件系统侧区间由主视图、侧视和初末姿态推断，无法排除未辨认出的毫米级短暂越界。评分器未报告DAG列出的七项原生数值，均填null。日志中的相对位姿量是agent脚本的计算，不当作物体真实状态；未将agent工作时间换算为视频时间，也未据agent自述断言辅助抓取在某一视频时刻断开。缺少精确物体姿态及接触对，不能进一步区分几何建模误差和抓持、跟踪误差各自的贡献。"
+ },
+ "summary_zh": "机器人先后抓住并抬起两件，随后双臂反复改姿尝试解脱；金色件按多视图空间关系判断留在系统侧，银色件始终未解脱且未完整入区，终态为1/2。断点在银色件入目标区的操作中，横档受金色槽口阻挡，日志支持相对位姿执行控制存在偏差。缺少状态日志和隐形分隔面的精确距离，金色件区间为视觉推断，原生阈值量未报告。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点值（金色件系统侧终态真、银色件入区终态假）、断点（silver_in_goal，操作中）与现象（两件互锁造成的被挡）均有视频帧支持，与评分器 success=0 一致；根因停在“控制”并注明证据不足以细分，较保守且��误。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": "2026-10-30T23:54:08Z",
+  "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/roboarc-cast-keyhole-03-i01-standard--egdgok4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--egdgok4/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "97187a6b97592b056692f4238ba30f5f61ebe0f46b995bbb076d340412ab7f36",
+   "video_sha256": "9bab5cf289b812eab306d75cc042fbcd9f5626cdaa747820d2fa1f1eb336135e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-standard--hwqggkq.json

@@ -0,0 +1,303 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-standard--hwqggkq",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 43831,
+   "live_rows": 43831,
+   "sim_steps": 43831,
+   "physics_steps": 438310,
+   "steps": 42598,
+   "elapsed_s": 9319.4,
+   "replay_wall_s": 2621.4,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-cast-keyhole-03-i01-standard--hwqggkq/replay.mp4",
+  "duration_s": 1095.867,
+  "fps": 15.0,
+  "frames": 16438,
+  "sha256": "9408dd615a55eadbf0a812e1ecb0a18c1c342aa9db1e75f0df771f6eb9441d6c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 0,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      466.5
+     ]
+    ],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色件在绿色目标框近侧；视频 10.0–460.0 s 的抽帧中金色件基本留在原位置，右臂主要接触银色件，左臂没有固定金色件。"
+     },
+     {
+      "t": 466.5,
+      "log_step": null,
+      "text_zh": "推断：视频 465.0–468.4 s 的密集帧显示右夹爪接触银色件时，两件一起转动，金色钩头向目标方向移出初态轮廓；按初态最靠目标点外移5毫米的固定分隔面，约在视频 466.5 s 失去系统侧关系，时间误差约1秒。"
+     },
+     {
+      "t": 1095.8,
+      "log_step": null,
+      "text_zh": "推断：视频 480.0–1080.0 s 的抽帧和视频 1094.8–1095.8 s 的末段显示金色钩头保持转向目标的位移，未退回原系统侧轮廓；log step 479、533、558 的点云分析也报告金色件最靠目标端向目标移动，但这些是 agent 的估计，不是评分器逐点判据。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 164,
+      "log_step": null,
+      "text_zh": "观察：视频 90.0–120.0 s 右夹爪在银色件上反复闭合和调整，视频 158.0–168.0 s 夹爪从银色件右端撤起，银色件仍在原桌面位置，没有随夹爪离桌；视频全程每20秒抽帧中均未见银色件整体进入绿框。"
+     },
+     {
+      "t": 624.3,
+      "log_step": 486,
+      "text_zh": "观察：视频 606.0–624.3 s 右臂在目标框上方移动，但银色件和金色件仍留在近侧桌面；log step 486 的控制脚本在脚本自设的“双指受阻”检测为 False 后仍执行抬升、转向和搬运。该步骤服务计数 control_steps=24972；推断：24972/40=视频 624.3 s，与此处空手停在目标上方的画面一致，43831/40=1095.775 s 也与视频终点一致。"
+     },
+     {
+      "t": 1038,
+      "log_step": null,
+      "text_zh": "观察：视频 1008.0–1038.0 s 右夹爪在银色件上闭合后向右移动，银色件主体仍在桌面近侧；log step 550、564 的 agent 自述“已分离”不代表银色件已入区。"
+     },
+     {
+      "t": 1095.8,
+      "log_step": 558,
+      "text_zh": "观察：视频 1080.0–1095.8 s 包括最后一秒，右夹爪在银色件上方反复压拖，银色长框仍明显伸在绿色目标框近侧之外；评分器 live_success=0、replay_success=0，与画面中的入区未完成一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "压/擦",
+      "note_zh": "观察：视频 90.0–120.0 s、320.0–460.0 s 及末段，右夹爪指端压在银色件表面或孔边，未见可靠的双指夹持搬运。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "观察：视频 465.0–468.4 s 右夹爪拨动银色件并使互锁两件一起转动；视频 1008.0–1095.8 s 又沿银色件做横向推拨及向目标方向的拖动尝试。左臂未参与操作。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 164,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 158.0–168.0 s 右夹爪在前面闭合尝试后从银色件右端空手撤起，银色件仍留在桌面；以后持续尝试抓取、解脱和推拖，视频 606.0–624.3 s 出现空手搬运，直到视频 1095.8 s 仍未入区。该步早先的失败始终没有通过完成入区得到修复；金色件约在视频 466.5 s 的被动转动是后续丢失的关系。"
+  },
+  "unmet": {
+   "gold_on_system_side": false,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 164,
+    "log_step": null,
+    "evidence_zh": "观察：视频 158.0–168.0 s 夹爪从银色件右端撤起而银色件留桌，视频 606.0–624.3 s 又空手走完搬运；视频 1080.0–1095.8 s 右夹爪继续在银色件处压拖，银色件仍未整体进入目标框。log step 557 输出夹指驱动 effort=-13.26，随后两次向目标运动的夹爪实测位移为 -1.81毫米、-0.01毫米，脚本以 no gain 停止；这些不是银色件的位移测量，也不是接触力牛顿数。推断：早期夹取没有建立能带走银色件的对向抓握接触，后续压拖也未完成搬运；未另证实独立的接触力异常，不将同一未完成接触操纵重复标注。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频 / agent 自述",
+    "evidence_zh": "观察：log step 4 服务规格将 proprio[56:58] 定义为按左右臂顺序给出的两个夹爪开口；log step 485、490、494 编辑的脚本却将 pr[56]、pr[57] 作为同一夹爪两侧的 STALL[1]、STALL[2] 反馈。log step 486 实际执行 log step 485 编写的 carry1.py，输出 both stalled: False、lift 阶段 conv=False 后仍执行转向和搬运，视频 606.0–624.3 s 显示右夹爪空手而银色件留桌；log step 495 执行 carry5.py 后则输出 WELD (both stalled): False、no weld - stop。推断：反馈字段解释和持物验证不可靠，且未验证持物仍继续运动，可解释空手搬运的控制行为；字段误读是否直接导致抓不起或影响仿真的辅助抓取，现有证据不能确定。竞争解释：agent 在 log step 490、494、496 自述条带约31毫米宽、两条带不等高、南侧条带比垫片可达位置远约20毫米，提示抓取接触几何可能不适配；这些尺寸与接触判断待核实，不能当作已观察到的物理原因。两种解释尚不能区分，保留上位“控制”，可行替代为未知。",
+    "alt_feasible": "未知"
+   }
+  ],
+  "gap_zh": "无状态日志或评分器的逐顶点分隔面距离、速度与保持时间；金色件越界及其约视频 466.5 s 的时刻由初末轮廓和转动帧推断，不能精确到毫米或物理步。孔口局部被夹爪遮挡，不能独立确认 agent 在 log step 550、564 声称的解脱是否完成；银色件明显未整体入区不受此缺口影响。七项原生数值指标均未报告。 抓取失败中反馈字段误读与接触几何不适配的贡献无法区分；几何尺寸及垫片接触位置仅有 agent 自述，未独立验证可行替代抓法。"
+ },
+ "summary_zh": "金色件留系统侧在开局成立，约视频 466.5 s 随银色件被拨转而丢失；银色件入目标区始终未达成，末态为0/2。右夹爪多次未建立可搬运的夹持，空手走过目标上方，后续压拖仍未完成入区；控制脚本误读左右夹爪开口且未可靠验证持物，抓取几何不适配是另一待核实的解释。无状态日志，金色件越界时刻、孔��解脱与具体抓取根因仍有缺口，因关键时刻和因果解释含推断将 conf 从7降为6；标准模式后写入脚本的未执行计划不计进度。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "进度节点（金色件约视频466.5 s失去、银色件始终未入区）、clock 换算（control_steps/40）和断点位置与视频一致；问题在 phenomena[1]「接触力异常」：证据只有夹指驱动 effort 读数（标签自己也承认不是力），没有力不对的可见或传感证据，且这一步是同一未建立持物的延续，应删去或改写。根因「控制」的因果链（字段误读导致没验证持物）是推断，与 agent 自述的垫片够不到南侧条带（抓取几何）竞争，建议在证据中写明并核对 conf。",
+    "n_issues": 2,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "金色件初态成立、约视频466.5 s转动越界丢失，银色件始终未入区，断点在silver_in_goal，与首末帧及468 s帧一致；一轮问题已修，现表述（没建立接触 + 控制，alt_feasible未知）有证据且标注了推断，未发现需改动处。",
+    "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": "2026-10-31T00:17:44Z",
+  "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/roboarc-cast-keyhole-03-i01-standard--hwqggkq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--hwqggkq/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--hwqggkq/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--hwqggkq/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c6b2a5cc458a613ea49e989f7bf013df04fd36c4a3d7b026a7e5d5aaa2e058d1",
+   "video_sha256": "9408dd615a55eadbf0a812e1ecb0a18c1c342aa9db1e75f0df771f6eb9441d6c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-standard--q3sknfo.json

@@ -0,0 +1,297 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-standard--q3sknfo",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 20599,
+   "live_rows": 20599,
+   "sim_steps": 20599,
+   "physics_steps": 205990,
+   "steps": 5408,
+   "elapsed_s": 10809.5,
+   "replay_wall_s": 1720.4,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_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/roboarc-cast-keyhole-03-i01-standard--q3sknfo/replay.mp4",
+  "duration_s": 515.067,
+  "fps": 15.0,
+  "frames": 7726,
+  "sha256": "e3643ea8f60125a89f331ec97bc375acbcc2915bb500403de59de2c991712584"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      9.5
+     ],
+     [
+      448.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色件在绿框近侧，与银色件套接；结合 DAG 明定的初态，金色件留系统侧从开局成立。"
+     },
+     {
+      "t": 9.5,
+      "log_step": 146,
+      "text_zh": "推断：视频 9.0–12.0 s 左夹爪转向并压低金色件，金色件展开后靠目标的边缘较初态前移，估计从视频约 9.5 s 起有部分越过仅距初态最前点 5 毫米的分隔面。log step 146 中 agent 根据相机深度重建的金色可见点也向目标方向扩展；该重建仅作支持，边界时刻是视觉估计。"
+     },
+     {
+      "t": 240.0,
+      "log_step": null,
+      "text_zh": "观察：视频 218.0–240.0 s 金色件被夹爪带动、转成左端朝目标的斜姿；视频 240.0 s 左端伸至绿框近边以内，视频 250.0–420.0 s 仍保留这一明显前伸姿态。"
+     },
+     {
+      "t": 448.0,
+      "log_step": null,
+      "text_zh": "推断：视频 432.0–447.0 s 金色件仍斜向目标，视频约 448.0 s 随右爪操纵银色件向近侧转移后，整个金色轮廓退到初态最前边缘的近侧，故估计在视频 448.0 s 恢复系统侧关系。"
+     },
+     {
+      "t": 515.0,
+      "log_step": null,
+      "text_zh": "观察：视频 450.0、465.0、480.0、500.0 s 及视频 514.1–515.0 s，金色件随银色件移到画面右下方，远离绿框及初态最前位置；末帧金色件仍在系统侧，但挂在银色件上。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "视频 240.0–300.0 s 左二指夹爪夹住金色件左端，右二指夹爪夹住银色件，配合拉动和转动；早期还有指尖抵推金色件。视频约 448.0 s 后左爪张开，金色件经与银色件的互锁被右臂间接带回近侧。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 40.0,
+      "log_step": null,
+      "text_zh": "观察：视频 32.0–58.0 s 右夹爪反复调整并夹持银色件，银色齿条/横杆仍穿在金色件开口内；银色件未离开原拼图一侧。"
+     },
+     {
+      "t": 240.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–510.0 s 的每隔约 10 秒抽帧以及视频 200.0–239.0、432.0–451.0 s 的逐秒检查，均未出现银色件解脱后完整进入绿框的形态；视频 240.0 s 双爪拉转两件时仍互锁。"
+     },
+     {
+      "t": 480.0,
+      "log_step": null,
+      "text_zh": "观察：视频 465.0–480.0 s 右臂抬起银色件，同时把金色件带起悬挂；视频 500.0 s 及视频 514.1–515.0 s，金色件仍扣在银色件下端齿槽处，两件在画面���下方，绿框内空着。"
+     },
+     {
+      "t": 515.0,
+      "log_step": null,
+      "text_zh": "观察：inputs/run.json 的评分字段 success=0、live_success=0、replay_success=0，且 live_matches_replay=1、deterministic=1；与视频末段未解脱、未入区相符。成功参考的 oracle 视频 675.0 s 则显示两件分开，银色件位于绿框上方，金色件由左爪留在近侧，仅用于辨认完成形态。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "视频 240.0–300.0 s 左二指夹爪夹住金色件，右二指夹爪夹住银色件，协同尝试齿槽解脱；视频 448.0–515.0 s 转为右爪单独持银色件，金色件仍套挂其上，左爪张开。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 40.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 40.0 s 起，右夹爪已在夹持并拉动银色件，但其齿槽与金色件开口持续相扣；视频 200.0–240.0 s 的拉转、视频 432.0–451.0 s 的换姿及视频 465.0–515.0 s 的抬起均未使两件解脱。银色件入区这一步始终未完成；金色件中途前伸后在近侧恢复，故断点仍为银色件入区。"
+  },
+  "unmet": {
+   "gold_on_system_side": true,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 40.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 40.0–58.0 s 右夹爪拉动时银色齿条仍卡在金色件弧形开口内；视频 240.0 s 的腕部视图仍能看到两件在齿槽处交叉套接，视频 465.0–480.0 s 右臂提起银色件时金色件也被带起。视频 514.1–515.0 s 金色件依然扣在银色开口内，解脱路径被两件自身的接触阻断。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "推理/规划",
+    "log_step": 194,
+    "evidence_zh": "推断：log step 189 的脚本用连续的 +X、-Y 小步推进，log step 194 接着以 +X 直拉作为脱出路径；其实际服务反馈在 log step 214 记录各段接触压力 47.9–64.8 N，log step 430 又记录同类 +X 分块拉动只到第二块便因 63.4 N 暂停。结合视频 40.0–440.0 s 齿槽始终相扣，以及视频 465.0–515.0 s 两件一起被提走，所选局部平移和转动顺序未形成让两件齿口依次通过的解锁路径，最终无法进入单独搬运银色件的阶段。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志或逐节点评分读数；分隔面不可见，且未提供冻结网格。金色件中段明显前伸、末段远离目标可判，但视频约 9.5 s 的最早越界及视频约 448.0 s 的恢复仅为姿态与初态相比较的估计，不能精确验证 5 毫米边界或排除早期微小往返；相邻时刻有约数秒不确定性。银色件明显未解脱且不在目标内，因此不需要凭画面估算其毫米级边距和速度阈值；全部原生连续指标未由评分器报告，填 null。根因是对执行路径的推断，缺少真实接触几何与期望轨迹的逐步对照。"
+ },
+ "summary_zh": "金色件留系统侧在初态成立，中途向目标方向前伸后于视频约 448.0 s 恢复，末段为 1/2 节点成立；银色件从未解脱并完整进入绿框。机器人反复拉动、转动仍被两件自身的齿槽互锁阻断，末段右臂把两件一起提起，依据脚本和可见结果推断主要问题为解锁路径的推理/规划。无状态日志，金色件最早越界及恢复的精确时刻、严格网格边界不能确定。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点值（金色件留系统侧末段成立、银色件未���区）、断点 silver_in_goal、被挡/碰撞现象与推理/规划根因均与视频及评分字段（success=0）相符；金色件越界/恢复的时刻本身是视觉估计，文件已如实标注。",
+    "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": "2026-10-31T00:01:57Z",
+  "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/roboarc-cast-keyhole-03-i01-standard--q3sknfo/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--q3sknfo/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "6f68b51b6e4fafcee7e764fd9cbe28cc5d1e6a753e365333d357361d84230ec7",
+   "video_sha256": "e3643ea8f60125a89f331ec97bc375acbcc2915bb500403de59de2c991712584",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_03/trajectories/roboarc-cast-keyhole-03-i01-standard--ttqvjxw.json

@@ -0,0 +1,314 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-03-i01-standard--ttqvjxw",
+ "taskKey": "roboarc/cast_keyhole_03",
+ "dag_sha256": "b3c76f087b01b1e9b92b7f49cd79a4319c6f7e66e449d9fd184127cc33a1417f",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "runtime_binding_ok": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "gpu_backend": 1,
+   "finalized": 1,
+   "n_actions": 3010,
+   "live_rows": 3010,
+   "sim_steps": 3010,
+   "physics_steps": 30100,
+   "steps": 737,
+   "elapsed_s": 9773.9,
+   "replay_wall_s": 161.6,
+   "video_rendered": 1,
+   "gpu_lost": 0,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "gave_up": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resume_state_mismatch": 0,
+   "resume_max_tcp_dev_m": 0.0,
+   "resume_max_joint_dev": 0.0,
+   "instance_mismatch": 0,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "grader_error": 0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.307,
+   "peak_processes": 7,
+   "peak_disk_gb": 0.381,
+   "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/roboarc-cast-keyhole-03-i01-standard--ttqvjxw/replay.mp4",
+  "duration_s": 75.333,
+  "fps": 15.0,
+  "frames": 1130,
+  "sha256": "cf4ff0b61affe683e4a6e703315bc8fc8767bf8b2c94c251d5a634a8ae5b4ef5"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色件与银色件套接，金色件在绿色目标框近侧；DAG 明确初态满足系统侧关系。"
+     },
+     {
+      "t": 20,
+      "log_step": null,
+      "text_zh": "观察：视频 10、12、20、28 s 金色件转平并有偏航，仍在原拼图位置附近与银色件套接；画面未显示金色件明显向绿色目标框平移。固定分隔面不可见，不能由这些画面确定毫米级越界。"
+     },
+     {
+      "t": 30,
+      "log_step": null,
+      "text_zh": "观察：视频 29–33 s 左夹爪将金色件拉转成立放；这是可见的姿态变化，无法据此确定此前曾越过固定分隔面或此刻恢复该关系。"
+     },
+     {
+      "t": 75.267,
+      "log_step": 298,
+      "text_zh": "观察：视频 33–75.267 s 金色件保持近乎竖直，左夹爪随后换抓左端并在末段向画面右侧移动金色件；视频 74.25、74.5、74.75、75.0、75.267 s 其轮廓仍留在原拼图一侧。"
+     },
+     {
+      "t": 75.267,
+      "log_step": null,
+      "text_zh": "推断：依据视频 0.0 s 的已满足初态，以及视频 10、12、20、28、33、48、70、72、74.5、75.267 s 金色件始终位于原拼图一侧的可见位置关系，将金色件留系统侧记为连续区间并保持 final=true；不再从 agent 自行拟合的点云估计构造失去与恢复事件。此判断是视觉推断，严格网格分隔面未获测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 5–9 s 左二指夹爪合拢夹金色件，提举时失去夹持；视频约 28–33 s 再夹持并拉转金色件，随后重新抓其左端，视频 43–75.267 s 持住金色件配合右臂操作。"
+     }
+    ]
+   },
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 11.9,
+      "log_step": 68,
+      "text_zh": "观察：视频 5–9 s 两夹爪分别合拢接触两件拼图；视频约 9.3–11.9 s 双臂抬起而两件均留在桌面，初次提举没有把拼图持续带离原处。log step 68 的动作名虽为提起两件，但不能据此记为已拿起或解脱。"
+     },
+     {
+      "t": 48,
+      "log_step": 146,
+      "text_zh": "观察：视频 30–48 s 左臂将金色件转成立放并换握，银色件仍横在桌面、与金色件交叠套接，绿色目标框为空。"
+     },
+     {
+      "t": 71.75,
+      "log_step": 290,
+      "text_zh": "观察：视频 68.5–71.5 s 右夹爪带着银色件右端抬起转动，左夹爪夹金色件向右移；视频约 71.75 s 银色件从右夹爪脱落，视频 72.0–72.5 s 回到桌面，未被送入绿色框。"
+     },
+     {
+      "t": 75.267,
+      "log_step": 298,
+      "text_zh": "观察：视频 73.5–75.267 s 右夹爪撤离后为空，银色件留在金色件右侧的桌面，整个绿色目标框内没有银色件；最后一秒各帧未见重新抓起或入区。评分字段 live_success=0、replay_success=0、live_matches_replay=1，与此一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 5–9 s 左夹爪夹金色件、右夹爪夹银色件后尝试一起提举；后续左臂多次单独换握金色件，视频约 60–71.5 s 左夹爪持金色件，右二指夹爪夹银色件中央附近并抬转，视频约 71.75 s 银色件脱手。两只夹爪各自夹住一件配合解脱。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "silver_horizontal_border_clearance",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_min_height_above_table",
+   "unit": "m",
+   "threshold": -0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_height_above_table",
+   "unit": "m",
+   "threshold": 0.24454,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "silver_max_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "remainder_max_signed_distance_goalward_of_separator",
+   "unit": "m",
+   "threshold": 0,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation_time",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 71.75,
+   "log_step": 290,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 68.5–71.5 s 机器人用左夹爪移动金色件、右夹爪抬转银色件进行解脱；视频约 71.75 s 银色件失去右夹爪夹持并落回桌面，视频 75.267 s 仍未重新抓起或进入绿色框，最终未达的是银色件入区步骤。"
+  },
+  "unmet": {
+   "gold_on_system_side": true,
+   "silver_in_goal": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 71.75,
+    "log_step": 290,
+    "evidence_zh": "观察：视频 69.0–71.5 s 银色件随右夹爪倾斜抬起，视频约 71.75 s 突然转平落回桌面，随后右夹爪空手上撤。log step 290 的脚本保持原夹爪闭合指令，log step 304 的该段传感记录却显示右夹爪开口由约38.9 mm缩至16.4 mm、夹持驱动力由约 -22 N降至接近0，与物体已经脱离夹爪一致；主动开爪撤退是在后续 log step 298。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 285 已读到右 TCP 目标约[30.58,-28.27,45.52] mm而实测约[36.23,-14.11,37.6] mm；log step 290 仍从指令关节值建立固定持物变换，整块提交双臂关节插值、沿程不据物体或接触反馈修正。该调用结束时左 TCP 的目标与实测 x 相差约50 mm，右 TCP 的目标与实测 z 相差约30 mm，服务接触监测峰值1261.9 N；视频 69–71.75 s 对应抬转中滑脱，说明控制未能维持预定双臂相对运动和夹持。证据不足以进一步区分接触设计、接触阻挡与跟踪控制，保留在“控制”层级。"
+   }
+  ],
+  "gap_zh": "无状态日志或逐物体接触对，评分字段未报告任何 DAG 原生几何、速度或共同保持时间指标。视频约 9–29 s 金色件转平并有偏航，是否越过固定分隔面无法由帧判定；log step 91、94、125 是 agent 自行编写程序拟合相机点云的结果，并非评分器网格或分隔面读数，故不据此标注失去与恢复。金色件连续留系统侧是依据初态与可见位置关系作出的视觉推断，严格网格边界仍未核实。夹爪附近有遮挡，无法确定末段滑脱由夹持几何、两臂接触还是负载下跟踪偏差主导。log step 311 记录模型提供方过滤中断；中断前视频已出现上述未修复的物理失败，因此按该物理失败标注，不把过滤或剩余预算当成根因。成功参考录像仅用于比较解脱和入区形态，不作本次事实证据。"
+ },
+ "summary_zh": "依据已满足的初态及视频中的位置关系，金色件留系统侧记为连续成立，最终为1/2节点成立；银色件始终未入绿色目标区。视频约 71.75 s 双臂解脱动作中银色件从右夹爪滑脱，落回桌面且未修复，日志中的执行偏差支持控制层面的归因。无状态日志，视频约 9–29 s 金色件偏航时是否越过隐形分隔面无法精确核实，连续区间为视觉推断，故置信度保持6；log step 311 的过滤中断发生在该物理失败之后。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "银色件未入区、断点在 silver_in_goal 的判断与帧一致；但 gold_on_system_side 在视频约 9.3–29.5 s 的“失去并恢复”只来自 agent 点云估计（叙述），帧中金色件始终留在原拼图位置，不应作为观测写入区间。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "第一轮问题已修正；银色件始终未入绿框、断点在 silver_in_goal、约 71.75 s 右夹爪脱落与帧一致，金色件连续成立作为视觉推断并在 gap_zh 说明，控制归因标注为推断且层级保守。",
+    "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": "2026-10-31T00:14:46Z",
+  "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/roboarc-cast-keyhole-03-i01-standard--ttqvjxw/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--ttqvjxw/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--ttqvjxw/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-03-i01-standard--ttqvjxw/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "25cb483ad56306bcd91e744208e23165d7362fc57bef0eef0a2c51ee3729efe8",
+   "video_sha256": "cf4ff0b61affe683e4a6e703315bc8fc8767bf8b2c94c251d5a634a8ae5b4ef5",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_04/trajectories/roboarc-cast-keyhole-04-i01-privileged--cfhdrvy.json

@@ -0,0 +1,296 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-04-i01-privileged--cfhdrvy",
+ "taskKey": "roboarc/cast_keyhole_04",
+ "dag_sha256": "70720774d01cbeadb5de0aa235fa5558388d350a894bac65e2e2f759e5f7b4c7",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 4339,
+   "submitted_actions": 4339,
+   "sim_steps": 4339,
+   "physics_steps": 43390,
+   "sim_time_s": 108.47500000000754,
+   "final_reward": 0.0,
+   "replay_wall_s": 388.3,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 59,
+   "agent_resets": 1,
+   "agent_sim_steps": 111017,
+   "agent_records": 3,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "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/roboarc-cast-keyhole-04-i01-privileged--cfhdrvy/replay.mp4",
+  "duration_s": 108.533,
+  "fps": 15.0,
+  "frames": 1628,
+  "sha256": "8b4b9bd9d63f502a60e0eaa3bb439c7ccce61fc986b41cf116c24cba83c68e13"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 0,
+  "nodes": {
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0 s，银色细长件与金色宽槽件在绿色目标框前方相互套接，银色件未入区。"
+     },
+     {
+      "t": 24.0,
+      "log_step": null,
+      "text_zh": "观察：视频 20–28 s，机器人左夹爪夹住金色件、右夹爪夹住银色件；视频 30–64 s 左夹爪退出后，两件仍在原处保持相互套接的排列，没有出现解锁后银色件单独移开的状态。"
+     },
+     {
+      "t": 75.0,
+      "log_step": null,
+      "text_zh": "观察：视频 74–86 s，机器人右夹爪抬起、转动银色件，金色件的圆端仍与银色件环部连在一起，金色件随之离开原位置并被带向目标框；主视和右腕视图均未见两件脱开。"
+     },
+     {
+      "t": 108.467,
+      "log_step": null,
+      "text_zh": "观察：视频 92–108.467 s，包括视频最后一秒的连续抽帧，右夹爪保持银色件近竖直，金色宽槽件仍挂在银色件下端并伸入目标框附近；银色件没有满足与金色件脱开后独立入区的完整判据。成功参考视频 56、64、66 s 中则能看到两件脱开、左夹爪持金色件、右夹爪单独搬运银色件。"
+     },
+     {
+      "t": 108.467,
+      "log_step": 294,
+      "text_zh": "推断：log step 294 的 rec3 保存了 4339 行动作，服务状态为仿真 108.475 s，均与本次评分器 n_actions=4339、sim_time_s=108.475 一致，且其末态与视频吻合，故用作交付对应关系的辅助证据；不把其他调试运行的“已分离”自检当成本次节点进度。评分器 strict_goal_reached=0。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 20–28 s，左二指夹爪夹住金色宽槽件，右二指夹爪夹住银色细长件，配合尝试分离；两爪分别夹持各自的物件。随后左爪松开退出，右爪继续夹持银色件，视频 74–108.467 s 的搬运又通过两件间的套接接触带走金色件。"
+     }
+    ]
+   },
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0.0,
+      76.0
+     ]
+    ],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 4,
+      "text_zh": "观察：视频 0 s 金色件位于目标框前方的初始拼图侧；log step 4 的初态金色件 y 上界为 0.08076 m，任务固定分隔面位于初始最靠近目标点之外 5 mm，初态留系统侧成立。"
+     },
+     {
+      "t": 64.0,
+      "log_step": 292,
+      "text_zh": "观察：视频 20–30 s 左夹爪夹住金色件并略向左移，随后松开退开；视频 32–64 s 金色件保持在目标框前的桌面位置。log step 292 的 rec3 诊断在松手后���录金色件 y 上界 0.0627 m，小于脚本使用的分隔面 0.0857 m，与画面位置一致。"
+     },
+     {
+      "t": 76.0,
+      "log_step": null,
+      "text_zh": "推断：视频 72 s 金色件仍在目标框前的原桌面位置；视频 74–80 s，右夹爪抬起银色件，仍套接的金色件离开原位、被斜向带起并向目标框移动。结合主视、侧视及腕视，以及固定分隔面仅在初态金色件最靠近目标的点之外 5 mm，将越过分隔面的节点丢失锚点估在视频 76 s，时间不确定约 ±2 s；该时刻为多视图运动线索支持的近似推断，并非直接看见不可见平面的毫米级越界。此后至视频 108.467 s 金色件一直随银色件停在目标方向，未返回原侧。"
+     },
+     {
+      "t": 108.467,
+      "log_step": 294,
+      "text_zh": "推断：视频 96–108.467 s 可见金色件下部已接近桌面、圆端仍连接银色件并位于目标方向；与评分动作数和仿真时长相符的 rec3 末态诊断（log step 294）给出金色件 y 上界 0.2726 m，远越 0.0857 m 分隔面，支持末态为假，而非由总分为零反推。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 20–30 s，机器人左二指夹爪夹住金色宽槽件并将其略向左移，随后张开退出；视频后段金色件由与右爪所持银色件的套接接触被动带走，未被左爪重新抓回。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target_horizontal_inset",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "target_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 75.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 75 s，机器人右夹爪正在抬起并搬运银色件，金色件仍套在银色件环部并被一起带起；视频 84–108.467 s 仍未脱开，且金色件随之移向目标区。失败发生在银色件分离入区过程中，该步始终没有补做完成。"
+  },
+  "unmet": {
+   "silver_in_goal": false,
+   "gold_on_system_side": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "被挡/碰撞",
+    "t": 75.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频 75–86 s，右夹爪所持银色件的环部仍与金色件圆端相互套住，搬运没有使银色件独立移走，反而抬起、转动并拖走金色件；视频 108.467 s 两件间的挂接仍在。这里是可见的件间接触阻止脱开，并带走另一任务件。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "推理/规划",
+    "log_step": 193,
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent 在 log step 193 明确依据自写检查声称“抓取后两件已相距 10 mm，抓取本身就拆开拼图”，据此把流程改为抓取后直接搬运，省去解锁曲线。该步展示的 clear_meas 代码仅用测得位姿更新金色件包围盒，实际参与 fcl.distance 的 Gobjs 却仍以单位变换初始化，未更新到金色件世界位姿，因而其返回的 10 mm 上限不能证明脱开；log step 292–294 的交付记录继续执行直接搬运。视频 20–30 s 没有完成脱开，视频 75–108.467 s 金色件仍随银色件运动，说明错误的阶段判断使必要的分离过程被跳过。"
+   }
+  ],
+  "gap_zh": "没有状态日志；金色件跨越不可见分隔面的准确时刻无法从现有画面逐毫米确定，视频 76 s 是根据视频 72 s 仍在原位、视频 74–80 s 离开原位并被带向目标框的多视图运动取的近似锚点，误差约 ±2 s。原生速度、完整网格内缩及共同确认时长未由评分器逐项报告，native_metrics 均保留 null。log step 313 把保留文件称作 rec5，但 log step 311 的文件时间仍对应较早录制；本次 4339 个评分动作及仿真 108.475 s 与 log step 294 的 rec3 完全相符，因此节点只按本次视频判断，其他录制结果不计入。"
+ },
+ "summary_zh": "初始金色件留系统侧成立，进度为 1/2；机器人双爪抓取后未将两件脱开，右爪搬运银色件时又把金色件带向目标，末态为 0/2。断点是银色件入区的分离搬运过程：agent 根据有位姿错误的自写距离检查跳过解锁流程，视频 75–108.467 s 始终能看到两件挂接。没有状态日志，金色件丢失系统侧条件的时刻仅能估在视频 76 s 附近（不确定约 ±2 s），毫米级边界和速度不能精确复核。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "总体（0/2、断点 silver_in_goal、推理/规划 根因）与视频一致，但 gold_on_system_side 的丢失时刻标成视频 84 s 偏晚：视频 72 s 金色件仍在原位，76–78 s 已被抬起并明显向目标框方向移动，80 s 已贴近目标框边。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "0/2、gold 丢失锚点约 76 s、断点 silver_in_goal 与视频一致；金色件始终随银色件挂接未脱开，末态未达成。根因依据日志的推断已标为推断，未发现须改之处。",
+    "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": "2026-10-31T00:16:01Z",
+  "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/roboarc-cast-keyhole-04-i01-privileged--cfhdrvy/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-04-i01-privileged--cfhdrvy/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-04-i01-privileged--cfhdrvy/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-04-i01-privileged--cfhdrvy/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "43a3319d99642108405a4cb1c4148c262f47ff9ceb63e1d6aa61e5966c398fb2",
+   "video_sha256": "8b4b9bd9d63f502a60e0eaa3bb439c7ccce61fc986b41cf116c24cba83c68e13",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_04/trajectories/roboarc-cast-keyhole-04-i01-privileged--cnbsdxk.json

@@ -0,0 +1,262 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-04-i01-privileged--cnbsdxk",
+ "taskKey": "roboarc/cast_keyhole_04",
+ "dag_sha256": "70720774d01cbeadb5de0aa235fa5558388d350a894bac65e2e2f759e5f7b4c7",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 0,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 576,
+   "submitted_actions": 576,
+   "sim_steps": 576,
+   "physics_steps": 5760,
+   "sim_time_s": 14.400000000000137,
+   "final_reward": 0.0,
+   "replay_wall_s": 85.4,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 1,
+   "success_in_hold_only": 0,
+   "agent_sessions": 33,
+   "agent_resets": 53,
+   "agent_sim_steps": 22323,
+   "agent_records": 14,
+   "agent_replays": 16,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.241,
+   "peak_memory_gb": 1.017,
+   "peak_processes": 12,
+   "peak_disk_gb": 0.993,
+   "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/roboarc-cast-keyhole-04-i01-privileged--cnbsdxk/replay.mp4",
+  "duration_s": 14.467,
+  "fps": 15.0,
+  "frames": 217,
+  "sha256": "693709c80ef003bea8270aca50f2dee943cb57e9e9f2eb41334b5e672373579a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 0,
+  "nodes": {
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 银色细长件与金色宽槽件相扣，两件均在目标方框外；视频 6.4 s 两爪分别夹住两件，原桌面位置随后腾空。"
+     },
+     {
+      "t": 8.2,
+      "log_step": 248,
+      "text_zh": "观察：视频 7.2–8.2 s 机器人双臂将两件一同抬起并移向绿色方框，两件保持相扣；视频 10.0–14.4 s 两件仍由各自夹爪持握，没有银色件脱开后单独进入目标体积的画面。log step 248 的已提交生成脚本对左右末端使用同一个 carry(delta)，未执行分离运动。"
+     },
+     {
+      "t": 14.4,
+      "log_step": 251,
+      "text_zh": "推断：视频 13.4–14.4 s 的末态仍是两件一起悬持，金色件随银色件跨到目标侧，故不满足本节点的脱开及共同确认条件。log step 251 对最终提交文件回放的自编几何检查报告银色件 all_inside=True、金色件 on_system_side=False，与画面一致；评分器 strict_goal_reached=0，仅作为旁证，不由零分单独判定节点未达。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频约 2.5–3.0 s 左夹爪合拢夹住金色件，视频约 6.0–6.8 s 右夹爪合拢夹住银色件；视频 7.2–14.4 s 两爪各握一件，将仍相扣的两件同步抬起、搬向目标方框并悬持。"
+     }
+    ]
+   },
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      7.7
+     ]
+    ],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 金色宽槽件在绿色方框近边之外的初始拼图侧；视频 0.0–6.8 s 金色件基本留在原位置，左夹爪下降并夹住它。"
+     },
+     {
+      "t": 7.7,
+      "log_step": 248,
+      "text_zh": "推断：视频 7.6–8.2 s 机器人将金色件和银色件同步朝方框搬运，金色件离开原侧并跨向目标侧；将条件丢失粗定于视频 7.7 s。分隔平面不可见，首次毫米级越界不能精确定位；本时间取可见整体搬运的起段，约有半秒不确定性。"
+     },
+     {
+      "t": 14.4,
+      "log_step": 251,
+      "text_zh": "观察：视频 8.2–14.4 s 金色件一直随银色件悬停在目标侧，没有回到初始拼图侧。log step 251 对最终提交文件的回放检查打印金色件 y 范围为 [0.2083,0.2748] m，分隔平面为 y≤0.085759 m，报告 on_system_side=False；这些是 agent 检查脚本的输出，与画面的大幅越界相符。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约 2.5–3.0 s 左侧二指夹爪夹住金色宽槽件，随后一直持握到视频 14.4 s；视频约 7.7 s 起左臂将金色件随银色件一起朝目标区搬运。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target_horizontal_inset",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "target_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "gold_on_system_side",
+   "t": 7.7,
+   "log_step": 248,
+   "where": "操作中",
+   "evidence_zh": "推断：视频约 7.7–8.2 s，机器人双臂各持一件同步向目标区搬运，原本成立的金色件留原侧关系在这一搬运中丢失；视频 8.2–14.4 s 始终没有恢复。选择这一首次丢失且未修复的节点为断点；银色件也始终没有脱开，不能合格入区。log step 248 明确同时搬运两件，log step 251 对最终提交文件的回放检查也报告金色件越过分隔平面。"
+  },
+  "unmet": {
+   "silver_in_goal": false,
+   "gold_on_system_side": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 7.7,
+    "log_step": 248,
+    "evidence_zh": "观察：视频 7.6–8.2 s 左夹爪把金色件随银色件一起移向目标方框，视频 13.4–14.4 s 金色件仍停在目标侧，两件仍相扣；要求留在原侧的金色件没有停在所需位置。两爪持续持物，画面没有脱手或跌落。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "推理/规划",
+    "log_step": 248,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 248 的最终生成脚本只按银色件包围盒中心误差计算位移，carry(delta) 将完全相同的平移同时施加给左右末端，继而保持 160 个控制步；脚本说明明确选择携带整对互锁件并让金色件随之移动。这种跨阶段方案没有形成分离运动，却主动破坏金色件留原侧条件，正对应视频 7.6–14.4 s 的同步搬运。log step 251 已检查到金色件 on_system_side=False，log step 264 仍将该文件称为最佳可达结果，log step 282 最终提交文件两次回放仍为失败；这支持规划取舍错误，而非把会话中其他分离实验的力或逆解问题当作本段现象。"
+   }
+  ],
+  "gap_zh": "没有 inputs/states.jsonl；分隔平面不在画面中，金色件第一次毫米级越界的精确时刻不明，视频约 7.7 s 是整体搬运起段的粗估。最终两件大幅同向移动及持续相扣足以判两节点为 false，不需以毫米边界或速度猜测是否达标。评分器未报告 DAG 的四项原生指标值，均记 null；log step 251 的数值属于 agent 自编回放检查，不能当作独立状态日志。agent 在 log step 264、283 声称实例无解，但搜索未找到路径不能排除全部合法方案，现有证据不支持任务设计问题；成功参考仅作相机可见性与脱开外观比较。"
+ },
+ "summary_zh": "机器人双爪分别抓住金色件和银色件，将仍相扣的两件同步抬起并搬向目标区，银色件合格入区始终未达。金色件留原侧从视频 0.0 s 成立，在视频约 7.7 s 的搬运中丢失且未恢复，最终为 0/2；提交脚本对两臂使用相同位移，根因是未实现分离却搬运整对零件的规划。无状态日志，首次越过隐形分隔平面的时刻只能粗估，不能据 agent 的无解声明认定任务设计问题。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "帧显示双爪各持一件、仍相扣地同步搬向方框，金色件越侧、银色件未脱开，节点值、断点和根因与证据一致；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": "2026-10-31T00:13:07Z",
+  "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/roboarc-cast-keyhole-04-i01-privileged--cnbsdxk/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-04-i01-privileged--cnbsdxk/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "6ea15d02d5258d649d00d2ab095f2b5a59d89b413e30436cd71c3f333b89ccca",
+   "video_sha256": "693709c80ef003bea8270aca50f2dee943cb57e9e9f2eb41334b5e672373579a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/roboarc/cast_keyhole_04/trajectories/roboarc-cast-keyhole-04-i01-privileged--kcwn9tq.json

@@ -0,0 +1,286 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "roboarc-cast-keyhole-04-i01-privileged--kcwn9tq",
+ "taskKey": "roboarc/cast_keyhole_04",
+ "dag_sha256": "70720774d01cbeadb5de0aa235fa5558388d350a894bac65e2e2f759e5f7b4c7",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "success_pending_review": 0,
+   "review_events": 0,
+   "review_max_depth_mm": 0.0,
+   "forcing_events": 3,
+   "all_success": 0,
+   "deterministic": 1,
+   "gpu_backend": 1,
+   "infra_error": 0,
+   "simulator_instability": 0,
+   "instance_ok": 1,
+   "instance_exact": 1,
+   "initial_state_generated": 0,
+   "instance_realized": 0,
+   "assets_exact": 1,
+   "runtime_binding_ok": 1,
+   "n_actions": 7461,
+   "submitted_actions": 7461,
+   "sim_steps": 7461,
+   "physics_steps": 74610,
+   "sim_time_s": 186.52500000002527,
+   "final_reward": 0.0,
+   "replay_wall_s": 662.4,
+   "replays": 2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "replay_hold_steps": 0,
+   "replay_tolerance_needed": 0,
+   "strict_goal_reached": 0,
+   "strict_goal_row": 0,
+   "strict_replay_required": 0,
+   "agent_sessions": 1,
+   "agent_resets": 19,
+   "agent_sim_steps": 0,
+   "agent_records": 2,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.825,
+   "peak_memory_gb": 1.222,
+   "peak_processes": 8,
+   "peak_disk_gb": 0.862,
+   "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/roboarc-cast-keyhole-04-i01-privileged--kcwn9tq/replay.mp4",
+  "duration_s": 186.6,
+  "fps": 15.0,
+  "frames": 2799,
+  "sha256": "3549563245e909a0c7c2fb97a2e4a9e1a1e15e3e80853c48f9968ddc8151d8b1"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 2,
+  "base": 1,
+  "peak": 1,
+  "final": 0,
+  "nodes": {
+   "silver_in_goal": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": null,
+      "text_zh": "观察：视频 0 s 银色细长件穿在金色宽槽件内，位于绿色目标方框近侧，尚未入区。"
+     },
+     {
+      "t": 30,
+      "log_step": 359,
+      "text_zh": "观察：视频 24–42 s 两爪分别接触两件，机器人反复转动右腕，银色件仍穿在金色槽中，未独立退出；log step 359 的服务读数仍为双手分别持有金、银件，相对旋转目标误差约78.35°。"
+     },
+     {
+      "t": 105.7,
+      "log_step": 429,
+      "text_zh": "推断：log step 328 是最后一次 reset 并开始录制，log step 533 保存7461行；40 Hz下7461/40=186.525 s，与本视频长度相符。log step 429 的 sim_steps=4228 对应视频 105.7 s，已在该帧换手释放后的银件悬立、金件下挂构型核对；评分器 deterministic=1。因此这里用该次录制中的服务读数辅助画面，而不使用保存后的会话动作。"
+     },
+     {
+      "t": 105.7,
+      "log_step": 429,
+      "text_zh": "观察：视频 105.7 s 银色件虽移到方框附近，金色件仍挂在其下端，二者没有脱开；log step 429 的银件包围盒为x=[-0.24801,-0.17932] m、y=[0.1454,0.29042] m，金件最大y=0.2166 m，服务 success=False。这一构型不能以银件投影在框内判为本节点完成。"
+     },
+     {
+      "t": 186.5,
+      "log_step": 532,
+      "text_zh": "观察：视频 185.6–186.5 s 两件仍相互套接并留在绿色方框近侧，银色件的右端明显在框外，末秒没有独立移入。log step 532（保存前最后一次动作）的银件包围盒y=[-0.05196,0.02008] m；目标区在log step 12的近边y约0.11476 m，故末态明显未入区。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频约 12–42 s 左夹爪夹住金色宽槽件、右夹爪夹住银色细长件，两臂拉开并尝试相对旋转；视频 48–100 s 又有双爪各持一件的配合动作。此后反复松爪、换手及用指尖推拨，视频末段右爪在银件端部反复闭合，但未再次形成能搬走银件的稳定夹持。"
+     }
+    ]
+   },
+   "gold_on_system_side": {
+    "method": "visual",
+    "intervals": [
+     [
+      0,
+      44.5
+     ],
+     [
+      92.5,
+      102.0
+     ],
+     [
+      169.5,
+      175.5
+     ]
+    ],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 9,
+      "text_zh": "推断：视频 0 s 金色件位于拼图原侧；log step 9 的冻结初态金件最大y=0.08076 m，结合任务所述向目标侧外移5 mm，固定分隔平面约为y=0.08576 m。视频 0–44 s 的位置、后续原侧拉动及该次录制log step 333、343、359的金件最大y约0.08189、0.03169、0.05810 m相符，初始节点成立。"
+     },
+     {
+      "t": 45,
+      "log_step": 374,
+      "text_zh": "推断：视频约 44.5–46 s 右臂抬起并转动银色件时，金色件随之向目标侧翻起，节点首次失去；log step 374 金件最大y=0.20916 m，明确超过固定平面。视频 48–90 s 可见两件仍耦合；该录制log step 385–405的多次读数中金件最大y仍在0.09404–0.14186 m等越界值，未把这些夹持停顿误判为恢复。"
+     },
+     {
+      "t": 94,
+      "log_step": 412,
+      "text_zh": "推断：视频约 92.5 s 两臂对拉后的微小位移使金件暂时返回原侧；视频 93–101.5 s 细查中该构型延续，log step 412、417、420的金件最大y分别约0.06548、0.06677、0.07094 m，均在平面内。视频约 102 s 换手时银件再带动金件向目标侧摆动；视频 105.7 s 与log step 429最大y=0.2166 m相符，恢复区间结束。隐形平面的交叉时刻为画面与这些读数结合的估计，不能精确到单帧。"
+     },
+     {
+      "t": 170,
+      "log_step": 512,
+      "text_zh": "推断：视频约 169.5 s 左臂将金件及相连银件一起转回平躺方向，金件再次返回原侧；log step 512最大y=0.06514 m，log step 519松爪后最大y=0.06037 m。视频 170–175 s 可见两件留在方框近侧；约视频 175.5 s 左臂移向银件时将整组推转，金件再次越界，对应log step 521最大y=0.15714 m。"
+     },
+     {
+      "t": 186.5,
+      "log_step": 532,
+      "text_zh": "推断：视频 176–186.5 s 金件保持最后的斜放构型，没有再返回先前平躺的原侧位置；保存前log step 532的金件最大y=0.14626 m，超过约0.08576 m的固定平面约60 mm，故末态为false，不能因主视图中它仍在目标方框近旁就计为留原侧。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：左夹爪先夹住金色件端部，配合右臂操作银件，之后多次重新夹持金件并向原侧拉回、旋转；视频约 112–146 s 还用松开或闭合的指尖接近、推拨金件。视频 169–173 s 左爪持金件转回原侧后松开；末段重新接近金件，未稳固夹住。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "target_horizontal_inset",
+   "unit": "m",
+   "threshold": 0.001,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "target_linear_speed",
+   "unit": "m/s",
+   "threshold": 0.001,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "target_angular_speed",
+   "unit": "rad/s",
+   "threshold": 0.02,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "joint_goal_confirmation",
+   "unit": "s",
+   "threshold": 0.5,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "silver_in_goal",
+   "t": 30,
+   "log_step": 359,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 24–42 s 机器人双爪分别夹持金件和银件，多次旋转、拉动后银件仍卡在金色槽内；log step 353–361反复给右臂相对旋转及拉出目标，实际相对姿态停止改善且接触压力增大。视频 44.5–46 s 又把金件一同带起，直到视频 186.5 s 两件仍未脱开。银色件分离入区这一步始终未修复，断点取其持续卡滞的操作阶段。"
+  },
+  "unmet": {
+   "silver_in_goal": false,
+   "gold_on_system_side": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "接触力异常",
+    "t": 30,
+    "log_step": 359,
+    "evidence_zh": "观察：视频 24–42 s 右夹爪对银件反复施加转动，银件在金色槽口内仅有小幅变化，未退出；log step 355、359、361的服务接触监测峰值压力依次为2218.4、2703.7、3581.1 N，而相对旋转误差停在约78–80°。这是已经接触并受阻的操作，不是从未抓到。"
+   },
+   {
+    "cat": "接触",
+    "item": "带动本体",
+    "t": 45.5,
+    "log_step": 374,
+    "evidence_zh": "观察：视频 44.5–46 s 机器人右臂持银色件抬转时，原在桌面的金色宽槽件随之翻起、向目标侧移动，未出现银件独立退出；视频 48 s 后金件仍与银件相连。log step 369到374金件中心y由0.01936 m变为0.14455 m、最高点升至1.06994 m，且左手已无辅助持物，说明操作银件带动了应留原侧的金件。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "没有状态日志；金色件的固定分隔平面不可见，恢复与越界区间由多视角画面和已提交录制的离散服务包围盒读数推断，边界约有1–2 s不确定性，未声称毫米级逐帧核实。日志显示脚本以固定相对90°目标反复旋转、拉动，但没有经过验证的完整脱锁路径，也缺少连续关节目标与实测轨迹，无法可靠区分槽口相对路径规划不当、抓取布局或逆解执行问题，故根因层留空。四项原生指标均未由评分器报告，value��null。"
+ },
+ "summary_zh": "金色件留原侧从视频 0 s 成立，操作中两次暂时恢复，末态仍越界；银色件始终没有完成分离入区，最终为0/2。机器人双爪抓住两件后，银件旋转、拉出受阻，随后又带动金件一起翻起，视频 186.5 s 两件仍相互套接。无连续状态日志，隐形分隔平面的区间边界为估计；现有证据不足以可靠区分规划和控制根因，根因层留空。",
+ "conf": 6,
+ "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（silver_in_goal 操作中）、两个现象与画面一致，末态两件仍套接且未入区；根因层留空并有 gap_zh，可接受。",
+    "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": "2026-10-31T00:06:01Z",
+  "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/roboarc-cast-keyhole-04-i01-privileged--kcwn9tq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/roboarc-cast-keyhole-04-i01-privileged--kcwn9tq/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "6662b026bb1c67a79d7ea3e161faed07d0ac6a2508ea149c3e568905a2dc345d",
+   "video_sha256": "3549563245e909a0c7c2fb97a2e4a9e1a1e15e3e80853c48f9968ddc8151d8b1",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}