Add 103 trajectories (phase 2, remaining runs; batch S07)

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-privileged--hbviuqx.json

@@ -0,0 +1,177 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-privileged--hbviuqx",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_check_success": 1,
+   "deterministic": 1,
+   "inserted": 1,
+   "tip_to_axis_m": 0.00112,
+   "axial_t": 0.549,
+   "socket_z": 0.0228,
+   "peg_z": 0.021,
+   "time_s": 56.2,
+   "n_actions": 2810,
+   "video_rendered": 1,
+   "grader_error": 0,
+   "agent_sessions": 68,
+   "agent_resets": 147,
+   "agent_cost_self_reported": 1,
+   "agent_traces": 90,
+   "agent_trace_videos": 6,
+   "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/mujoco_playground_aloha_single_peg_insertion-privileged--hbviuqx/replay.mp4",
+  "duration_s": 14.12,
+  "fps": 25.0,
+  "frames": 353,
+  "sha256": "2f9df38d954e23e25bbe68e5266d4315db61cc41822679bd9a01a15075f2e587"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [
+     [
+      12.2,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 的两个全景视图显示红色插销和蓝色套筒分开放在桌面，插入关系初始不成立。"
+     },
+     {
+      "t": 6.6,
+      "log_step": null,
+      "text_zh": "观察：视频 6.40–6.80 s，机器人右夹爪的两指夹住红色插销并将其带离原桌面位置，随后移向蓝色套筒；左夹爪没有拿起套筒，蓝管仍在原桌面位置。"
+     },
+     {
+      "t": 12.2,
+      "log_step": null,
+      "text_zh": "推断：视频 11.68–12.40 s 的右腕视图显示右夹爪继续将红色插销送入蓝色套筒，红条插入端被蓝管遮住，外露长度进一步缩短；将视频约 12.2 s 记为达到验收深度的近似起点，时间不确定度约 0.5 s。无逐帧状态日志，图像不能独立精确测量 5 毫米轴距和 12 毫米深度阈值；末态达标由评分器原生读数确认。"
+     },
+     {
+      "t": 14.08,
+      "log_step": 146,
+      "text_zh": "观察：视频 13.12–14.08 s 逐帧抽查中，右夹爪仍握住插销，红条端部保持在蓝管内，外露长度继续缩短，未见退出；左夹爪仍未持套筒。log step 146 对最终交付的 2810 行动作重新生成并独立重放，均返回 inserted=True、tip_to_axis_m=0.00112、axial_t=0.549；inputs/run.json 的 reward 同样给出 inserted=1、success=1、deterministic=1，末态满足 DAG 的几何判据，且不要求松手。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 6.40–6.80 s，机器人右侧二指夹爪夹住并搬运红色插销；视频 8.00–14.08 s 持续握持插销并向留在桌面上的蓝色套筒送入。左夹爪未抓握套筒，本次实际为单夹爪抓握插销完成插入。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00112
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 0.549
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 0.549
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人右夹爪抓住并搬运红色插销，向桌面上的蓝色套筒逐步插入，视频约 12.2 s 达标并保持至末帧，最终完成 1/1 节点；评分器确认末态轴距为 0.00112 米、轴向参数为 0.549。任务设计问题：inputs/run.json 的 task_text 要求双臂分别抓住并抬起两件物体，但视频 0.00–14.08 s 中蓝色套筒始终留在桌面、左夹爪未拿取，评分器仍判成功，故按实际几何验收标注。没有状态日志，毫米级首次达标时刻只能结合腕部画面近似判断，末态判定有评分器和 log step 146 的最终交付重放支持。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点 peg_inserted 标注（约12.2 s 起、末态成立、单夹爪抓握）与帧和评��器读数一致；腕部视图中红条自约11 s 起进入蓝管并逐步缩短，12.2 s 的近似起点在不确定度内，任务设计问题说明合理。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:21:47Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-privileged--hbviuqx/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--hbviuqx/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "ca5e3fa264a3965314200229040f36c80e7b9b005ec9fe24c8022339547f5f54",
+   "video_sha256": "2f9df38d954e23e25bbe68e5266d4315db61cc41822679bd9a01a15075f2e587",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-privileged--kbikpjm.json

@@ -0,0 +1,171 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-privileged--kbikpjm",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_check_success": 1,
+   "deterministic": 1,
+   "inserted": 1,
+   "tip_to_axis_m": 0.00258,
+   "axial_t": 0.4887,
+   "socket_z": 0.1872,
+   "peg_z": 0.1852,
+   "time_s": 28.8,
+   "n_actions": 1440,
+   "video_rendered": 1,
+   "grader_error": 0,
+   "agent_sessions": 7,
+   "agent_resets": 13,
+   "agent_cost_self_reported": 1,
+   "agent_traces": 4,
+   "agent_trace_videos": 4
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/mujoco_playground_aloha_single_peg_insertion-privileged--kbikpjm/replay.mp4",
+  "duration_s": 14.48,
+  "fps": 25.0,
+  "frames": 362,
+  "sha256": "49df866ab4e92644044e29c1e5c1e29ff0f0a0fae149df83dc6d938f60e908aa"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [
+     [
+      13.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s，蓝色套筒与红色插销分开放在桌面上，机器人两只夹爪尚未夹持物体，插入关系不成立。"
+     },
+     {
+      "t": 7.6,
+      "log_step": null,
+      "text_zh": "观察：视频 4.8–6.4 s，两侧二指夹爪分别在蓝色套筒和红色插销两侧合拢；视频 6.8–8.4 s，两物体随各自夹爪上升并离开原来的桌面位置，正面视图可见两物体悬空。"
+     },
+     {
+      "t": 11.6,
+      "log_step": null,
+      "text_zh": "观察：视频 9.0–11.6 s，机器人左夹爪持套筒，右夹爪持插销并横向移至套筒开口方向，随后沿轴向推进；仅对准的阶段不计完成。"
+     },
+     {
+      "t": 13.0,
+      "log_step": null,
+      "text_zh": "推断：结合视频 12.4、12.6、12.8、13.0、13.2、13.4 s 的连续腕部视图，右夹爪所持红色插销端部逐渐进入蓝色套筒内，端部被管壁遮住；约视频 13.0 s 达到所需插入深度，首次达标时刻有约 ±0.5 s 的视觉不确定性。无逐帧状态日志，不能从像素单独精确测量轴距和深度阈值；末态达标由本次评分器读数支持。"
+     },
+     {
+      "t": 14.44,
+      "log_step": 22,
+      "text_zh": "观察：视频 13.4–14.44 s，两夹爪持续持物，插销端部仍在蓝色套筒内，未见退出；inputs/run.json 的 reward 给出 success=1、inserted=1、deterministic=1，端点轴距 tip_to_axis_m=0.00258，小于 0.005 m，axial_t=0.4887，处于 [-0.875,0.75]。log step 19–22 记录最终动作文件保存并从初态重放，log step 22 的重放为 1440 个控制步，读数与评分器一致；不以早期 trial 的失败结果标记本次视频。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 4.8–8.4 s，机器人左侧二指夹爪夹住蓝色套筒、右侧二指夹爪夹住红色插销并各自抬起；视频 9.0–14.44 s，左夹爪持续持套筒，右夹爪持插销完成对准和推进，两臂协同插入，末态均未松手。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00258
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 0.4887
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 0.4887
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人双臂分别抓握套筒和插销并抬离桌面，约视频 13.0 s 完成插入，保持至视频 14.44 s，唯一目标节点末态为真。评分器确认成功且重放确定，末态端点轴距为 2.58 毫米、轴向参数为 0.4887，均满足验收。无逐帧状态日志，毫米级首次达标时刻只能由连续画面估计，约有半秒误差，末态几何判断由评分器读数补足。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "唯一目标节点 peg_inserted 的区间 [13.0, null]、末态为真、双臂协作抓握均与画面和评分器读数一致；成功运行无断点，无异常。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:14:05Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-privileged--kbikpjm/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--kbikpjm/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "d4d53290611e757ec0717a683befa4b7dc6767b04f5f232c280d04d094d5f5d7",
+   "video_sha256": "49df866ab4e92644044e29c1e5c1e29ff0f0a0fae149df83dc6d938f60e908aa",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-privileged--kehyyib.json

@@ -0,0 +1,187 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-privileged--kehyyib",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_check_success": 1,
+   "deterministic": 1,
+   "inserted": 1,
+   "tip_to_axis_m": 0.0006,
+   "axial_t": 0.1096,
+   "socket_z": 0.1746,
+   "peg_z": 0.1778,
+   "time_s": 13.2,
+   "n_actions": 660,
+   "video_rendered": 1,
+   "grader_error": 0,
+   "agent_sessions": 41,
+   "agent_resets": 55,
+   "agent_cost_self_reported": 1,
+   "agent_traces": 15,
+   "agent_trace_videos": 6,
+   "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/mujoco_playground_aloha_single_peg_insertion-privileged--kehyyib/replay.mp4",
+  "duration_s": 13.28,
+  "fps": 25.0,
+  "frames": 332,
+  "sha256": "fde313c7dd9d095ff297dc35865e2003f01137eaae7324d19cf92a0bae6994ba"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.4,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的外部视图中，蓝色套筒与红色插销分开放在桌面，插销端点未进入套筒。"
+     },
+     {
+      "t": 4.8,
+      "log_step": null,
+      "text_zh": "观察：视频 3.2–4.0 s，机器人两只二指夹爪分别从两侧夹住蓝色套筒和红色插销；视频 4.4–5.6 s 的正面视图中，两物体随各自夹爪升起，原桌面位置空出，物体与桌面之间出现间隙。"
+     },
+     {
+      "t": 10.4,
+      "log_step": null,
+      "text_zh": "观察：视频 9.0–9.6 s，机器人持红色插销的夹爪将其端点对准另一夹爪所持蓝管的开口，端点仍在开口外；视频 9.8–11.2 s 的左腕视图中，红色端点进入蓝管，露在管外的红条长度持续缩短。"
+     },
+     {
+      "t": 10.4,
+      "log_step": null,
+      "text_zh": "推断：结合视频 9.8–11.2 s 的连续深入过程及最终评分器 inserted=1，将达到规定插入深度的起点估在视频 10.4 s，误差约 0.4 s；无逐帧状态日志，图像不能精确确定轴距小于 5 毫米且轴向参数首次不大于 0.75 的瞬间。"
+     },
+     {
+      "t": 13.24,
+      "log_step": 58,
+      "text_zh": "观察：视频 12.2、12.4、12.6、12.8、13.0、13.2 和 13.24 s，外部视图中机器人继续持住两物体，未见退出或掉落，腕部视图的插入口被另一夹爪遮挡；inputs/run.json 的 reward 给出 success=1、inserted=1、deterministic=1、tip_to_axis_m=0.0006、axial_t=0.1096。log step 57 重新生成交付文件并重放成功，log step 58 再次重放所得终态几何读数与评分器一致。"
+     },
+     {
+      "t": 13.24,
+      "log_step": null,
+      "text_zh": "推断：评分器的末态轴距 0.0006 米小于 0.005 米，轴向参数 0.1096 位于 [-0.875,0.75] 内；结合视频持续插入并保持的过程，节点末态为真。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 3.2–5.6 s，机器人左臂二指夹爪夹持蓝色套筒外壁，右臂二指夹爪夹持红色插销，两物体均随夹爪离桌；视频 9.8–11.2 s，左爪持稳套筒，右爪持插销向开口推进完成插入，之后双爪继续持物至视频 13.24 s。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.0006
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 0.1096
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 0.1096
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人双臂分别抓握并抬起套筒和插销，在空中对准后于约视频 10.4 s 达到插入状态，唯一节点末态完成（1/1），无失败断点。评分器判定成功且重放确定，末态轴距为 0.6 毫米、轴向参数为 0.1096，均达标。无逐帧状态日志，首次跨越毫米级验收阈值的时刻只能近似判断，视频末段插入口受夹爪遮挡，末态几何判定结合评分器字段。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点 peg_inserted 的末态为真、约 10.4 s 到达，与视频帧（9.0 s 端点在开口外，10.4 s 红条已进入蓝管）及评分器 inserted=1、tip_to_axis 0.0006 一致；成功运行无断点，contact_actual（双臂协作+抓握）与帧相符。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:16:50Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--kehyyib/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--kehyyib/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "68011c507e6a35f79e19a33da07de3420d6cc46f3ee65401550648abdeb8e8b1",
+   "video_sha256": "fde313c7dd9d095ff297dc35865e2003f01137eaae7324d19cf92a0bae6994ba",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-privileged--qknakkt.json

@@ -0,0 +1,177 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-privileged--qknakkt",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_check_success": 1,
+   "deterministic": 1,
+   "inserted": 1,
+   "tip_to_axis_m": 0.00275,
+   "axial_t": 0.6549,
+   "socket_z": 0.0228,
+   "peg_z": 0.0207,
+   "time_s": 11.94,
+   "n_actions": 597,
+   "video_rendered": 1,
+   "grader_error": 0,
+   "agent_sessions": 40,
+   "agent_resets": 48,
+   "agent_cost_self_reported": 1,
+   "agent_traces": 15,
+   "agent_trace_videos": 6,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.084,
+   "peak_processes": 14,
+   "peak_disk_gb": 0.24,
+   "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/mujoco_playground_aloha_single_peg_insertion-privileged--qknakkt/replay.mp4",
+  "duration_s": 12.0,
+  "fps": 25.0,
+  "frames": 300,
+  "sha256": "5d9df3acf1fde936194ce208c0f3120a2dca56ad15ed8efa858d873bbfee9add"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 四视图显示红色插销与蓝色套筒分离，初始插入关系不成立。"
+     },
+     {
+      "t": 4.8,
+      "log_step": null,
+      "text_zh": "观察：视频 3.2–4.0 s 右侧机械臂的二指夹爪在红色插销两侧合拢；视频 4.8–5.2 s 红销随右夹爪离开原来桌面位置并抬起，随后向蓝色套筒开口移动；蓝色套筒仍在桌面上，左夹爪未接触它。"
+     },
+     {
+      "t": 10.0,
+      "log_step": null,
+      "text_zh": "推断：逐帧对比视频 7.0–11.8 s（每隔视频 0.2 s 取帧），右腕视图中红销端部从开口进入蓝管遮挡区域，露出长度逐渐缩短；结合末态通过插入验收的评分器读数，将满足深度要求的起点估为视频约 10.0 s，误差约 0.5 s。视频 11.0–11.96 s 保持插入，未见退出；无逐帧几何状态日志，首次越过毫米级阈值的时刻为视觉估计。"
+     },
+     {
+      "t": 11.96,
+      "log_step": 39,
+      "text_zh": "观察：视频 11.96 s 末帧仍见红销端部被蓝管遮住、右夹爪夹持红销，形态与参考成功视频 14.08 s 相符；参考仅用于识别可见线索。本次 run.json 的评分器 reward 为 success=1、inserted=1、deterministic=1、tip_to_axis_m=0.00275、axial_t=0.6549，满足轴距严格小于 0.005 m 且轴向参数位于 [-0.875,0.75]；log step 39 对已保存的 597 行动作独立重放得到相同末态读数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频 3.2–5.2 s 右机械臂二指夹爪合拢夹住红色插销并将其抬离桌面，视频 7.0–11.96 s 继续握持并送入蓝色套筒；套筒由桌面承托，另一夹爪未抓握套筒。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00275
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 0.6549
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 0.6549
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人用右夹爪抓起红色插销，向桌面上的蓝色套筒对准并插入，约在视频 10.0 s 达成唯一节点，末态评分器确认成功（轴距 2.75 毫米，轴向参数 0.6549），无失败断点。任务设计问题：inputs/run.json 的任务文本要求双臂分别抓持并抬起两件物体，但本次视频 0.0–11.96 s 套筒留在桌面、单臂插入仍被评分器接受，验收仅检查插入几何关系。无逐帧状态日志，首次达到精确深度阈值的时刻只能由视频估计，约有 0.5 s 误差，末态几何数值采用评分器字段。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点 peg_inserted 的值、约 10.0 s 的起点估计（已标注为推断，误差 0.5 s）、末态 final=true、contact_actual（右夹爪抓握）与评分器读数一致；成功 run 无失败断点，任务设计问题说明有 run.json 任务文本支撑。未发现需修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:19:01Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-privileged--qknakkt/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--qknakkt/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "50959e39294563f78256502012db84876512795887aa36c9a99add4fd853afa4",
+   "video_sha256": "5d9df3acf1fde936194ce208c0f3120a2dca56ad15ed8efa858d873bbfee9add",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-privileged--sd6tel3.json

@@ -0,0 +1,177 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-privileged--sd6tel3",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_check_success": 1,
+   "deterministic": 1,
+   "inserted": 1,
+   "tip_to_axis_m": 0.00122,
+   "axial_t": -0.0009,
+   "socket_z": 0.1512,
+   "peg_z": 0.1504,
+   "time_s": 10.4,
+   "n_actions": 520,
+   "video_rendered": 1,
+   "grader_error": 0,
+   "agent_sessions": 15,
+   "agent_resets": 25,
+   "agent_cost_self_reported": 1,
+   "agent_traces": 4,
+   "agent_trace_videos": 4,
+   "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/mujoco_playground_aloha_single_peg_insertion-privileged--sd6tel3/replay.mp4",
+  "duration_s": 10.48,
+  "fps": 25.0,
+  "frames": 262,
+  "sha256": "3982528519bddb769ab10758d7ee05a3fd65a1c03ad70fe4ea1971f309e2ddb4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [
+     [
+      7.6,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频起始四视图中，红色插销与蓝色套筒分开放在桌面上，插销端尚未进入套筒。"
+     },
+     {
+      "t": 3.4,
+      "log_step": null,
+      "text_zh": "观察：视频 2.2–3.4 s，机器人两侧二指夹爪分别闭合夹住蓝色套筒与红色插销；视频 3.4–4.5 s，两件随各自夹爪离开原桌面位置并抬到桌面上方，随后相向对准。"
+     },
+     {
+      "t": 7.6,
+      "log_step": null,
+      "text_zh": "推断：视频 6.8–8.2 s 的逐段帧显示机器人持红色插销沿蓝色套筒开口持续送入；腕部视图中红条插入端逐渐被蓝管遮住、外露长度缩短，视频约 7.6 s 已有明显插入深度，故以该时刻作为达标起点，精确越过 12 毫米深度阈值的时间约有 ±0.5 s 误差。后续视频至 10.44 s 未见退出或错开；无逐帧状态日志，不能逐帧直接核验毫米级轴距。"
+     },
+     {
+      "t": 10.44,
+      "log_step": 26,
+      "text_zh": "观察：视频 9.48–10.44 s（含末帧），机器人双爪持续持有两件，红色插销一端保持进入蓝色套筒内。inputs/run.json 的评分器 reward 给出 success=1、inserted=1、deterministic=1、tip_to_axis_m=0.00122、axial_t=-0.0009，满足轴距小于 0.005 米且轴向参数在 [-0.875,0.75]；log step 26 对最终保存的 520 个动作执行独立重放，报告相同读数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 2.2–4.5 s，两侧二指夹爪分别合拢夹持蓝色套筒与红色插销，并共同将两件抬离桌面；视频 6.8–10.44 s，持套筒的夹爪固定套筒，另一夹爪持插销沿开口送入并保持，两爪各自夹住一件。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00122
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": -0.0009
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": -0.0009
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人双爪分别抓握并抬起套筒和插销，完成空中对准与插入；插入节点约从视频 7.6 s 起成立，并保持至视频 10.44 s 末帧，最终完成 1/1。评分器确认成功且末态轴距为 1.22 毫米、轴向参数为 -0.0009；无逐帧状态日志，首次达到精确几何阈值的时刻仅能由视频估计，约有半秒误差。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点 peg_inserted 成功，评分器 inserted=1、末态轴距 1.22 mm；视频帧显示红条露出长度在 6.8–8.2 s 逐渐缩短，7.6 s 起点在半秒误差内可接受；接触 双臂协作+抓握 符合帧所见；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": "2026-10-30T03:16:25Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--sd6tel3/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--sd6tel3/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "83dddcd782425a832bfee7a8042057cfe9dc116d3e8bad49331e7ae63caf19f4",
+   "video_sha256": "3982528519bddb769ab10758d7ee05a3fd65a1c03ad70fe4ea1971f309e2ddb4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-privileged--tzyml4h.json

@@ -0,0 +1,182 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-privileged--tzyml4h",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_check_success": 1,
+   "deterministic": 1,
+   "inserted": 1,
+   "tip_to_axis_m": 0.00097,
+   "axial_t": 0.2118,
+   "socket_z": 0.1548,
+   "peg_z": 0.1588,
+   "time_s": 13.3,
+   "n_actions": 665,
+   "video_rendered": 1,
+   "grader_error": 0,
+   "agent_sessions": 4,
+   "agent_resets": 6,
+   "agent_cost_self_reported": 1,
+   "agent_traces": 1,
+   "agent_trace_videos": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.492,
+   "peak_processes": 8,
+   "peak_disk_gb": 0.239,
+   "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/mujoco_playground_aloha_single_peg_insertion-privileged--tzyml4h/replay.mp4",
+  "duration_s": 13.36,
+  "fps": 25.0,
+  "frames": 334,
+  "sha256": "6f57f1a6c194e1f08a250842d4f463ce68ceb163b9036726f833eb86730f8296"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 的外部视图显示蓝色套筒与红色插销分开放在桌面，尚未建立插入关系。"
+     },
+     {
+      "t": 4.8,
+      "log_step": null,
+      "text_zh": "观察：视频 3.0–3.8 s 两个二指夹爪分别在套筒和插销两侧闭合；视频 4.2–4.8 s 两件物体随各自夹爪上升，离开原桌面位置，随后由双臂持物移向彼此。"
+     },
+     {
+      "t": 10.0,
+      "log_step": null,
+      "text_zh": "推断：视频 9.56 s 红色插销端仍在蓝色套筒开口外，视频 9.68–10.16 s 左腕视图显示红端沿开口进入内腔、被蓝管遮住且外露长度逐渐缩短；据此将满足插入深度的首次时刻近似标在视频 10.0 s（约有 0.3 s 不确定性）。初次进入开口不等同于达到至少 12 毫米深度，画面不能精确测量毫米级阈值。"
+     },
+     {
+      "t": 13.32,
+      "log_step": null,
+      "text_zh": "观察：视频 12.00、12.36、12.60、12.84、13.08 和 13.32 s 的左腕视图仍可见红色插销伸入蓝管，外部视图显示两件物体保持离桌、两爪继续握持，未见退出或掉落；评分器 reward 报告 success=1、inserted=1、deterministic=1、tip_to_axis_m=0.00097、axial_t=0.2118，确认末态同时满足轴距与深度判据。"
+     },
+     {
+      "t": null,
+      "log_step": 8,
+      "text_zh": "观察：log step 7 的脚本输出记录成功重放后保存 665 行动作；log step 8 在新进程读取已保存文件，重放结果仍为 success=True、inserted=True、tip_to_axis_m=0.00097、axial_t=0.2118，与本次评分器末态字段一致。此处用于核对交付对象，未将 agent 工作时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 3.0–4.8 s，机器人左臂二指夹爪夹住蓝色套筒，右臂二指夹爪夹住红色插销，分别抬离桌面；视频 9.56–13.32 s 左臂持套筒、右臂持插销送入并保持，两爪均以对向夹持方式抓握各自物体。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00097
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 0.2118
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 0.2118
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人双臂分别抓握并抬起套筒与插销，视频约 10.0 s 完成插入，保持至末帧视频 13.32 s，最终完成 1/1 个节点。评分器确认末态轴距为 0.00097 米、轴向参数为 0.2118，交付文件独立重放成功（log step 8），无失败断点。无逐帧状态日志，腕部视图后段有夹爪遮挡，毫米级验收依赖评分器原生读数，首次达成时间为视觉近似值。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "peg_inserted 在视频约10 s 达成并保持到末帧，与评分器读数（inserted=1，tip_to_axis 0.00097，axial_t 0.2118）及帧一致；接触双臂协作+抓握正确，无失败断点。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:14:21Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-privileged--tzyml4h/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-privileged--tzyml4h/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "fd6cfde9e47e3bdaa06a61c1c31b2d182f0390584284c7d090c859d8fbe0a270",
+   "video_sha256": "6f57f1a6c194e1f08a250842d4f463ce68ceb163b9036726f833eb86730f8296",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-standard--8g8mjwh.json

@@ -0,0 +1,238 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-standard--8g8mjwh",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "control_steps": 5910,
+   "calls": 127,
+   "elapsed_s": 986.8,
+   "video_rendered": 1,
+   "infra_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "sim_error": 0,
+   "resets": 0,
+   "reset_budget": 0,
+   "rows": 880,
+   "done": 1,
+   "video_done": 1,
+   "episodes": 1,
+   "control_steps_total": 5910,
+   "live_success": 0,
+   "inserted": 0,
+   "tip_to_axis_m": 0.00029,
+   "axial_t": 2.4482,
+   "socket_z": 0.0228,
+   "peg_z": 0.0231,
+   "time_s": 118.2,
+   "video": 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/mujoco_playground_aloha_single_peg_insertion-standard--8g8mjwh/replay.mp4",
+  "duration_s": 26.32,
+  "fps": 25.0,
+  "frames": 658,
+  "sha256": "2576ab6a9fdc21f469b8fa8431920b648bfe9ed1a02384554b7a6338885c9126"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s，蓝色套筒与红色插销分开放在桌面上，尚无插入关系。"
+     },
+     {
+      "t": 4.6,
+      "log_step": null,
+      "text_zh": "观察：视频 2.5 s 双臂分别夹持并抬起蓝色套筒和红色插销；视频 4.4–5.2 s 两件物体靠拢，红条靠套筒的一端被蓝管遮住，随后腕部视图被机器人遮挡。该外观只能证明一端进入，不能辨认命名站点 peg_end2。"
+     },
+     {
+      "t": 8.4,
+      "log_step": null,
+      "text_zh": "观察：视频 8.2–8.4 s 红色插销从右夹爪下方落到桌面；视频 10.5–15.0 s 机器人重新抓取并抬起插销，视频 15.4 s 蓝色套筒也落到桌面。"
+     },
+     {
+      "t": 21.5,
+      "log_step": null,
+      "text_zh": "观察：视频 17.5–19.25 s 右夹爪持红色插销向桌上的蓝色套筒推进，视频 19.5–21.0 s 右夹爪松开后以闭合指端从外端推红条，视频 21.5 s 再夹住红条；左夹爪此时悬空。"
+     },
+     {
+      "t": 26.28,
+      "log_step": 67,
+      "text_zh": "观察：视频 24.5–26.28 s 右夹爪仍夹持红色插销，蓝管遮住其靠套筒的一段，外端仍明显伸出；成功参考视频 12.0–14.08 s 也有相似遮挡外观，因此不能凭露出长度给本次受检端点计分。评分器末态 inserted=0、tip_to_axis_m=0.00029 m、axial_t=2.4482，轴距合格但受检端点远在开口之外，超过轴向上界 0.75。"
+     },
+     {
+      "t": 26.28,
+      "log_step": 55,
+      "text_zh": "推断：log step 55 的视觉脚本以 xend-0.096 作为插入端，log step 67 承认假定朝套筒的负 x 端就是 peg_end2；该估计的轴向参数约 0.37 与评分器的 2.4482 不同，支持实际进入的是另一端。视频全过程未见机器人将插销掉头或改从另一端插入，故不认定存在满足指定端点判据的区间；末态为假由原生几何读数直接决定。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "视频约 2.5–8.2 s，左二指夹爪夹持蓝色套筒，右二指夹爪夹持红色插销，抬起并相向靠拢尝试插入；随后插销掉落。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "视频约 14.0 s 右夹爪重新夹住红色插销；蓝色套筒于视频约 15.4 s 掉落后由桌面承托，右夹爪单独搬运、推进和调整插销，末态仍夹持红条。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "视频约 19.5–21.0 s，右夹爪先松开红色插销，再用闭合的指端顶住红条露出的外端向桌上套筒方向推；此阶段不是夹住红条搬运。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00029
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 2.4482
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 2.4482
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "peg_inserted",
+   "t": 4.6,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "推断：视频约 4.6 s 双臂首次将红条一端送进蓝管，后来掉落、重新抓取并改用桌面插入，但直到视频 26.28 s 都未见掉头换端。末态评分器 axial_t=2.4482>0.75，指定端点仍在套筒开口外；首次插入方向的端点偏差没有得到纠正。"
+  },
+  "unmet": {
+   "peg_inserted": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 24.5,
+    "log_step": null,
+    "evidence_zh": "推断：视频 24.5–26.28 s 红色插销一段进入蓝管，另一端仍露在套筒外，机器人保持该姿态结束；评分器记录受检端点距轴 0.00029 m、轴向参数 2.4482，说明该端点虽对轴却没有进入套筒规定深度。结合 log step 55、67 的端点假设，实际进入的一端与受检端点不一致。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "感知",
+    "log_step": 55,
+    "source_zh": "日志 / 评分器字段",
+    "evidence_zh": "推断：log step 55 的脚本把可见红条最大 x 坐标减去 0.096 当作 tip，log step 62 得到该点约 x=-0.0742；log step 67 明确说未确认哪一端是 peg_end2，只假定负 x 端是受检端。评分器末态却给出 axial_t=2.4482，而 agent 在 log step 67 估计约 0.37；视频 24.5–26.28 s 末态仍有红条外端露出，支持受检端点识别错误，导致控制器围绕错误端点调整。"
+   },
+   {
+    "cat": "认知",
+    "item": "误以为完成",
+    "log_step": 67,
+    "source_zh": "agent 自述 / 日志",
+    "evidence_zh": "推断：agent 在 log step 67 自述插销已插入、视觉估计符合评分条件，并明确说已停止发送命令；虽承认没有成功信号及端点身份风险，仍以该估计接受末态。log step 65 的服务 time_left_s 仍为 878.8，任务预算为 1800 秒，说明停止并非预算耗尽；评分器 inserted=0 与该完成判断矛盾。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或沿程原生 inserted/轴向读数；视频 5.4–10.0 s 腕部视图受遮挡，且图像不能直接辨认 peg_end2 站点或精确核验毫米阈值。因此全过程未达成及插错端的判断结合了可见运动方向、脚本端点假设和末态评分器，瞬时是否擦过阈值不能完全排除；没有将 agent 自估的插入深度当作节点达成证据。"
+ },
+ "summary_zh": "机器人先双臂持件尝试空中插入，掉落后重新抓取插销并改在桌面推进、调整，最终插入节点为 0/1。红条虽有一端进入蓝管，评分器受检端点轴向参数仍为 2.4482，未进入允许范围；日志支持识别错端并据错误完成估计停止。缺少逐帧几何状态，端点身份及全过程未达成包含推断，末态未达成有原生读数直接支持。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "单节点 peg_inserted 末态为假由评分器 inserted=0、axial_t=2.4482 直接决定，视频帧（红条始终从外端伸出、无掉头）与之相符；接触标注、断点、位姿不准、感知（log 55/62/67 端点假设，估计 0.37 vs 评分器 2.45，且红条约 96 mm 长，几何上吻合插错端）与误以为完成均有证据支持，时钟与引用均无问题。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:14:00Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-standard--8g8mjwh/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--8g8mjwh/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "842e31ddf92072288cce4a46683a5ad87ca7d853eef45f1e302a6293f2219db2",
+   "video_sha256": "2576ab6a9fdc21f469b8fa8431920b648bfe9ed1a02384554b7a6338885c9126",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-standard--8qb9v4r.json

@@ -0,0 +1,186 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-standard--8qb9v4r",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "finalized": 1,
+   "control_steps": 550,
+   "calls": 3,
+   "elapsed_s": 1416.9,
+   "video_rendered": 1,
+   "infra_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "sim_error": 0,
+   "resets": 0,
+   "reset_budget": 0,
+   "rows": 550,
+   "done": 1,
+   "video_done": 1,
+   "episodes": 1,
+   "control_steps_total": 550,
+   "live_success": 1,
+   "inserted": 1,
+   "tip_to_axis_m": 0.00057,
+   "axial_t": 0.0745,
+   "socket_z": 0.1675,
+   "peg_z": 0.1665,
+   "time_s": 11.0,
+   "video": 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/mujoco_playground_aloha_single_peg_insertion-standard--8qb9v4r/replay.mp4",
+  "duration_s": 11.08,
+  "fps": 25.0,
+  "frames": 277,
+  "sha256": "749b4c2cb11fe0294dff94fbc4b4b05cb705a50364b7919e26b637a49458389d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [
+     [
+      9.2,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频开头红色插销与蓝色套筒分开放在桌面上，插入关系尚未成立。"
+     },
+     {
+      "t": 4.8,
+      "log_step": 44,
+      "text_zh": "观察：视频 3.2–4.0 s 两个二指夹爪分别闭合夹住蓝色套筒和红色插销；视频 4.4–5.6 s 两件物体随各自夹爪离开桌面，原放置处变空，随后机器人将二者移到桌面上方相对的位置。"
+     },
+     {
+      "t": 9.2,
+      "log_step": 47,
+      "text_zh": "推断：对照视频 7.8–8.2 s 的开口外对准与视频 8.4–9.4 s 的连续推进，腕部视图中红色插销逐渐进入蓝色套筒，露出长度缩短，至视频 9.4 s 插入端被蓝管遮住；将达到规定插入深度的起点估为视频 9.2 s，误差约 0.3 s。无逐帧状态日志，不能由像素精确确定毫米级阈值越界时刻。"
+     },
+     {
+      "t": 11.04,
+      "log_step": null,
+      "text_zh": "观察：视频 10.0、10.2、10.4、10.6、10.8、11.0 和 11.04 s，机器人双爪仍各持一件，红色插销保持在蓝色套筒内，未见退回或掉落；评分器 reward 报告 success=1、inserted=1、tip_to_axis_m=0.00057、axial_t=0.0745，末态满足距轴线小于 0.005 m 且轴向参数位于 [-0.875,0.75] 的验收。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 3.2–5.6 s 左臂二指夹爪夹住蓝色套筒、右臂二指夹爪夹住红色插销并将两者抬起；视频 8.4–10.0 s 左爪持稳套筒、右爪持销推进完成插入，双爪保持抓握至视频 11.04 s。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.00057
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 0.0745
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 0.0745
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人双爪分别抓握并抬起套筒与插销，对准后推进，约在视频 9.2 s 达成插入，保持至视频 11.04 s，最终完成 1/1 个节点。评分器确认插入成功，末态距轴线 0.57 毫米、轴向参数 0.0745，均满足验收，无失败断点。没有状态日志，首次达标时刻由视频估计，误差约 0.3 s；末态毫米级几何关系由评分器读数佐证。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "标注与视频一致：双爪各夹套筒与插销抬起、对准后约视频 9.2 s 插入，至 11.04 s 保持，评分器读数佐证 final=true；contact_actual 双臂协作+抓握正确，无断点，校验通过。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:20:20Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--8qb9v4r/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--8qb9v4r/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f832ce33190a2ec78504eba58835cc29f067bcf972da69999e5f0abe35658191",
+   "video_sha256": "749b4c2cb11fe0294dff94fbc4b4b05cb705a50364b7919e26b637a49458389d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-standard--aabtuh3.json

@@ -0,0 +1,229 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-standard--aabtuh3",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "control_steps": 3820,
+   "calls": 76,
+   "elapsed_s": 1827.8,
+   "video_rendered": 1,
+   "no_verdict": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "resets": 0,
+   "reset_budget": 0,
+   "rows": 62,
+   "done": 1,
+   "video_done": 1,
+   "episodes": 1,
+   "control_steps_total": 3820,
+   "n_actions": 3820,
+   "video": 1,
+   "video_pending": 0,
+   "inserted": 0,
+   "tip_to_axis_m": 0.03436,
+   "axial_t": -0.58,
+   "socket_z": 0.0237,
+   "peg_z": 0.0112,
+   "time_s": 76.4,
+   "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/mujoco_playground_aloha_single_peg_insertion-standard--aabtuh3/replay.mp4",
+  "duration_s": 25.52,
+  "fps": 25.0,
+  "frames": 638,
+  "sha256": "e92e064bf5d3ff1d0800efab15d6326345af90282c9a47d66de6e3a60f3df4e3"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 四视图中红色插销与蓝色套筒分离。视频 0.0–25.48 s 的已查帧中，机器人先抓取套筒，再反复夹取红销，后将两件移近；未见红销端部沿开口进入内腔的可见特征。"
+     },
+     {
+      "t": 12.48,
+      "log_step": 118,
+      "text_zh": "观察：视频 12.16–12.32 s 红销随右夹爪离桌，视频 12.40–12.80 s 右夹爪继续抬升而红销脱离夹爪落回桌面，蓝色套筒仍在另一夹爪中；log step 118 的动作保持右夹爪闭合并要求继续抬升红销。"
+     },
+     {
+      "t": 21.6,
+      "log_step": null,
+      "text_zh": "观察：视频 20.4–22.0 s 腕部视图显示红销与蓝管并排，红销端部在蓝管外侧，未从管口进入。成功参考录像视频 8.0–12.0 s 的相同腕部视图可见红销一端被蓝管遮住、露出段缩短，仅用于比较达成特征。"
+     },
+     {
+      "t": 25.48,
+      "log_step": null,
+      "text_zh": "观察：视频 25.48 s 末帧腕部视图显示红销仍沿蓝管外侧，开口处未形成插入关系；本次评分器 reward 报告 inserted=0、tip_to_axis_m=0.03436、axial_t=-0.58。末态轴向参数在允许范围，但距轴线 34.36 毫米超过严格小于 5 毫米的阈值，故节点末态为假。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 6.0–12.32 s 左侧机械臂二指夹爪夹持蓝色套筒，右侧机械臂夹爪反复闭合夹取红销；视频 12.16–12.32 s 红销短暂随右夹爪离桌，随后滑脱。两臂各自夹持零件参与同一插入步骤。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "观察：视频 17.0–22.0 s 右夹爪接近桌面推动红销，使其向蓝色套筒靠近；红销在桌面上移位，并未保持为稳定的悬空抓持。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.03436
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": -0.58
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": -0.58
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "peg_inserted",
+   "t": 12.48,
+   "log_step": 118,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 12.40–12.80 s，机器人右夹爪抬升时红销脱离并落回桌面，插入操作中的持物失败；随后机器人重新接近、推移并调整两件，视频 20.4–22.0 s 及视频 25.48 s 仍显示红销位于蓝管外侧，直到结束也没有完成插入。"
+  },
+  "unmet": {
+   "peg_inserted": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 12.48,
+    "log_step": 118,
+    "evidence_zh": "观察：视频 12.16–12.32 s 红销被右夹爪夹起，视频 12.40–12.80 s 夹爪继续上移而红销滑脱落回桌面；log step 118 的右夹爪目标仍为闭合值 0.002，没有计划松手。"
+   },
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 21.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频 20.4–22.0 s，机器人把红销移至套筒旁，但红销与蓝管并排、端部偏在管口外侧；视频 25.48 s 仍是外侧错开关系。评分器末态 tip_to_axis_m=0.03436，支持最终横向偏差超限。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 112、118 连续要求闭爪抬升红销，视频 12.16–12.80 s 却显示夹持失效、红销滑落；log step 168、171、173 转为朝套筒推动红销，视频 20.4–22.0 s 及视频 25.48 s 仍显示两件并排错开。机器人未能实现稳定持物与相对位姿控制，因此归于控制；现有证据不能再区分夹持几何、接触力或轨迹跟踪问题。",
+    "source_zh": "视频 / 日志"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或接触力数据，无法精确判定夹持滑脱的力学原因，也不能从像素测量全过程的毫米级轴距；部分后段腕部画面被机械臂遮挡。末态由视频 25.48 s 的清晰外侧关系及评分器几何读数共同确认，未将日志中的抓取动作名称当作已抓稳的证据。"
+ },
+ "summary_zh": "机器人夹持蓝色套筒并反复夹取红销，但视频 12.40–12.80 s 红销在抬升中滑脱，随后推移和对准也未形成插入，唯一节点全程未见达成、末态为 0/1。视频 25.48 s 红销仍位于蓝管外侧，评分器端点距轴线 34.36 毫米，失败表现为没保持住及位姿不准，根因归于控制。缺少逐帧状态及接触力数据，无法进一步确定控制子类或精确测量中途几何误差。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "唯一节点 peg_inserted 末态为假，与视频 25.4 s 腕部视图（红销在蓝管外侧）及评分器 inserted=0、tip_to_axis_m=0.03436 一致；断点（视频约 12.5 s 红销滑脱）、现象和根因「控制」均有证据支持，未发现需要改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:21:36Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-standard--aabtuh3/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--aabtuh3/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "a49aae85c8ab26df9844c7889fecdb2126e5052410a19c0108c81bb4f0f28334",
+   "video_sha256": "e92e064bf5d3ff1d0800efab15d6326345af90282c9a47d66de6e3a60f3df4e3",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-standard--cur5s7c.json

@@ -0,0 +1,203 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-standard--cur5s7c",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "control_steps": 5006,
+   "calls": 108,
+   "elapsed_s": 3084.2,
+   "video_rendered": 1,
+   "infra_error": 0,
+   "grader_error": 0,
+   "no_verdict": 0,
+   "sim_error": 0,
+   "rows": 5006,
+   "done": 1,
+   "video_done": 1,
+   "live_success": 0,
+   "inserted": 0,
+   "tip_to_axis_m": 0.01083,
+   "axial_t": 1.0123,
+   "socket_z": 0.0228,
+   "peg_z": 0.0137,
+   "time_s": 100.12,
+   "video": 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/mujoco_playground_aloha_single_peg_insertion-standard--cur5s7c/replay.mp4",
+  "duration_s": 25.08,
+  "fps": 25.0,
+  "frames": 627,
+  "sha256": "d2bfe376dc66f7d5551dbca06be575f5c8e045419f7d6269c4b030672a0fa9d6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 四视图中，红色插销与蓝色套筒分离，均在桌面上；初态不满足插入关系。"
+     },
+     {
+      "t": 21.5,
+      "log_step": null,
+      "text_zh": "观察：视频 0–20.75 s 的全程粗看及视频 8–25.04 s 的加密帧中，机器人先多次向红销合爪、抬升和重新抓取，随后于视频约 20.75–21.5 s 把红销移至蓝管口；此前红销与蓝管分开，之后红销仍大段外露，后续修正未出现成功参考视频 8–14.08 s 中红条深入蓝管、露出长度明显缩短的特征。"
+     },
+     {
+      "t": 25.04,
+      "log_step": 219,
+      "text_zh": "观察：视频 24.25、24.5、24.75、25.04 s 中，机器人右夹爪仍夹着红销停在蓝管口附近，红条未深入至成功参考的可见位置；inputs/run.json 的评分器末态字段 inserted=0、tip_to_axis_m=0.01083、axial_t=1.0123，分别不满足轴距严格小于 0.005 m 及轴向参数不大于 0.75 的要求。log step 219 的机器人服务末态状态为 control_steps=5006、sim_time_s=100.12；这里视频时间取自所看帧，未用 agent 工作时间换算。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：视频约 20.75–21.5 s，右臂二指夹爪夹持红色插销，将其从原位置移至蓝色套筒入口附近，随后至视频 25.04 s 保持夹持并尝试推送、修正；蓝色套筒留在桌面，另一夹爪未抓住套筒。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.01083
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 1.0123
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 1.0123
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "peg_inserted",
+   "t": 21.5,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 21.5 s，机器人右夹爪持红销靠向蓝管入口，之后反复推进和微调至视频 25.04 s，仍未形成成功参考中红条深入管内的关系；评分器末态 inserted=0、轴距 0.01083 m、轴向参数 1.0123，插入操作始终未完成。"
+  },
+  "unmet": {
+   "peg_inserted": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 21.5,
+    "log_step": null,
+    "evidence_zh": "观察：视频 21.2–25.04 s，机器人已将红销带到蓝管口，但红条在右腕视图中仍几乎整段外露，推送和侧向微调后也未出现成功参考视频 8–14.08 s 中端部被蓝管明显遮住的深入关系。评分器给出的末态轴距 10.83 毫米、轴向参数 1.0123 确认最终既偏离轴线又深度不足。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "轨迹或逆解",
+    "t": null,
+    "log_step": 191,
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent 的插入控制未找到并执行满足目标的臂位姿轨迹。log step 158 的运动学程序���打印多次 IK miss；log step 174 指定腕相机末端高度 0.163 m，执行后的估计高度为 0.1504 m。log step 191 读取右臂关节并计算限位裕量，腕角约 -1.869 rad、裕量约 -0.001 rad，且其旋转搜索中最佳位置残差仍为 0.02044 m；log step 202 指令沿负 x 推进并向上抬，执行后腕相机估计位置从 [-0.0018,0.0588,0.1478] m 变为 [-0.0069,0.0735,0.1448] m，推进很小且高度反降。结合视频 21.5–25.04 s 插销停留在入口附近，这些诊断支持逆解及受限轨迹执行问题。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或全过程原生几何读数，无法用像素精确量出中途毫米级轴距和深度；节点区间依据四视图可见关系及成功参考判断，末态由评分器几何字段明确判定。根因中的腕位姿和逆解残差来自 agent 自建运动学程序的诊断，缺少独立关节轨迹和接触力记录，无法进一步分离限位、跟踪偏差与接触影响。"
+ },
+ "summary_zh": "机器人多次重新抓取后，以右侧夹爪将红销送到蓝色套筒口，但未完成插入，节点末态为 0/1。视频约 21.5–25.04 s 的推进与微调始终未修正插入位姿，评分器末态轴距 10.83 毫米、轴向参数 1.0123；日志中的腕关节限位与逆解残差支持轨迹或逆解问题。缺少逐帧状态和接触力记录，根因细分仍有上述诊断依赖。",
+ "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 peg_inserted 末态 false、无区间、抓握接触、断点 peg_inserted(21.5 s 附近，操作中) 与现象“位姿不准”都与视频帧和评分器字段(inserted=0, tip_to_axis 0.01083, axial_t 1.0123)一致；根因“轨迹或逆解”已标为推断，log step 158/174/191/202 的内容与引用相符，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": "2026-10-30T03:21:35Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-standard--cur5s7c/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--cur5s7c/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8b4ad35ebacd8fe49f2a9c033fcaef538f11038c6f02e5f7bb68a5ec03403b39",
+   "video_sha256": "d2bfe376dc66f7d5551dbca06be575f5c8e045419f7d6269c4b030672a0fa9d6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-standard--ekz8zex.json

@@ -0,0 +1,223 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-standard--ekz8zex",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "control_steps": 7231,
+   "calls": 166,
+   "elapsed_s": 1686.8,
+   "video_rendered": 1,
+   "no_verdict": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "resets": 0,
+   "reset_budget": 0,
+   "rows": 7231,
+   "done": 1,
+   "video_done": 1,
+   "episodes": 1,
+   "control_steps_total": 7231,
+   "n_actions": 7231,
+   "video": 1,
+   "video_pending": 0,
+   "inserted": 0,
+   "tip_to_axis_m": 0.03181,
+   "axial_t": 0.3314,
+   "socket_z": 0.0363,
+   "peg_z": 0.0108,
+   "time_s": 144.62,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.307,
+   "peak_processes": 8,
+   "peak_disk_gb": 0.521,
+   "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/mujoco_playground_aloha_single_peg_insertion-standard--ekz8zex/replay.mp4",
+  "duration_s": 24.16,
+  "fps": 25.0,
+  "frames": 604,
+  "sha256": "bf897ddcb4acce78e14cdc7f71ac37a4cc8044b77dabfb9960bda4b484e3601d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 四视图中红色插销与蓝色套筒分开放在桌面，初态没有插入关系。"
+     },
+     {
+      "t": 7.8,
+      "log_step": null,
+      "text_zh": "观察：视频 7.4–8.8 s 机器人两爪各夹一件并靠近，腕部视图中的红条与蓝管存在侧向及高度错位，红条露在管外；随后视频 9.2–18.4 s 的多次调整仍可见红条沿蓝管外侧伸出，没有成功参考视频末段那种从开口进入、露出长度缩短的关系。"
+     },
+     {
+      "t": 20.8,
+      "log_step": null,
+      "text_zh": "观察：视频 18.6–21.2 s 机器人再次调整两臂，红条留在桌面附近，蓝管转向并移到其旁边；视频 21.4–24.12 s 两臂保持最后姿态，右腕视图的夹爪前没有夹住红条，俯视及正视仍显示红条偏在蓝管外。"
+     },
+     {
+      "t": 24.12,
+      "log_step": null,
+      "text_zh": "观察：末帧视频 24.12 s 未见插销端从套筒开口进入；inputs/run.json 的在线末态评分字段 inserted=0、tip_to_axis_m=0.03181（米），大于严格小于 0.005 米的要求，axial_t=0.3314 则在允许范围内。该节点末态为假，不能从轴向参数单项合格认定插入。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "抓握",
+      "note_zh": "观察：视频 7.4–8.8 s 左侧二指夹爪夹持蓝色套筒，右侧二指夹爪夹持红色插销，两臂共同移动并尝试对准送入；此前先后取物，后续又反复重抓、调整。视频 24.12 s 右爪前已没有夹住红条。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.03181
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 0.3314
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 0.3314
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "peg_inserted",
+   "t": 7.8,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 7.4–8.8 s 机器人正在用两爪将红色插销送向蓝色套筒，但红条与开口错位，仍在管外；之后多轮调整直至视频 24.12 s 都未完成该插入步骤。末态评分 inserted=0、端点轴距 31.81 毫米，确认这一步最终没有修复。"
+  },
+  "unmet": {
+   "peg_inserted": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 7.8,
+    "log_step": null,
+    "evidence_zh": "观察：视频 7.8 s 两腕视图显示红色插销与蓝色套筒没有同轴，红条贴近管外而非从开口进入；视频 12.4–18.4 s 红条仍偏在蓝管一侧，视频 24.12 s 最终红条与套筒也未形成插入关系。末态评分 tip_to_axis_m=0.03181 米，轴距超出 5 毫米要求。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "视频 / 日志",
+    "evidence_zh": "推断：机器人在视频 7.4–8.8 s、9.8–18.4 s 反复执行靠近和送入，但没有把两件物体的相对位姿控制到插入所需精度。log step 145、304、323、328 的脚本明确以对准并插入为目标、执行关节插值轨迹，视频中实际物体仍在管外，说明问题发生在插入的具体操控执行；缺少逐帧物体姿态及接触记录，控制子类不能确定。"
+   },
+   {
+    "cat": "认知",
+    "item": "误以为完成",
+    "source_zh": "日志 / 评分器字段 / 视频",
+    "evidence_zh": "推断：agent 的错误完成判断促使它保留失败姿态并停止纠偏。log step 334 将控制器改为仅保持当前关节角，称其为 winning pose；log step 337 实际执行保持，log step 339 最终声称已核验轴距小于 1.5 毫米且插入成功，随后结束，而 log step 338 的资源记录仍有 124.9 秒预算。视频 23.04–24.12 s 保持姿态未形成插入，在线最终评分却为 inserted=0、轴距 31.81 毫米，否定其完成判断。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态日志或接触力记录；多视图足以判断可见的错位和末态失败，但无法精确测量全过程的毫米级轴距、插深，亦无法确定具体控制偏差来自持物相对位姿变化、轨迹求解还是跟踪。节点区间按已检查画面判定，末态使用原生评分字段；agent 自算的点云指标及最终成功声明不当作评分器读数，也未把 agent 时间换算成视频时间。"
+ },
+ "summary_zh": "机器人曾用双臂分别抓握蓝色套筒和红色插销并多次尝试对准，唯一的插入节点未见达成，末态为 0/1。断点是插入操作中的位姿错位：最终轴距 31.81 毫米超限，agent 却误判成功并保持该姿态后结束。无逐帧状态和接触记录，具体控制子类及全过程毫米级几何无法精确核验。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点 peg_inserted 末态为假、末态评分 inserted=0 与视频一致；断点、位姿不准现象及两个根因有视频和日志依据，时钟未换算，未发现需要修改的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:10:57Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-standard--ekz8zex/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--ekz8zex/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "ada6ef14f1237ffb3adbdbe200d53830e9b361fc6a00e44e725da2c1baca2a43",
+   "video_sha256": "bf897ddcb4acce78e14cdc7f71ac37a4cc8044b77dabfb9960bda4b484e3601d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-standard--m4vtq8s.json

@@ -0,0 +1,192 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-standard--m4vtq8s",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "control_steps": 12605,
+   "calls": 891,
+   "elapsed_s": 1695.2,
+   "video_rendered": 1,
+   "grader_error": 0,
+   "sim_error": 0,
+   "rows": 2461,
+   "done": 1,
+   "live_success": 0,
+   "inserted": 0,
+   "tip_to_axis_m": 0.08908,
+   "axial_t": 1.6346,
+   "socket_z": 0.0228,
+   "peg_z": 0.0108,
+   "time_s": 252.1,
+   "video": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/mujoco_playground_aloha_single_peg_insertion-standard--m4vtq8s/replay.mp4",
+  "duration_s": 24.08,
+  "fps": 25.0,
+  "frames": 602,
+  "sha256": "087f118f5adb54973076430e142c39ebd194dc0472227e65f8d9aae8f8fa529c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的两幅外部视图显示红色插销和蓝色套筒分开放在桌面，初始没有插入。"
+     },
+     {
+      "t": 4.2,
+      "log_step": null,
+      "text_zh": "观察：视频 3.8–5.2 s，右侧机械臂的空夹爪从物体附近抬起并向外摆动，红色插销仍在桌面；视频 6.0–15.6 s 的反复末端运动也未带起插销，未出现红端进入蓝色套筒的关系。"
+     },
+     {
+      "t": 22.4,
+      "log_step": null,
+      "text_zh": "观察：视频 18.0–22.0 s，右夹爪在红色插销附近反复调整，随后视频 22.2–23.4 s 向桌边偏离；插销留在桌面，未随夹爪移动。"
+     },
+     {
+      "t": 24.04,
+      "log_step": null,
+      "text_zh": "观察：视频 23.2、23.4、23.6、23.8、24.04 s，外部视图一直显示两件物体分离，双臂最终悬停；没有参考成功视频 13.8 s 所见的红条一端深入蓝管并被遮住的特征。评分器末态字段 inserted=0、tip_to_axis_m=0.08908、axial_t=1.6346，与可见未插入一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "抓握",
+      "note_zh": "观察：右臂二指夹爪在插销附近调整并尝试夹取红色插销（视频 18.0–22.4 s），但未见夹住或带离桌面；此处记录失败的抓取尝试，不表示已建立抓握。左夹爪未见抓起蓝色套筒，未形成双臂各持一件的协作。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "value": 0.08908,
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better"
+  },
+  {
+   "name": "axial_t",
+   "value": 1.6346,
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better"
+  },
+  {
+   "name": "axial_t",
+   "value": 1.6346,
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "peg_inserted",
+   "t": 4.2,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 4.2 s，右臂空夹爪抬离桌上物体区域，红色插销未随之拿起；之后虽多次恢复臂姿并重试，直到视频 24.04 s 仍未建立持销并插入的关系。该时刻是抓取阶段偏离的粗略导航锚点，不是由 agent 时钟换算；抓取属于 DAG 的 peg_inserted 操作阶段。"
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没建立接触",
+    "t": 4.2,
+    "log_step": null,
+    "evidence_zh": "观察：视频 3.8–5.2 s 的右夹爪抬离，以及视频 18.0–23.4 s 的再次靠近和偏离，均未使红色插销离开桌面或随夹爪运动；视频 24.04 s 红销与蓝套筒仍分离。推断：插入所需的持销抓握未建立，而非持住后滑落；不另将同一空夹事件重复标为位姿不准。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "反馈控制",
+    "source_zh": "日志",
+    "evidence_zh": "观察：单局在线会话的 log step 79 中，agent 脚本明确把右夹爪目标设在红销上方，自己的伺服误差输出从 0.2033 m 增至 0.3624 m，并在后续轮次仍不收敛；log step 113 再次接近时误差从 0.0596 m 增至 0.3023 m，脚本闭爪前后的红销估计位置不变。视频 4.2–5.2 s 和 22.2–23.4 s 也可见空夹爪向外偏离，物体仍留在桌面。推断：接近抓取目标的视觉反馈运动未收敛，导致所需抓握没有建立；这些日志数值用于解释控制失败，不作为精确物体状态或视频时间映射。"
+   }
+  ],
+  "gap_zh": "无状态日���，不能逐帧确认夹爪是否曾短暂擦碰物体，也不能从像素测出毫米级插入几何。视频已检查全程每约 0.2 s 的帧及视频 24.04 s 末帧；本次没有可见插入特征，末态另有评分器几何字段支持。日志仿真末态为仿真 252.1 s，而视频仅 24.08 s，未建立逐控制步到视频的映射，故不将日志动作与视频时刻一一对应。agent 在 log step 106、117 自述末端特征误跟到桌框、雅可比不可靠，但缺少可独立核验的逐帧估计真值，无法进一步拆分感知误差与雅可比辨识误差。",
+  "unmet": {
+   "peg_inserted": false
+  }
+ },
+ "summary_zh": "本次插入节点全程未达成，末态为 0/1；机器人反复调整夹爪，但未见红色插销被抓住并送入蓝色套筒，视频 24.04 s 两件物体仍分离。断点位于插入步骤的抓取操作阶段，现象是所需抓握未建立，日志支持接近目标的视觉伺服反复发散。无状态日志且缺少日志到视频的精确映射，不能进一步核验短暂擦碰或区分末端定位与雅可比估计误差。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "peg_inserted 全程未达成（评分器 inserted=0、tip_to_axis 0.089 m）与视频一致；断点在插入步骤的抓取阶段，现象“没建立接触”与根因“反馈控制”有日志（id 79、113、117）和视频支持，时钟没有把 agent 时间换算成视频时间。未发现需要改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:13:43Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--m4vtq8s/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--m4vtq8s/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "19e16ac6ace858addd9756e6ebc0a3ed826df55f03445a52c686c6cfd5db8f82",
+   "video_sha256": "087f118f5adb54973076430e142c39ebd194dc0472227e65f8d9aae8f8fa529c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/aloha_single_peg_insertion/trajectories/mujoco_playground_aloha_single_peg_insertion-standard--yfzetnw.json

@@ -0,0 +1,234 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_aloha_single_peg_insertion-standard--yfzetnw",
+ "taskKey": "mujoco-playground/aloha_single_peg_insertion",
+ "dag_sha256": "9709e85780759f3204a58aae58bbc9f03fd320c1c8d74829d8d296fb394797b0",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "control_steps": 1512,
+   "calls": 6,
+   "elapsed_s": 1819.1,
+   "video_rendered": 1,
+   "no_verdict": 0,
+   "instance_mismatch": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "resets": 0,
+   "reset_budget": 0,
+   "rows": 1512,
+   "done": 1,
+   "video_done": 1,
+   "episodes": 1,
+   "control_steps_total": 1512,
+   "n_actions": 1512,
+   "video": 1,
+   "video_pending": 0,
+   "inserted": 0,
+   "tip_to_axis_m": 0.14157,
+   "axial_t": 8.9131,
+   "socket_z": 0.0228,
+   "peg_z": 0.0108,
+   "time_s": 30.24,
+   "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/mujoco_playground_aloha_single_peg_insertion-standard--yfzetnw/replay.mp4",
+  "duration_s": 30.32,
+  "fps": 25.0,
+  "frames": 758,
+  "sha256": "0e6773892d2aa868610ad664c1c8e41e827959baa5c43bfb12e4a041bdabfa65"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "peg_inserted": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 红色插销与蓝色套筒分别位于桌面中央附近，插销未进入套筒。"
+     },
+     {
+      "t": 2.8,
+      "log_step": 30,
+      "text_zh": "观察：视频 0.2–3.2 s 左夹爪附近的蓝套筒被带动、翻动并移向桌面左后方；视频 12.8–13.2 s 右夹爪扫过红销，红销在桌面滑动转向。两者始终分开；视频全程的两臂运动未出现持红销向蓝管开口对准并送入的过程。log step 30 的实际提交代码为双臂随机关节扫描。"
+     },
+     {
+      "t": 30.28,
+      "log_step": null,
+      "text_zh": "观察：视频 29.2–30.28 s 两物体仍分离留在桌面，最后一帧蓝套筒在左后方、红销在较前方；未出现成功参考录像视频约 8–14 s 所示红条端部进入蓝管且露出长度缩短的特征。评分器末态 inserted=0、tip_to_axis_m=0.14157、axial_t=8.9131，轴距和深度均不满足本节点。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "手指/夹爪",
+      "how": "推/扫",
+      "note_zh": "观察：标定扫描中，视频约 2.4–3.2 s 左侧二指夹爪碰拨蓝色套筒，使其移向桌面左后方；视频约 12.8–13.2 s 右侧二指夹爪扫动红色插销，使其在桌面滑动并转向。这些是先后发生的单夹爪接触，未见双臂各持一件配合插入。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "tip_to_axis_m",
+   "unit": "m",
+   "threshold": 0.005,
+   "direction": "lower_better",
+   "value": 0.14157
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": -0.875,
+   "direction": "higher_better",
+   "value": 8.9131
+  },
+  {
+   "name": "axial_t",
+   "unit": "1",
+   "threshold": 0.75,
+   "direction": "lower_better",
+   "value": 8.9131
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "peg_inserted",
+   "t": null,
+   "log_step": 30,
+   "where": "首步前",
+   "evidence_zh": "观察：评分视频 0.0–30.28 s 两臂做广泛姿态扫描并偶然拨动物体，没有开始面向套筒开口的插销抓取、对准和送入操作。log step 30 的代码生成 54 组随机双臂关节姿态，分四批提交 476+476+476+84=1512 个控制步；评分器 control_steps=1512、resets=0，未记录后续操作提交。推断：该评分尝试止于控制标定，插入步骤的定向操作尚未启动，不能把扫描中的碰拨当作插入尝试；没有可定位的单一插入失败时刻。"
+  },
+  "unmet": {
+   "peg_inserted": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": null,
+    "log_step": 30,
+    "evidence_zh": "观察：视频 0.0–30.28 s 两臂反复摆动，蓝套筒与红销没有接合；log step 30 唯一实际运动脚本是随机关节扫描，后续 log step 31–40 仅做运动学拟合和代码修正。推断：机器人没有开始有目标的抓取对准插入操作，物体在扫描中被拨动���构成该操作的启动。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "认知",
+    "item": "推理/规划",
+    "source_zh": "日志",
+    "evidence_zh": "观察：inputs/run.json 的 log step 9–24 反复处理相机坐标、点云和标定约定，log step 25–27 转向腕相机方案；log step 28 中 agent 自述“Starting with the sweep first”，决定先拟合各臂运动学再执行任务。log step 30 的实际代码生成 54 组随机双臂关节姿态，四批提交 476+476+476+84=1512 个控制步；评分器 reward.control_steps=1512、episodes=1、resets=0，log step 31–40 只继续拟合及排错，没有后续运动提交。视频 0.0–30.28 s 两臂广泛扫动、先后拨动物体，未见持红销对准蓝套筒并送入。推断：agent 将相机标定和完整运动学拟合置于任务操作之前，反复排错后仍将评分回合内全部实际运动用于随机标定扫描，未转入目标操作；这是跨阶段策略与执行顺序的规划问题。此归因依据具体标定路线和动作提交内容，而非以时间耗尽或未交付本身作根因；没有证据证明某次面向插入的轨迹、逆解或跟踪控制失败。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 中有完整评分视频和运行日志，无逐帧状态日志；视频能确认两物体全程未插合及夹爪拨动，但视频约 0.2–1.4 s 左夹爪与蓝套筒的短时接触部分遮挡，不能确证是否曾形成对向夹持。log step 40 未给出最终完整运动学拟合结果，且没有提交面向抓取与插入的控制，无法评估该模型最终能否支持任务控制。"
+ },
+ "summary_zh": "评分轨迹仅包含双臂随机标定扫描，夹爪先后拨动蓝套筒和红销，插入节点从未达成，末态为 0/1，轴距 0.14157 米、轴向参数 8.9131 均超出验收范围。插入的定向操作尚未启动；日志中的相机标定、运动学拟合路线与实际扫描提交支持“认知／推理/规划”根因：反复标定排错后，评分回合内的全部运动仍用于扫描。无状态日志，初段蓝套筒是否曾短时被夹住无法确定；未提交任务控制，不能判断其控制能力。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 7,
+    "summary_zh": "WHERE 标签（peg_inserted 未达成、contact_actual、首步前/没尝试）与视频一致，评分器 inserted=0 也支持。问题在根因：把「控制」归因于 agent 未把运动学建模落实为控制器，实质是预算耗尽/未交付的复述，证据不支持控制子类；日志显示的是策略选择（唯一的评分回合被用来做标定扫描）。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "peg_inserted 未达成、contact_actual、首步前/没尝试、认知·推理/规划根因均与视频、log step 28/30/31 及评分器 inserted=0 一致；上一轮的控制根因问题已修正。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:28:34Z",
+  "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/mujoco_playground_aloha_single_peg_insertion-standard--yfzetnw/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--yfzetnw/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--yfzetnw/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_aloha_single_peg_insertion-standard--yfzetnw/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8dbb36c0a866cbe0be157135ce25d2f6792dbd9ab9bfc2b0b133525cfa6db89d",
+   "video_sha256": "0e6773892d2aa868610ad664c1c8e41e827959baa5c43bfb12e4a041bdabfa65",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-privileged--b8k85uu.json

@@ -0,0 +1,198 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-privileged--b8k85uu",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 0,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.3178,
+   "worst_yaw_err": 0.7281,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 33,
+   "agent_resets": 85,
+   "agent_sim_steps": 171215,
+   "agent_replays": 9,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.474,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.241,
+   "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/mujoco_playground_barkour_joystick-privileged--b8k85uu/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "1e0c15e8a5d5f301beb2c9dcfeb52c73aa9a0a70dd6a3e37b14e7cfb36d136a7"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.000–10.143 s中，Barkour四足机器人在棋盘格地面上交替迈腿并移动，前半段机身航向有来回偏摆，视频5.143–10.143 s继续移动并逐渐转向；视频9.143–10.143 s逐帧检查仍见机身由腿支起，没有倒地。"
+     },
+     {
+      "t": 3.286,
+      "log_step": null,
+      "text_zh": "观察：视频2.500–4.500 s逐帧可见机器人一边迈步一边改变航向，随后又转回；成功参考视频同一前半段的机身朝向较稳定。参考仅用于识别行进和转向外观，不替代本次评分。"
+     },
+     {
+      "t": 10.143,
+      "log_step": 62,
+      "text_zh": "观察：log step 62对最终交付的trajectory.npz执行评分器回放，输出steps=1000、fell_at=null、tracking_ok=false；两段平移误差分别为0.2927和0.3178 m/s，偏航误差分别为0.5540和0.7281 rad/s。run.json评分字段也给出n_actions=1000、fell=0、deterministic=1及相同最差误差。"
+     },
+     {
+      "t": 10.143,
+      "log_step": 62,
+      "text_zh": "推断：完整日程虽已运行且机器人未倒地，但两段速度跟踪误差均超过平移0.2 m/s、偏航0.3 rad/s的验收阈值，因此“按指令行进”从未满足完整判据，final=false；不是仅凭success=0作此判断。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.143–10.143 s，Barkour四条腿的足端交替落在棋盘格地面，承重并推动躯干行进和转向；视频末秒仍由足腿支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "value": 0.3178,
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better"
+  },
+  {
+   "name": "worst_yaw_err",
+   "value": 0.7281,
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 3.286,
+   "log_step": 62,
+   "where": "操作中",
+   "evidence_zh": "观察：视频2.500–4.500 s机器人正在连续迈步，机身航向明显偏摆后回转；该处取视频3.286 s为偏差的粗略导航锚点。log step 62最终交付回放表明第一段平移/偏航均值误差0.2927 m/s、0.5540 rad/s已经超标，第二段也未达标；机器人最终未倒地但整个按指令行进步骤始终未修复完成。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 3.286,
+    "log_step": 62,
+    "evidence_zh": "观察：视频2.500–4.500 s中机器人迈步时航向来回变化，视频5.143–10.143 s虽继续移动和转向，但与成功参考后半段相比转向进展较小；视频末秒仍保持足腿支撑。log step 62及评分器记录确认两段的平移与偏航速度均未跟上指令。这是未倒地的行进偏差。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 55,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 55的生成代码将当前指令与速度、偏航反馈叠加，生成���期足端扫动轨迹并以IK转换成关节目标后录制；log step 56保存的存活配置与log step 62最终交付回放的两段误差完全一致，仍有显著速度跟踪残差。结合视频2.500–4.500 s的航向偏摆，支持该步态控制未能实现所需平移和转向，而非未启动或途中摔倒；现有证据不能在接触设计、反馈控制及参数之间可靠细分。"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志；视频能够确认迈步、转向及未倒地，精确速度均值采用与最终评分字段吻合的log step 62交付文件回放结果。不能定位均值误差首次超阈值的瞬间，视频3.286 s仅作可见偏摆的导航锚点，也不能进一步确认足端滑移、接触力分配或某个增益参数的独立因果作用。会话中其他配置的摔倒不属于本次交付回合。"
+ },
+ "summary_zh": "机器人运行完整1000步，视频末秒仍由足腿支撑，但两段平移和偏航速度误差均超标，唯一节点“按指令行进”最终为0/1。断点在该步骤操作中，现象为未倒地的步态跟踪偏差；日志支持步态控制未实现所需速度，不能可靠细分具体控制机制。无逐步状态日志，精确误差来自最终交付文件回放及评分字段，偏差开始时刻只能给粗略视频锚点。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "唯一节点“按指令行进”final=false、断点（操作中）、步态现象与控制根因均有视频和 log step 62 的评分器回放（steps=1000、fell_at=null、两段误差 0.2927/0.3178、0.554/0.7281，与 run.json 一致）支持；视频全程未倒地，与参考视频相比无矛盾。时钟、contact_actual（腿/躯干·支撑）和 native_metrics 均正确。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:03:32Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--b8k85uu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--b8k85uu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c015a416e86a69cff78ae57f33b0063775aac6e232bde6c95dc1273762e374e6",
+   "video_sha256": "1e0c15e8a5d5f301beb2c9dcfeb52c73aa9a0a70dd6a3e37b14e7cfb36d136a7",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-privileged--flcfwwu.json

@@ -0,0 +1,197 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-privileged--flcfwwu",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 0,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.3568,
+   "worst_yaw_err": 0.8218,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 85,
+   "agent_resets": 3778,
+   "agent_sim_steps": 1279530,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.884,
+   "peak_processes": 18,
+   "peak_disk_gb": 0.981,
+   "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/mujoco_playground_barkour_joystick-privileged--flcfwwu/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "b9b42ccde9af38e61b0f2caa4a99065454b3d68a97f0ee6ca9092b9be94684d5"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0 s起，Barkour四足机器人站在棋盘格地面上；视频0.5–5.0 s足端交替落地、机身平移且由腿支起。起始站立及前半段移动不满足节点要求的完整指令日程。"
+     },
+     {
+      "t": 5.5,
+      "log_step": null,
+      "text_zh": "观察：视频5.0、5.5、6.0及7.0 s，机器人继续迈步，机身仍大致保持原先侧向朝向；视频8.0–10.14 s才逐渐转成近端面朝相机，未见倒地。成功参考在其视频5.5–10.0 s呈现持续且大幅的转弯；参考仅用于辨认视角中的转向表现。"
+     },
+     {
+      "t": 10.14,
+      "log_step": 149,
+      "text_zh": "观察：log step 147保存最后选中的轨迹；log step 149由评分工具在新进程回放该文件，报告1000步、fell_at=null，第一段lin_err=0.0905 m/s、yaw_err=0.0945 rad/s，第二段lin_err=0.3568 m/s、yaw_err=0.8218 rad/s。inputs/run.json的reward给出相同的两项最差误差、n_actions=1000、fell=0和deterministic=1；视频10.14 s机器人仍由足端支撑。"
+     },
+     {
+      "t": 10.14,
+      "log_step": 149,
+      "text_zh": "推断：最终交付文件回放结果与评分器字段相符，且视频支持完整行进但转向不足的表现；第二段平面速度误差0.3568 m/s超过0.2 m/s、偏航角速度误差0.8218 rad/s超过0.3 rad/s，因此完整日程节点始终未达到，终值为false。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.5–10.14 s，Barkour四条腿的足端交替接触棋盘格地面，支撑躯干并在迈步中移动；视频末秒机身仍离地，未见躯干着地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "value": 0.3568,
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better"
+  },
+  {
+   "name": "worst_yaw_err",
+   "value": 0.8218,
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 5.5,
+   "log_step": 149,
+   "where": "操作中",
+   "evidence_zh": "推断：log step 149对最终交付文件的回放表明第一段两项误差均达标，第二段两项均超阈值；视频5.5–7.0 s机器人换段后继续以接近原先朝向迈步，随后到视频10.14 s也未呈现所需的持续大幅转弯。断在按指令行进步骤，视频5.5 s是换段后表现的粗略导航锚点，不是由agent工作时间换算。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 5.5,
+    "log_step": 149,
+    "evidence_zh": "观察：视频5.5–7.0 s机器人迈步但朝向变化很小，视频8.0–10.14 s才缓慢转成近端面朝相机；其转弯幅度明显小于成功参考的对应阶段，全程及末秒仍有足端支撑。log step 149的最终文件回放确认第二段平移和偏航跟踪均未达标，属于未倒地的行进偏差。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 143中的生成脚本把当前指令s.commands[t]、速度和角速度送入Trot.act并记录关节目标；log step 147进一步选择并保存了轨迹，log step 149回放该最终文件仍得到第二段lin_err=0.3568 m/s、yaw_err=0.8218 rad/s，与评分器一致。结合视频5.5–10.14 s持续迈步却未形成所需转弯，支持该步态控制方案未能实现第二段速度指令；没有足够证据细分为逆解、反馈或某个参数错误。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志，无法确定第二段误差的逐时变化或归因到具体落足、逆解、反馈参数；视频只能支持运动及转向表现，精确误差采用评分器与最终文件回放记录。断点视频5.5 s为粗略锚点，不能视为首次瞬时超差时刻。"
+ },
+ "summary_zh": "机器人运行满1000个控制步，视频末秒仍由足端支撑，评分器也记录fell=0，但唯一的“按指令行进”节点未完成（0/1）。第一段跟踪达标，第二段转弯表现不足，平面速度及偏航误差分别为0.3568 m/s和0.8218 rad/s，断点在换段后的行进操作中，现象为步态，根因保留为控制。缺少逐步状态日志，不能进一步锁定控制器的具体失效机制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成（终值false）、断点在换段后的行进、步态现象和控制根因均有log step 149评分器回放（第二段lin_err 0.3568、yaw_err 0.8218）与视频支持；时钟均为视频时间，无换算问题。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:02:32Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--flcfwwu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--flcfwwu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "967b33e9312f82c2c0c1757eb112383455a6b3efd57b5799ae7cb335f92af238",
+   "video_sha256": "b9b42ccde9af38e61b0f2caa4a99065454b3d68a97f0ee6ca9092b9be94684d5",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-privileged--gcdbhfc.json

@@ -0,0 +1,202 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-privileged--gcdbhfc",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 0,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.3056,
+   "worst_yaw_err": 0.4985,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 227,
+   "agent_resets": 424,
+   "agent_sim_steps": 1936815,
+   "agent_replays": 12,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.734,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.833,
+   "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/mujoco_playground_barkour_joystick-privileged--gcdbhfc/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "119a8ab825c29c0764649961832eafbba0df1603162fbe84226f3c3771f8499e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.429,
+      "log_step": null,
+      "text_zh": "观察：视频0.429–3.214 s，Barkour机器人反复抬落腿，足端接触棋盘格地面并支撑躯干，地面相对机器人移动，行进尝试已开始。"
+     },
+     {
+      "t": 4.714,
+      "log_step": null,
+      "text_zh": "观察：视频3.429–5.357 s，机器人仍由腿支起，但黄色前端从偏向画面右侧转为较正对相机，前半段已出现明显朝向变化。"
+     },
+     {
+      "t": 9.5,
+      "log_step": null,
+      "text_zh": "观察：视频6.000–10.143 s，机器人机身保持离地，后半段朝向变化有限，视频9.071–10.143 s逐帧检查未见翻倒；成功参考视频的后半段呈持续明显转向，仅用于外观对照。"
+     },
+     {
+      "t": null,
+      "log_step": 186,
+      "text_zh": "观察：log step 186对最终交付文件trajectory.npz运行回放，报告1000/1000控制步、fell_at=null、tracking_ok=false；第0段平移/偏航误差为0.3056 m/s、0.3785 rad/s，第1段为0.3028 m/s、0.4985 rad/s。inputs/run.json评分器记录deterministic=1、fell=0，最差两项误差与该回放一致，均超过DAG阈值0.2 m/s和0.3 rad/s。"
+     },
+     {
+      "t": null,
+      "log_step": 186,
+      "text_zh": "推断：该节点要求完整日程及每段误差全部合格；交付回放和评分器均证实两段跟踪未达标，因此节点从未完成、终值为false，不能因机器人迈步且未倒地而计为达成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.429–8.357 s，Barkour四条腿的足端反复抬落、接触地面，承重并推动机身移动；视频9.071–10.143 s机身仍由足端支撑，未见躯干落地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.3056
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.4985
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 4.714,
+   "log_step": 186,
+   "where": "操作中",
+   "evidence_zh": "观察：视频4.286–5.357 s，机器人在执行前半段行进时已有明显朝向漂移；视频6.000–10.143 s后半段也未呈持续充分转向。log step 186最终交付回放的两段平移及偏航误差均超标，评分器最差误差为0.3056 m/s和0.4985 rad/s；视频4.714 s仅作可见偏航偏差的导航锚点，不代表误差均值首次超阈值时刻。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 4.714,
+    "log_step": 186,
+    "evidence_zh": "观察：视频3.429–5.357 s，Barkour在迈步中明显改变朝向，而任务前半段要求近乎不转向的平移；视频6.000–10.143 s后半段转向又很有限，机器人始终未倒地。log step 186交付回放确认两段速度跟踪均不合格：这是行进方向和转向未跟上指令的步态问题。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 105记录agent明确区分前半段前进/正侧移与后半段负侧移/正偏航的目标；log step 181记录生成器采用带相位偏移的逆运动学小跑，并给平移、偏航设置分段增益。经log step 183修正生成设置后，log step 186对最终文件的回放仍在两段都出现平移和偏航误差超标，且最差值与评分器一致；结合视频4.286–5.357 s的朝向漂移及后半段转向不足，归为关节轨迹未实现所需机身速度的控制能力不足，证据不足以细分控制子类型。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有本次逐帧状态日志或完整速度时间序列；视频可见行进、朝向变化和支撑，但不能量出各帧机体系速度、足地接触力或精确累计误差。两段均值来自log step 186的最终交付回放，最差值由评分器字段核对；无法定位误差均值首次越界时刻，也无法进一步区分轨迹设计、反馈或具体参数的贡献。"
+ },
+ "summary_zh": "机器人完成了1000个控制步，视频可见迈步及足端支撑，评分器记录未翻倒，但“按指令行进”节点未达成，最终为0/1。前半段出现明显朝向漂移，后半段转向不足，最差平移误差0.3056 m/s、偏航误差0.4985 rad/s均超阈值，断点为行进操作中的步态跟踪失败，根因归为控制。缺少逐帧状态与速度序列，不能确定误差首次越界时刻或更具体的控制子类型。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成（final=false）、断点和根因与评分器字段（worst_lin_err 0.3056、worst_yaw_err 0.4985、fell=0）及 log step 186 回放一致；视频帧可见持续迈步、未倒地，未发现需修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:03:06Z",
+  "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/mujoco_playground_barkour_joystick-privileged--gcdbhfc/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--gcdbhfc/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "2d4aeb73ff7784b0a372b19b40efe365535bf9c9cc359ae763ecf8ace2bb2379",
+   "video_sha256": "119a8ab825c29c0764649961832eafbba0df1603162fbe84226f3c3771f8499e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-privileged--ieyzivc.json

@@ -0,0 +1,173 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-privileged--ieyzivc",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 1,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.131,
+   "worst_yaw_err": 0.0539,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 4,
+   "agent_resets": 6,
+   "agent_sim_steps": 15000,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.678,
+   "peak_processes": 8,
+   "peak_disk_gb": 0.828,
+   "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/mujoco_playground_barkour_joystick-privileged--ieyzivc/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "8b16175b9799740b1fde0b75c77daa3b25f6aafbebde919affac999ae002350d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.143,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.571,
+      "log_step": null,
+      "text_zh": "观察：视频 0.000 s 机器人处于四足站姿；视频 0.214–0.571 s 连续帧显示足端轮换抬落，机身由腿支起，地面棋盘格相对机器人移动；视频 0.857–4.857 s 的抽帧中机器人继续迈步平移。"
+     },
+     {
+      "t": 6.0,
+      "log_step": null,
+      "text_zh": "观察：视频 4.571–6.214 s 的逐帧检查显示机器人持续迈步，随后机身朝向发生变化；视频 6.286–9.143 s 的抽帧中转向进一步扩大，平移方向相对机身也改变，未见倒地。"
+     },
+     {
+      "t": 10.143,
+      "log_step": null,
+      "text_zh": "观察：视频 9.143–10.143 s 的逐帧检查中机器人继续交替落足，躯干始终离开地面，末帧仍保持腿部支撑。inputs/run.json 的 reward 记录 success=1、replay_success=1、deterministic=1、n_actions=horizon=1000、fell=0，worst_lin_err=0.131 m/s、worst_yaw_err=0.0539 rad/s，分别小于任务阈值 0.2 m/s 与 0.3 rad/s。"
+     },
+     {
+      "t": 10.143,
+      "log_step": 11,
+      "text_zh": "推断：该节点要求完成整个日程，不能从开局站立或前半段行走计入。log step 10 的脚本保存关节目标，log step 11 的完整日程检查报告 1000 步且两段误差达标，log step 12 随后调用所存文件的回放；这些与本次评分器字段及视频终态一致，因此节点在回合末达成，导航时间取已检查的末帧视频 10.143 s（约数）。这里未把 agent 工作时间或仿真 10.0 s 直接换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.214–0.571 s、4.571–6.214 s 及 9.143–10.143 s 的连续帧可见 Barkour 足端交替着地与抬起，落地足承重支撑躯干，机器人借此平移并转向；躯干未落地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.131
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.0539
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人完成“按指令行进”节点，最终进度为 1/1：视频前半段持续迈步平移，后半段明显转向，逐帧检查视频 9.143–10.143 s 至末帧仍保持腿部支撑。评分器确认全部 1000 个控制步完成且未触发终止条件，最大分段平均平面速度误差为 0.131 m/s、偏航角速度误差为 0.0539 rad/s，均达标。无状态日志，精确误差与终止历史依据本次评分器字段；完成时间取视频末帧约 10.143 s，未假定仿真时钟与视频时钟相同。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "单节点已达成，final=1、无failure与评分器字段（success=1、fell=0、误差0.131/0.0539达标）及视频一致；时间取视频末帧，未换��agent时间，接触标注合理。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:02:36Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--ieyzivc/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--ieyzivc/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "533eb3fb3f691ffaf4177273567eb7768d98ee47607ee5f03063c2c6f71dcb36",
+   "video_sha256": "8b16175b9799740b1fde0b75c77daa3b25f6aafbebde919affac999ae002350d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-privileged--oqmtteh.json

@@ -0,0 +1,199 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-privileged--oqmtteh",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 0,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.4832,
+   "worst_yaw_err": 0.7525,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 15,
+   "agent_resets": 46,
+   "agent_sim_steps": 96610,
+   "agent_replays": 5,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.455,
+   "peak_processes": 12,
+   "peak_disk_gb": 0.241,
+   "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/mujoco_playground_barkour_joystick-privileged--oqmtteh/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "77fc0a48811eb008e6b28bbaccdf1c23e91beb6284c3eee0dbd179c18c8cb105"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.0,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.0–10.143 s，机器人在棋盘格地面上不断屈伸、交替迈腿，足端落地支撑躯干；视频 0.286–1.714 s 可见机身随步伐小幅摆向，全程未见侧翻或躯干落地。"
+     },
+     {
+      "t": 5.5,
+      "log_step": null,
+      "text_zh": "观察：视频 4.714–6.286 s 及视频 6.857–10.143 s 中，机器人继续迈步，黄色头部仍大致朝画面右侧，只有小幅往复摆向，未出现后半程持续明显转向。作为外观参照，成功示范视频 5.5–9.5 s 的机器人从朝右逐渐转至朝左、再朝向相机；该示范仅用于辨识转向外观。"
+     },
+     {
+      "t": 10.143,
+      "log_step": 16,
+      "text_zh": "观察：log step 16 记录机器人轨迹执行和回放均为 1000 步、fell_at=null，随后保存 (1000,12) 轨迹；两段平移误差为 0.4113、0.4832 m/s，偏航误差为 0.7525、0.6784 rad/s。最终评分器同样报告 n_actions=1000、fell=0、worst_lin_err=0.4832、worst_yaw_err=0.7525、deterministic=1；视频最后一秒至视频 10.143 s 仍在支撑迈步。"
+     },
+     {
+      "t": 10.143,
+      "log_step": 16,
+      "text_zh": "推断：已保存回放的最差误差与最终评分完全一致，且视频中缺乏持续转向的表现相符，故评分片段对应这条基线轨迹；后续 log step 22、31、37 的短程跌倒试验不是评分片段。完整日程结束仍不满足平移 0.2 m/s 和偏航 0.3 rad/s 两项阈值，因此“按指令行进”从未完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.286–1.714 s、4.714–6.286 s 和 9.214–10.143 s，Barkour前后腿持续交替屈伸，足端接触棋盘格地面并支撑、推动躯干；未见躯干翻落地面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "value": 0.4832,
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better"
+  },
+  {
+   "name": "worst_yaw_err",
+   "value": 0.7525,
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 1.0,
+   "log_step": 16,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.286–1.714 s 机器人已在迈步，机身反复小幅摆向；视频 5.071–10.143 s 继续行走却未形成后半段所需的持续转向。log step 16 保存的完整回放两段速度误差均超阈值，最终评分器确认 worst_lin_err=0.4832 m/s、worst_yaw_err=0.7525 rad/s 且 fell=0。推断：失败发生在持续执行“按指令行进”期间，视频 1.0 s 仅作早段异常行进的粗略导航锚点，并非均值判据的瞬时越线时间。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 1.0,
+    "log_step": 16,
+    "evidence_zh": "观察：视频 0.286–1.714 s 及 4.714–6.286 s 中机器人不断迈步，机身随步伐往复摆向；视频 6.857–10.143 s 头部仍大致朝右，未实现持续转向，��到末帧仍由腿支起。log step 16 的已保存回放与最终评分的误差数值一致，且最终 fell=0。推断：这是未能实现所需平移、转向跟踪的行进步态，未见翻倒。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": null,
+    "log_step": 17,
+    "source_zh": "日志",
+    "evidence_zh": "观察：log step 10 的脚本以步态相位、足端目标和逆运动学生成关节轨迹，并用速度、偏航反馈修正目标；log step 16 运行后保存基线。log step 17 在修改该控制器前重跑诊断，第一段平均平移速度约为 [0.123,-0.004] m/s，对应指令 [0.459,0.221] m/s；第二段平均速度约为 [0.166,0.073] m/s，对应指令 [0.090,-0.389] m/s，第二段平均偏航速度约 -0.0039 rad/s，对应指令 +0.643 rad/s。该诊断是会话证据；评分片段本身的跟踪失败另由 log step 16 的已保存回放、最终评分器误差字段以及视频 5.071–10.143 s 缺乏持续转向共同支持。推断：脚本虽设置了正确的速度日程，生成的动作未能产生所需的机体速度与转向，根因可归于控制；现有证据不足以唯一细分接触设计、反馈律或参数问题。"
+   }
+  ],
+  "gap_zh": "没有逐帧仿真状态日志，不能独立重算速度误差均值、足地接触力或足端净空；精确验收依赖最终评分器及 log step 16 的已保存回放。log step 18 以后关于下沉、摆腿拖地等解释来自会话诊断和 agent 自述，未在本次评分片段中独立验证，故不据此指定更细控制根因。连续误差的失败没有唯一瞬时事件，断点时刻仅为视频导航锚点。"
+ },
+ "summary_zh": "机器人跑满 1000 个控制步并持续由腿支撑，但“按指令行进”未完成，最终进度为 0/1；后半程没有形成明显持续转向，最差平移误差 0.4832 m/s、偏航误差 0.7525 rad/s 均超过阈值。断点位于该步骤操作中，现象为步态未达到要求，日志支持控制层面的速度跟踪不足；本次交付是 log step 16 保存的基线，后续会话试验中的跌倒不属于评分视频。缺少状态日志，无法进一步确认足地接触及更细控制机制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成（0/1）、断点、步态现象与控制根因均有视频、log step 16/17 与评分器字段支持；未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:03:04Z",
+  "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/mujoco_playground_barkour_joystick-privileged--oqmtteh/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--oqmtteh/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "6e3b9ab4d15a2cc1071211bf8b04ede193aa7559c8a62e22b0fb80ab40cf28e2",
+   "video_sha256": "77fc0a48811eb008e6b28bbaccdf1c23e91beb6284c3eee0dbd179c18c8cb105",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-privileged--r6czfhq.json

@@ -0,0 +1,173 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-privileged--r6czfhq",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 1,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.1631,
+   "worst_yaw_err": 0.28,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 584,
+   "agent_resets": 1418,
+   "agent_sim_steps": 1468815,
+   "agent_replays": 18,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 6.747,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.256,
+   "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/mujoco_playground_barkour_joystick-privileged--r6czfhq/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "12fd5cd903a233ed92b026c73b509786d6586394dce689d5b2b29d5a827418c2"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.14,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–5.0 s 的连续抽帧中，Barkour 四足机器人在棋盘格地面上交替迈腿，地面相对机器人移动，躯干由腿支起；起始站立尚未完成整段日程。"
+     },
+     {
+      "t": 5.5,
+      "log_step": null,
+      "text_zh": "观察：视频 4.50–6.50 s 逐帧核对显示机器人继续迈步并改变机身朝向；视频 7.0–10.14 s 可见持续转向后的行进，视频 9.14–10.14 s 逐帧未见倒地，足端落地支撑躯干。"
+     },
+     {
+      "t": null,
+      "log_step": 187,
+      "text_zh": "观察：log step 181 的生成脚本在执行后保存关节目标轨迹，log step 182 生成完成；log step 183 对该交付文件的两次官方回放均成功且 state_sha 相同，log step 187 最终回放报告 steps=1000、horizon=1000、success=true、fell_at=null、tracking_ok=true；两段 lin_err 分别为 0.1631、0.1617 m/s，yaw_err 分别为 0.2662、0.2800 rad/s，与本次 reward 的最差误差及成功字段一致。"
+     },
+     {
+      "t": 10.14,
+      "log_step": null,
+      "text_zh": "推断：结合本次评分器 n_actions=horizon=1000、fell=0、replay_success=1，以及视频全程行进、后半段转向和视频 10.14 s 最后帧仍受腿支撑的表现，完整日程验收成立，节点 final=true；达成时间以视频末帧约 10.14 s 为锚点（约半秒精度），不把会话中的早期试验计入本次进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–10.14 s，Barkour 四条腿的足端交替落在棋盘格地面上，支撑躯干完成平移及转向；视频 4.50–6.50 s 和 9.14–10.14 s 逐帧可见迈腿与落足。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.1631
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.28
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人完成“按指令行进”节点（最终 1/1）：视频前半段持续平移，后半段明显转向，直到视频 10.14 s 末帧仍由足端支撑，评分器确认执行满 1000 个控制步且成功。最差平面速度误差 0.1631 m/s、偏航角速度误差 0.28 rad/s 均达标；本次没有状态日志，精确跟踪误差及终止判据采用评分器字段和交付文件的最终官方回放记录，节点达成时间仅以视频末尾近似定位。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "成功运行，唯一节点 final=true、contact 腿/躯干·支撑、无失败层，均与视频帧及 log step 182/183/187 的回放结果（steps=1000、success、lin_err 0.1631、yaw_err 0.28）一致，未发现需要修改之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:05:40Z",
+  "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/mujoco_playground_barkour_joystick-privileged--r6czfhq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--r6czfhq/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7a55154ba6a4a74417493da0f84d9d453be1dd3fcd637d3f7daf5d5e4242013f",
+   "video_sha256": "12fd5cd903a233ed92b026c73b509786d6586394dce689d5b2b29d5a827418c2",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-privileged--zvvvnnm.json

@@ -0,0 +1,173 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-privileged--zvvvnnm",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 1,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.0852,
+   "worst_yaw_err": 0.1101,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 52,
+   "agent_resets": 938,
+   "agent_sim_steps": 305790,
+   "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/mujoco_playground_barkour_joystick-privileged--zvvvnnm/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "8205ffc08a4db65fe8d9e40f4b79d4cc898129e614818011df4930ebd3a0d479"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.143,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.5,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–5.0 s 的连续抽帧中，Barkour四足机器人在棋盘格地面上交替迈腿，地面纹理相对跟随相机移动，机身保持离地支起。"
+     },
+     {
+      "t": 5.5,
+      "log_step": null,
+      "text_zh": "观察：视频 4.786–5.857 s 的逐帧画面显示机器人继续迈步并开始改变机身朝向；视频 6.0–10.143 s 的画面中转向持续，视频 9.0–10.143 s 逐帧检查未见倒地，末帧机身仍由腿支撑。"
+     },
+     {
+      "t": null,
+      "log_step": 55,
+      "text_zh": "观察：log step 55 中，agent脚本重新生成并保存最终交付轨迹，随后对该文件进行两次新进程回放；工具均报告steps=1000、horizon=1000、success=true、fell_at=null、tracking_ok=true，最终状态哈希均为311c62c55717。两段平面速度误差均值分别为0.0527、0.0852米/秒，偏航角速度误差均值分别为0.0837、0.1101弧度/秒；这些数值与本次reward的最大误差及成功结果一致。"
+     },
+     {
+      "t": 10.143,
+      "log_step": null,
+      "text_zh": "推断：结合本次评分器reward的success=1、replay_success=1、deterministic=1、n_actions=horizon=1000、fell=0，以及视频末帧仍在支撑行进，完整日程及误差验收已通过；节点只在整段日程结束时计入，达到时间以已检查的末帧视频 10.143 s 为近似锚点，不把开局站立或前半段行进计为完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–10.143 s，Barkour机器人的足端随四条腿交替落向棋盘格地面，支撑躯干完成平移与转向；视频 4.786–5.857 s 的密集画面可见抬足与落足交替。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.0852
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.1101
+  }
+ ],
+ "failure": null,
+ "summary_zh": "本次完成唯一节点“按指令行进”（最终1/1）：视频前半段持续迈步平移，后半段明显转向，视频 10.143 s 末帧仍由腿支撑。评分器确认完整执行1000步、fell=0且确定性回放成功，最大分段平面速度误差0.0852米/秒、偏航角速度误差0.1101弧度/秒均低于阈值，无失败断点。没有状态日志，精确误差与全程终止条件依据本次评分器结果，完成时间以视频末帧作近似锚点。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "成功运行，唯一节点按指令行进 final=true 有评分器结果（success=1、fell=0、1000步、误差低于阈值）支持，无失败断点；log step 55 的数值与 reward 一致，时钟与接触标注无误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:02:00Z",
+  "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/mujoco_playground_barkour_joystick-privileged--zvvvnnm/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-privileged--zvvvnnm/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "3e40a1ec1f2d38d15765ecaac1c494831f83e3e6735e58911da9b8d6c020e3b9",
+   "video_sha256": "8205ffc08a4db65fe8d9e40f4b79d4cc898129e614818011df4930ebd3a0d479",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--8u3fim4.json

@@ -0,0 +1,190 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--8u3fim4",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "physics_steps": 5000,
+   "steps": 783,
+   "elapsed_s": 615.8,
+   "video_rendered": 1,
+   "fell": 0,
+   "worst_lin_err": 0.1045,
+   "worst_yaw_err": 0.1371,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 2,
+   "reset_budget": 10,
+   "episodes": 3,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.527,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.829,
+   "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/mujoco_playground_barkour_joystick-standard--8u3fim4/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "d6f123ac9f29e917ffb91ac4faf3a661f7dc1def5ceb66ef1e03061bb9455178"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.14,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s机器人从棋盘格地面上的站立姿态开始；视频0.21–5.07 s跟随视角中四条腿交替迈步，地面格纹相对机器人连续移动，躯干由足端支撑。开局站立尚未完成完整指令日程。"
+     },
+     {
+      "t": 5.5,
+      "log_step": null,
+      "text_zh": "观察：视频5.14–6.79 s机器人继续迈步，地面相对移动方向发生改变，机身朝向逐渐转动；视频7.00–10.14 s持续转向行进。逐帧检查视频9.21–10.14 s，机器人仍由腿支撑，未见倒地。"
+     },
+     {
+      "t": null,
+      "log_step": 44,
+      "text_zh": "观察：log step 28的脚本重置后开始第三回合；log step 36服务状态为episode=3、control_steps=400、done=False，log step 37脚本接续该现场回合，未再重置；log step 44服务最终状态为episode=3、control_steps=1000、done=True、reason=horizon。前两回合调试中的失败不是本次评分视频的事件。"
+     },
+     {
+      "t": null,
+      "log_step": null,
+      "text_zh": "观察：inputs/run.json的reward报告success=live_success=replay_success=1、live_matches_replay=1、deterministic=1、n_actions=sim_steps=1000、fell=0；worst_lin_err=0.1045 m/s，小于0.2 m/s；worst_yaw_err=0.1371 rad/s，小于0.3 rad/s。精确误差采用评分器记录，而非agent在log step 45自行计算的抽样均值。"
+     },
+     {
+      "t": 10.14,
+      "log_step": null,
+      "text_zh": "推断：完整视频的持续行进、后半段换向和转向与本次评分器的1000控制步成功记录相符，故按指令行进节点最终成立；该节点要求完整日程完成，以检查过的末帧视频10.14 s作为完成时间近似锚点，不将agent工作时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.21–10.14 s，Barkour四条腿的足端依次落在棋盘格地面并交替抬起，在平移、侧移及转向时支撑躯干；未见躯干倒地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "value": 0.1045,
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better"
+  },
+  {
+   "name": "worst_yaw_err",
+   "value": 0.1371,
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "本次评分为第三回合，机器人完成完整指令日程，唯一节点“按指令行进”在视频末段约10.14 s达到，最终1/1；前两回合调试失败不计入本次轨迹。视频显示机器人持续由足端支撑，前半段平移、后半段换向并转向，评分器确认1000控制步、fell=0，平移与偏航误差分别为0.1045 m/s和0.1371 rad/s，均达标。无状态日志，精确误差与历史终止条件依赖本次评分器记录，完成时刻取检查过的末帧作近似锚点。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "单节点“按指令行进”final true，与评分器 success=1、fell=0、1000 控制步、误差 0.1045/0.1371 一致；视频中机器人全程站立迈步，换向转向可见。接触与时钟标注无误，无失败标签。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:20:47Z",
+  "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/mujoco_playground_barkour_joystick-standard--8u3fim4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--8u3fim4/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "87e37e120553e61622c94f43349485354422b7da73e93c5c0c69d275ec47b77d",
+   "video_sha256": "d6f123ac9f29e917ffb91ac4faf3a661f7dc1def5ceb66ef1e03061bb9455178",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--a2jyusj.json

@@ -0,0 +1,206 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--a2jyusj",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 524,
+   "sim_steps": 524,
+   "physics_steps": 2620,
+   "steps": 3422,
+   "elapsed_s": 870.0,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": 0.489,
+   "worst_yaw_err": 0.4328,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.21,
+   "peak_processes": 6,
+   "peak_disk_gb": 0.286,
+   "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/mujoco_playground_barkour_joystick-standard--a2jyusj/replay.mp4",
+  "duration_s": 5.28,
+  "fps": 25.0,
+  "frames": 132,
+  "sha256": "a948a57b31621b02291d3d9003249eb318a3d04f0e60406c70fab43b61405a41"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 152,
+      "text_zh": "观察：视频0.00–5.24 s中，机器人四足交替落地迈步并移动；log step 152显示脚本重置到episode=11、resets_left=0，log step 164、166均报告该回合control_steps=524、done=True、reason=fell。此前回合的测试不作为本节点进度。"
+     },
+     {
+      "t": 1.2,
+      "log_step": 163,
+      "text_zh": "观察：视频约1.12–2.88 s中机器人由侧向逐渐转为面对跟随相机，视频3.20–4.96 s继续偏转；未保持前半段近乎固定的朝向。log step 163所载本回合控制步120、180的脚本传感汇总分别为vx=-0.20、-0.21 m/s，目标vx约0.46 m/s，控制步420的wz=-0.80 rad/s，目标约0.02 rad/s。"
+     },
+     {
+      "t": 5.24,
+      "log_step": 164,
+      "text_zh": "推断：log step 9给出100 Hz控制率，log step 164的本回合524步对应仿真5.24 s；已与视频5.24 s末帧和5.28 s片长核对（末帧显示时长约0.04 s）。机器人尚未完成1000步/10秒完整日程就被服务终止；评分器worst_lin_err=0.489 m/s、worst_yaw_err=0.4328 rad/s又分别超过0.2和0.3阈值，因此节点全程未达成，不能把开局站立或部分迈步计为完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–5.24 s中，Barkour的四条腿以足端交替接触棋盘格地面，承重并推动机身移动；可见近侧足端反复落地、抬起。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "value": 0.489,
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better"
+  },
+  {
+   "name": "worst_yaw_err",
+   "value": 0.4328,
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 1.2,
+   "log_step": 163,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约1.2 s起，机器人仍在迈步，却在第一段指令尚未切换时逐渐大幅改变朝向，至视频2.24–2.88 s面对跟随相机，之后继续偏转；log step 163记录本回合前向速度多次与目标反号。该行进步骤没有完成；log step 164显示服务最终于524控制步以fell终止。断点时间取持续跟踪偏离的粗略起点。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 1.2,
+    "log_step": 163,
+    "evidence_zh": "观察：视频1.12–4.96 s中机器人不断迈步，但在前半段持续偏转，从侧向转为正对相机再转向另一侧，没有沿近乎不变的朝向平移。log step 163的目标偏航角速度约0.02 rad/s，而控制步300、420的脚本传感汇总分别为-0.67、-0.80 rad/s；这是行进方向和转向跟踪偏离。视频4.08–5.24 s逐帧查看未见明确侧翻，故不凭fell字段另加平衡现象。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 1.2,
+    "log_step": 150,
+    "source_zh": "日志 / 视频 / 评分器字段",
+    "evidence_zh": "推断：log step 150中agent编写对角小跑关节位置控制脚本，设置0.24秒步态周期、12步提交块，并读取IMU线速度、角速度和重力反馈；log step 151–152启动该脚本并重置为最终第11回合。视频1.12–4.96 s的持续转向偏离与log step 163中多次反向的前向速度、负偏航速度相符，评分器两项速度误差也均超限，支持该控制器未实现所要求的速度跟踪。没有足够证据把问题进一步定位到逆解、接触布局或某个特定参数。"
+   }
+  ],
+  "gap_zh": "观察：inputs中没有状态日志；run.json中log step 150的控制脚本中段已被截去，且最终服务仅报告reason=fell，评分器仅给出fell=1，未区分躯干翻倒、实际关节越界或躯干过低。视频末秒已逐帧检查，末帧仍可见机身在腿上、没有清楚侧翻，无法据此确定触发终止的具体条件，也无法从画面复算速度误差均值；根因只保留控制层级。log step 164的0.4834、0.3921是agent脚本自行汇总，原生指标采用评分器的0.489、0.4328。",
+  "unmet": {
+   "joystick_schedule_followed": false
+  }
+ },
+ "summary_zh": "评分回合为最后一次重置后的第11回合：机器人确实以足端支撑并持续迈步，但视频约1.2 s起明显偏转，未完成“按指令行进”，节点最终0/1。控制器未能实现速度跟踪，两项评分误差均超限，服务在524控制步（对应视频约5.24 s）以fell结束回合。缺少状态日志和具体终止条件，且视频未清楚显示侧翻，因此现象标为步态偏离，根因保留为控制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点joystick_schedule_followed为false、final 0与评分器（success=0, fell=1, 两项误差超限）一致；断点、步态现象和控制根因均有视频与log step 163/164支持，时钟采用控制步524=5.24 s并已核对，未见需改动之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:10:03Z",
+  "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/mujoco_playground_barkour_joystick-standard--a2jyusj/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--a2jyusj/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "546a980e9b1e9ca9962866375d73cd8af5beabf269d6bcf5de0a4b30e8558c8b",
+   "video_sha256": "a948a57b31621b02291d3d9003249eb318a3d04f0e60406c70fab43b61405a41",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--jreayho.json

@@ -0,0 +1,204 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--jreayho",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 117,
+   "sim_steps": 117,
+   "physics_steps": 585,
+   "steps": 2042,
+   "elapsed_s": 997.5,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": 0.2772,
+   "worst_yaw_err": 0.4842,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "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/mujoco_playground_barkour_joystick-standard--jreayho/replay.mp4",
+  "duration_s": 2.36,
+  "fps": 25.0,
+  "frames": 59,
+  "sha256": "af697b08fed010cc4c72a5f16b8e5468ec8edb28cd8b5f66fcbbc0d1b28c1bd2"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.6,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–1.20 s中，Barkour四足机器人在棋盘格地面上反复迈腿并移动，足端接触地面，已开始行进尝试；开局站立不满足完整日程验收。"
+     },
+     {
+      "t": 2.32,
+      "log_step": 102,
+      "text_zh": "观察：视频1.20–2.32 s逐帧可见机器人腿部折低、躯干下沉，最后仍保持近水平朝向，未出现参考成功视频后半段的持续转向过程。log step 102记录最终回合control_steps=117、sim_time_s=1.17、done=true、reason=fell；log step 104与106确认该回合为episode 11且无剩余重置，故此前测试回合不计入节点。"
+     },
+     {
+      "t": null,
+      "log_step": null,
+      "text_zh": "观察：inputs/run.json的评分器字段n_actions=117、fell=1、worst_lin_err=0.2772 m/s、worst_yaw_err=0.4842 rad/s，live_matches_replay=1。完整验收要求1000个控制步以及两项误差分别不大于0.2 m/s、0.3 rad/s，本次没有完成完整日程。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–2.32 s，机器人四条腿的足端先后落地支撑并推动机身移动；末段腿部折低、机身下沉。画面未足以确认躯干本身触地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "value": 0.2772,
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better"
+  },
+  {
+   "name": "worst_yaw_err",
+   "value": 0.4842,
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 2.16,
+   "log_step": 102,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约2.00–2.32 s，正在迈腿行进的机器人腿部进一步折低，躯干靠近地面，未恢复正常支起高度即结束；log step 102的最终回合由服务在control step 117以reason=fell终止，未进入第二段指令。断点锚点取视频2.16 s的下沉过程，不将日志中的仿真1.17 s当作视频时间。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 2.16,
+    "log_step": 102,
+    "evidence_zh": "观察：视频1.92–2.32 s中机器人腿部屈折、躯干持续下沉，末帧机身明显低伏，未能维持由腿支起的行进姿态；log step 102记录服务将本回合终止为fell。画面支持支撑塌低，未显示机身翻转超过90度。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 54、76、84、92、96及100记录agent自编的足端轨迹、姿态反馈和关节位置控制器；log step 102以最后一组步态参数执行最终回合，打印的前向速度从control step 50的0.32 m/s变为control step 100的0.71 m/s，目标为0.4594 m/s，随后服务以fell终止。结合视频末段的机身下沉，控制器未能同时维持行进跟踪和机身支撑；证据不足以确定具体增益、逆解或足端承载布局哪一项导致失稳。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "未提供逐步状态日志或最终回合的躯干世界高度、全部实际关节角及首次终止谓词；fell同时覆盖翻倒、关节越界和躯干过低，不能仅凭此字段认定具体触发项。log step 102打印的final up_z实际来自重力方向分量，不能直接当作机身向上轴；agent在log step 116关于高度规则或关节规则的解释属于待核实推测。视频总长2.36 s，而最终回合服务记录仿真1.17 s，未建立精确的逐步时钟映射，所有视频锚点均直接取自已检查的帧；评分器仅给出误差汇总，不能补出未运行的第二段误差。"
+ },
+ "summary_zh": "最终被评分的是第11回合，机器人已迈腿移动，但视频末段腿部折低、机身下沉，服务在control step 117以fell终止，唯一节点“按指令行进”始终未完成（0/1）。评分器两项速度误差均超过阈值，控制器未能维持稳定支撑与跟踪；缺少逐步状态，无法确定fell具体由高度、关节边界还是翻倒规则触发，且视频时钟与仿真时钟不能直接等同。",
+ "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、日志一致：第11回合在control step 117以fell终止，唯一节点未达成；断点、平衡现象与控制根因均有证据且对不确定处已在gap_zh中说明，未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:02:24Z",
+  "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/mujoco_playground_barkour_joystick-standard--jreayho/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--jreayho/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c324f31672fa65de4ca846ea593393c0897295bba04d457f37399366e2eb8555",
+   "video_sha256": "af697b08fed010cc4c72a5f16b8e5468ec8edb28cd8b5f66fcbbc0d1b28c1bd2",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--mbzvhxu.json

@@ -0,0 +1,209 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--mbzvhxu",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "physics_steps": 5000,
+   "steps": 6025,
+   "elapsed_s": 1067.9,
+   "video_rendered": 1,
+   "fell": 0,
+   "worst_lin_err": 0.3604,
+   "worst_yaw_err": 0.7504,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "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/mujoco_playground_barkour_joystick-standard--mbzvhxu/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "ac6bbcd9d5603357db8ea2dbf676c516b9997cb367a734fc2cc747755afe868a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 47,
+      "text_zh": "观察：log step 47 的机器人服务返回 reset -> 11 0，随后运行最终回合；log step 48 确认 episode=11、control_steps=1000、reason=horizon。视频从站立开始，连续行进至结尾，没有中途重置；前面调试回合的摔倒不属于本次评分轨迹。"
+     },
+     {
+      "t": 5.5,
+      "log_step": 47,
+      "text_zh": "观察：视频 0.0–5.0 s 机器人在棋盘格地面上小幅迈步平移；视频 5.5–9.0 s 继续移动，黄色头部从画面右侧逐渐转向相机，机身由腿支起。log step 7 的服务指令表要求仿真 5.01 s 后偏航角速度为 +0.6428 rad/s；log step 47 最终回合在仿真 6、7、8、9 s 打印的服务 IMU 陀螺仪偏航速度分别为 -0.35、-0.34、-0.28、-0.21 rad/s，均为反向。"
+     },
+     {
+      "t": 10.143,
+      "log_step": 48,
+      "text_zh": "观察：逐帧查看视频 9.0–10.143 s，机器人仍由足端支撑，末段迈步减弱，未见翻倒；log step 48 服务终态为完成 1000 个控制步并因 horizon 结束。评分器记录 n_actions=1000、fell=0、worst_lin_err=0.3604 m/s、worst_yaw_err=0.7504 rad/s，后两项分别超过 0.2 m/s 和 0.3 rad/s 阈值。"
+     },
+     {
+      "t": 10.143,
+      "log_step": 33,
+      "text_zh": "推断：该节点要求完整日程结束且两项跟踪误差均合格，因此本回合从未达到节点，不能因走满日程而记为完成。agent 在 log step 49 自述仅偏航失败，但其 log step 33 的 score 函数对平移误差采用分量绝对值均值，并仅使用每 25 个控制步采样的结果，区别于评分器的逐步平面误差向量模长均值；log step 47 的自算数值不能替代评分器记录。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–10.143 s，Barkour 四条腿的足端在棋盘格地面交替落足，支撑躯干并使其平移、转向；末段趋于站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.3604
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.7504
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 5.5,
+   "log_step": 47,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 5.5–9.0 s，机器人继续小步移动，黄色头部逐渐转向相机，未见翻倒；log step 47 最终回合的仿真 6–9 s IMU 偏航速度持续为负，而 log step 7 服务给出的第二段目标为 +0.6428 rad/s。推断：换指令后的反向转动是未修复的行进跟踪失败的明显表现，取视频 5.5 s 为导航锚点；该回合最终走满 1000 个控制步，但评分器两项误差均超阈值，节点未完成。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 5.5,
+    "log_step": 47,
+    "evidence_zh": "观察：视频 5.5–9.0 s，机器人虽保持足端支撑并移动，黄色头部却从画面右侧逐渐转向相机；末段视频 9.0–10.143 s 未见翻倒。log step 47 的最终回合服务 IMU 读数在仿真 6–9 s 给出负偏航速度，与服务指令表（log step 7）的正偏航目标相反。推断：行进中的方向跟踪偏离，属于未倒地的步态失败。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent 的行进控制器未能把正确的速度目标实现为机体运动。log step 33 的脚本包含正确指令表、速度及偏航反馈，并通过足端轨迹和逆解发送关节位置；log step 44、46 的修改在最终回合恢复对角迈步方案，log step 47 确认执行至 1000 个控制步，但该回合 IMU 偏航速度在仿真 6–9 s 仍持续为负，评分器记录平移和偏航误差均超标。证据支持控制实现不足，尚不能区分足端接触设计、轨迹生成或反馈参数的具体责任。"
+   }
+  ],
+  "gap_zh": "没有状态日志或逐控制步的完整传感记录，评分器也未给出各段的精确均值；不能从跟随相机与头部相机复算误差，或确定首次偏离的精确时刻及具体控制机制。agent 自算平移误差与评分器采用不同统计方法，节点和原生指标以评分器结果为准；不据 agent 关于足端打滑的自述进一步细分根因。"
+ },
+ "summary_zh": "最终评分的第 11 回合走满 1000 个控制步，视频至 10.143 s 仍由腿支撑，评分器 fell=0，但“按指令行进”节点未完成（0/1）。换指令后出现反向偏航，评分器平移误差 0.3604 m/s、偏航误差 0.7504 rad/s 均超阈值，标为全身·步态，根因归于控制。缺少逐步状态及分段精确评分，不能细分控制失效机制；agent 自算平移合格的结论未采用。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未完成（评分器 worst_lin_err 0.3604、worst_yaw_err 0.7504 超阈值，fell=0，1000 步 horizon）、断点、步态现象和控制根因均有 log 与评分器字段支持，视频时间在范围内，未发现需改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:03:34Z",
+  "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/mujoco_playground_barkour_joystick-standard--mbzvhxu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--mbzvhxu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "71d2809db674d12b37c3631b1b8ec82f8c97a105435805f9f408de5e16038dbc",
+   "video_sha256": "ac6bbcd9d5603357db8ea2dbf676c516b9997cb367a734fc2cc747755afe868a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--nksvqsy.json

@@ -0,0 +1,205 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--nksvqsy",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "infra_error": 0,
+   "n_actions": 1000,
+   "sim_steps": 1000,
+   "physics_steps": 5000,
+   "steps": 1832,
+   "elapsed_s": 1695.8,
+   "video_rendered": 1,
+   "fell": 0,
+   "worst_lin_err": 0.5119,
+   "worst_yaw_err": 0.6466,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "instance_mismatch": 0,
+   "no_verdict": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 4,
+   "reset_budget": 10,
+   "episodes": 5,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.321,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.298,
+   "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/mujoco_playground_barkour_joystick-standard--nksvqsy/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "91db0fa4bbbd0a820e3410e3aef654517181e1cf65ad9fb3574f9f3387f2dee6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.5,
+      "log_step": 45,
+      "text_zh": "观察：视频 0.00–1.43 s 的逐帧画面显示，Barkour四足机器人由足端支起躯干，仅有小幅姿态变化，没有启动交替迈步；log step 45 的脚本重置到第5回合后，重复同一组站姿关节目标600个控制步。"
+     },
+     {
+      "t": 5.0,
+      "log_step": null,
+      "text_zh": "观察：视频全程按约0.5 s抽帧，并逐帧检查视频 4.64–5.71 s，机器人仍在原地站立；后半段未出现持续侧移或明显转向。成功参考视频约 5–9.5 s 可见连续迈步和机身朝向明显改变，该参考仅用于辨认视觉线索。"
+     },
+     {
+      "t": 10.14,
+      "log_step": 46,
+      "text_zh": "观察：视频 9.00–10.14 s 的逐帧画面仍显示机器人站立，未见跌倒或行进；log step 46 的脚本继续重复站姿目标400步，服务返回第5回合 control_steps=1000、done=True、reason=horizon。评分器记录 fell=0、worst_lin_err=0.5119 m/s、worst_yaw_err=0.6466 rad/s，分别超过0.2 m/s及0.3 rad/s的阈值，完整日程的行进验收未通过。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–10.14 s 中，Barkour足端接触棋盘格地面，四腿支起躯干；实际仅保持站姿并小幅调整，没有形成行进所需的连续交替落足。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "value": 0.5119,
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better"
+  },
+  {
+   "name": "worst_yaw_err",
+   "value": 0.6466,
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 0.5,
+   "log_step": 45,
+   "where": "首步前",
+   "evidence_zh": "观察：视频 0.00–1.43 s 的逐帧画面及后续视频至 10.14 s 均未见持续迈步，机器人只保持站姿；log step 45–46 的第5回合脚本仅重复固定关节目标，服务最终报告1000步到达horizon。评分器两项速度误差超标；先前测试回合已被重置，不计为本次评分轨迹的跌倒。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.5,
+    "log_step": 45,
+    "evidence_zh": "观察：视频 0.00–10.14 s 中机器人始终由四腿支撑站立，未启动按指令平移和转向的行进；log step 45–46 的最终回合只发送600步及400步重复站姿关节目标。因此本次评分回合没有尝试行进，不能将此前测试回合的失稳当成本回合现象。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 39,
+    "evidence_zh": "推断：log step 39 的步态测试中，agent脚本记录前向速度仅约0.11–0.12 m/s，低于指令约0.459 m/s，偏航速度��约−0.94变到+0.87 rad/s，服务在第2回合293个控制步返回reason=fell；log step 42、44 的调参测试也分别在246及108个控制步返回reason=fell。随后log step 45–46 改为仅维持站姿，视频确认最终没有行进；这些会话测试支持未能实现可用行进控制这一根因，但不能证明牵引、力矩或某个特定参数有问题。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有状态日志；最终回合仅有站姿命令、服务终态和评分器汇总指标，无法定位此前步态控制失败的具体动力学机制，故根因保留为“控制”，不细分子类。此前测试回合的reason=fell不是本次评分视频中的跌倒证据。"
+ },
+ "summary_zh": "评分第5回合中机器人保持站立并运行满1000步，但未启动按指令行进，节点最终为0/1；评分器记录平面速度误差0.5119 m/s、偏航角速度误差0.6466 rad/s，均超阈值，fell=0。断点在行进首步前，现象为没尝试；此前步态测试反复失败后，最终回合只维持站姿，根因归为控制。没有状态日志，无法进一步确定具体控制失效机制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点未达、断点首步前/没尝试、根因控制（来自会话测试证据，依 FAIL-22 做法）均与视频、log 及评分器字段一致；视频 0–10.2 s 全程站立。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:22:12Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--nksvqsy/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--nksvqsy/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9f6cbaad96fa7c2058010fb5cf3db93bba7f7149655dbafaf7654575a65409c4",
+   "video_sha256": "91db0fa4bbbd0a820e3410e3aef654517181e1cf65ad9fb3574f9f3387f2dee6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--p3ejwxv.json

@@ -0,0 +1,189 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--p3ejwxv",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 365,
+   "sim_steps": 365,
+   "physics_steps": 1825,
+   "steps": 494,
+   "elapsed_s": 1638.1,
+   "video_rendered": 1,
+   "fell": 0,
+   "worst_lin_err": 0.2614,
+   "worst_yaw_err": 0.2735,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 1,
+   "reset_budget": 10,
+   "episodes": 2,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.179,
+   "peak_processes": 6,
+   "peak_disk_gb": 0.26,
+   "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/mujoco_playground_barkour_joystick-standard--p3ejwxv/replay.mp4",
+  "duration_s": 4.88,
+  "fps": 25.0,
+  "frames": 122,
+  "sha256": "3eb954a851c5c2fad47064697476e70809d026625c3d1cf50e5efffe5e6237d7"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.4,
+      "log_step": null,
+      "text_zh": "观察：已检查本次视频全部122帧（视频0.00–4.84 s）；机器人从站立、轻微调整发展为迈步移动，视频2.4–4.84 s腿部交替运动，末段机身明显倾斜，最后仍由腿支起，未见翻倒。画面没有呈现完整日程结束。"
+     },
+     {
+      "t": null,
+      "log_step": 57,
+      "text_zh": "观察：log step 42重置为episode 2；log step 54、55、57连续推进该回合，log step 57最后读数为control_steps=365、done=False。评分器reward记录n_actions=365、sim_steps=365、fell=0、live_matches_replay=1、success=0；这与DAG要求完整1000个控制步不符，不能把先前episode 1的终止或本地仿真测试计入评分回合。"
+     },
+     {
+      "t": null,
+      "log_step": 59,
+      "text_zh": "推断：结合最后回合仅365/1000控制步及视频未显示完整日程，按指令行进节点始终未达到，final=false；log step 59的模型提供方过滤器中断发生在服务仍未结束回合之后。控制步计数只用于确认日程未完成，未将仿真365×0.01=3.65 s直接当成视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频2.4–4.84 s，Barkour四条腿的足端交替落在棋盘格地面，支撑躯干并随迈步移动；末段倾斜时仍有足端支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.2614
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.2735
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": null,
+   "log_step": 59,
+   "where": "操作中",
+   "evidence_zh": "观察：机器人已在视频中迈步，最后检查到的视频4.84 s仍未翻倒；log step 57现场回合停留在控制步365且done=False，log step 58报告尚余185 s墙钟预算，log step 59随即出现“This request was blocked by Gemini's filters”并结束日志。推断：中断时正在进行按指令行进，未完成完整日程；过滤器触发无法定位到某一视频帧，故t=null。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [],
+  "root_causes": [],
+  "gap_zh": "观察：inputs/有本次与参考视频及完整run.json，但没有状态日志；log step 59记载Gemini提供方过滤器中断，之前log step 57仍为done=False，评分器fell=0，未见评分回合已经不可逆判负。推断：这是结果尚未固定时的外因截断，应留待重测，不纳入智能体失败归因，现象和根因均留空。评分器worst_lin_err=0.2614 m/s高于0.2 m/s，worst_yaw_err=0.2735 rad/s低于0.3 rad/s，但这些是未走完整日程的记录，不能据此断言继续执行后仍会失败。视频有末段倾斜，未见翻倒；无法从画面确认实际关节角、躯干高度及逐步速度误差，也未建立控制步到视频钟的精确映射。"
+ },
+ "summary_zh": "评分回合中机器人已迈步移动，视频末段明显倾斜但仍由腿支撑；现场回合只执行365/1000个控制步，按指令行进未完成，最终节点0/1。log step 57仍为done=False，随后log step 59出现Gemini提供方过滤器中断，属于结果尚未固定时的外因截断，保留目录success=0而不标智能体失败现象或根因。缺少状态日志，未建立控制步到视频钟的精确映射，精确速度及终止判据不能由画面补全。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标签合理：log step 57 现场回合 control_steps=365、done=False，log step 58 仍余185 s，log step 59 为Gemini提供方过滤器中断，结果未固定，属外因截断；节点false、现象与根因留空并写gap_zh，符合FAIL-5/D37。视频无翻倒，时间均未把智能体时间换算成视频时间。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:04:04Z",
+  "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/mujoco_playground_barkour_joystick-standard--p3ejwxv/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--p3ejwxv/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f1453e9f7c850f1603605071841bd386cf6ef73a78e272bbec68ae785cc50b08",
+   "video_sha256": "3eb954a851c5c2fad47064697476e70809d026625c3d1cf50e5efffe5e6237d7",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--shxsrs8.json

@@ -0,0 +1,204 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--shxsrs8",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 99,
+   "sim_steps": 99,
+   "physics_steps": 495,
+   "steps": 1466,
+   "elapsed_s": 527.2,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": 0.7971,
+   "worst_yaw_err": 0.3627,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.291,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.239,
+   "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/mujoco_playground_barkour_joystick-standard--shxsrs8/replay.mp4",
+  "duration_s": 2.0,
+  "fps": 25.0,
+  "frames": 50,
+  "sha256": "9dac77ada6b9ee1e6913c7527b4527df317720a6d860e063533eb7e0a573afa1"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 机器人由足端支起，随后腿部动作并伴随机身移动；视频 1.40–1.96 s 机身逐渐前倾，后部抬高，黄色前部降至地面附近，未恢复正常支撑。整段没有完成两段指令的连续行进及后半段转向。"
+     },
+     {
+      "t": 1.96,
+      "log_step": 24,
+      "text_zh": "观察：log step 24 明确开始最后一次重置后的评分回合，服务返回 episode=11、control_steps=99、done=true、reason=fell、resets_left=0；评分器记录 sim_steps=99、fell=1、live_matches_replay=1。该回合未达到所需1000控制步；评分器 worst_lin_err=0.7971 m/s、worst_yaw_err=0.3627 rad/s，分别超过0.2 m/s和0.3 rad/s阈值。"
+     },
+     {
+      "t": 1.8,
+      "log_step": 24,
+      "text_zh": "推断：结合视频末段前倾失稳与最后评分回合的服务终止记录，该节点从未完成。视频为2.00 s，而99控制步对应仿真0.99 s；没有已验证的时钟映射，证据锚点直接取所检查的视频画面，不把log step 24打印的仿真时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–1.40 s，Barkour足端与棋盘格地面接触并支撑机身，腿部尝试迈步；视频 1.60–1.96 s，后部抬高、前腿仍接地，机身向前下方倾倒。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.7971
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.3627
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 1.8,
+   "log_step": 24,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.60–1.96 s 机器人行进尝试中机身持续前倾，后部抬起、前部下沉，未恢复支撑；log step 24 的最后评分回合在99控制步后以 reason=fell 终止，未进入后半段命令。断点锚点取视频约1.80 s的明显失稳画面。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.8,
+    "log_step": 24,
+    "evidence_zh": "观察：视频 1.60–1.96 s，Barkour后部逐步抬高，躯干前倾、黄色前部降到地面附近，结束前没有恢复四足支起的正常姿态；log step 24 服务返回 done=true、reason=fell，评分器 fell=1。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：agent的足端轨迹与关节位置控制没有实现所需的速度跟踪和支撑稳定。log step 12 将控制改为每10个控制步批量提交再读取反馈；log step 22–23 又在逆解目标上加入按站立关节偏差估计的承载补偿���log step 24 实际以该控制器运行最后评分回合，脚本读取的速度从仿真约0.49 s的(-0.14,-0.05) m/s变为仿真约0.99 s的(-0.98,-0.03) m/s，与第一段(+0.4594,+0.2212) m/s命令相反，并收到fell终止；视频 1.60–1.96 s可见支撑失稳。现有证据支持控制失效，但不能单独证实某个增益、反馈周期或承载补偿就是原因。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有状态日志，无法确定fell究竟由躯干高度、关节边界还是翻倒判据首次触发，也无法检验agent在log step 25提出的软执行器下沉、摆动足拖地解释。视频2.00 s与99控制步对应的仿真0.99 s不同，缺少已核实的时钟映射，所有视频锚点均直接来自逐帧检查；根因保留在控制大类。"
+ },
+ "summary_zh": "评分回合尝试迈步但未完成“按指令行进”，最终0/1节点；视频约1.80 s出现明显前倾失稳，log step 24确认最后回合在99控制步后以fell终止。评分器平移和偏航误差均超阈值，日志与画面支持控制失败，但不能确定具体控制子因。没有状态日志，且视频2.00 s与该回合仿真0.99 s的映射未经核实，失败锚点采用视频画面。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点未达成、断点取该节点、平衡现象与控制大类根因均有视频和评分器/服务读数支持；时钟处理谨慎，未把仿真步数换算成视频时间。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:04:14Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--shxsrs8/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--shxsrs8/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8f01f69dcb22e73c081c1a81759b66d0c170d355fc140cf4e6d385b425f9d1ad",
+   "video_sha256": "9dac77ada6b9ee1e6913c7527b4527df317720a6d860e063533eb7e0a573afa1",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard--v5f2nhs.json

@@ -0,0 +1,209 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard--v5f2nhs",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 123,
+   "sim_steps": 123,
+   "physics_steps": 615,
+   "steps": 2944,
+   "elapsed_s": 851.8,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": 0.4717,
+   "worst_yaw_err": 0.501,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "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/mujoco_playground_barkour_joystick-standard--v5f2nhs/replay.mp4",
+  "duration_s": 2.48,
+  "fps": 25.0,
+  "frames": 62,
+  "sha256": "17ca1c88f853ff082f5bf7c7c100343afc7cdfe20b98dfe239f6f258bc2d055b"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–1.44 s，Barkour四足机器人由足端支撑在棋盘格地面上，机身先下沉并有小幅腿部运动；没有完成持续平移、随后换向转弯的整段日程。"
+     },
+     {
+      "t": 1.8,
+      "log_step": null,
+      "text_zh": "观察：视频 1.80–2.44 s，机器人躯干明显倾斜，一端抬高、另一端下沉，多个足端抬离地面；末帧仍未恢复平稳支撑。"
+     },
+     {
+      "t": null,
+      "log_step": 55,
+      "text_zh": "观察：log step 55 的服务重置结果为 episode 11、resets_left 0，agent随后执行控制脚本；此前回合均被重置，不属于本次评分轨迹。"
+     },
+     {
+      "t": null,
+      "log_step": null,
+      "text_zh": "推断：inputs/run.json 的评分器记录 sim_steps=123、fell=1、live_matches_replay=1、success=0；本次评分回合未运行满所需1000个控制步，且 worst_lin_err=0.4717 m/s、worst_yaw_err=0.501 rad/s 均超过节点阈值，因此该完整日程节点始终未达成。123个控制步仅用于判断日程未完成，不转换成视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–1.76 s，Barkour的足端接触棋盘格地面承重，腿部屈伸并尝试迈步；视频 1.80–2.44 s，足端支撑变得不稳定，躯干持续倾斜。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.4717
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.501
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 1.8,
+   "log_step": 56,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 1.80 s 起，正在尝试行进的机器人机身倾斜加剧、部分足端离地，至视频 2.44 s 仍未恢复稳定支撑。log step 55 启动最后的第11回合；log step 56 的同回合脚本输出显示 IMU姿态偏差增长、接触足数减少；评分器最终记录123个控制步及 fell=1，整段行进日程未完成。断点时间直接取视频帧。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.8,
+    "log_step": 56,
+    "evidence_zh": "观察：视频 1.80–2.44 s，机器人躯干由近水平变为明显倾斜，一端下沉，多足抬离地面，无法保持稳定站立。log step 56 的脚本读取到控制步90时足接触为[0,1,0,1]，控制步110时为[0,0,0,1]；评分器记录 fell=1。画面可见失衡，但未显示完整翻转至倒扣。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 55 ���录agent在最后回合运行调整后的步态控制脚本；log step 56 的脚本传感读数中，指令约为[0.46,0.22,0.02]，控制步80的平面速度约[-0.01,-0.04]，控制步110约[0.02,-0.09]，IMU重力方向的水平分量由控制步80的[0.10,-0.02]增至控制步110的[0.42,0.28]，足接触由四足减至单足。结合视频 1.80–2.44 s 的持续倾斜，所执行控制未能同时建立所需运动并稳定机身；证据支持行进控制失稳，但不足以指定控制律或参数子类。"
+   }
+  ],
+  "gap_zh": "无逐帧状态日志；评分器的 fell 汇总字段未区分躯干翻倒、实际关节越界和躯干过低，视频也不能量出这些精确阈值，因此不声称触发了哪一条底层终止条件。日志中的控制脚本修改仅保留不完整的调用文本，无法核实具体控制律、参数或推扰发生时刻；agent在 log step 58 自述“第一次推扰中跌倒”不能单独证明推扰是原因。本视频时长2.48秒与123个控制步对应的仿真1.23秒不同，未提供明确播放映射，所有视频锚点均直接读帧。"
+ },
+ "summary_zh": "评分的第11回合只完成起始支撑及短暂腿部运动，“按指令行进”未达成，最终0/1；此前跑满日程的测试已被重置。机器人在视频约 1.80 s 起失衡且未恢复，评分器记录123个控制步、fell=1，两项速度误差也超阈值，日志支持行进控制失稳。缺少逐帧状态及完整控制脚本，无法确定具体终止判据或控制参数原因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点取视频约1.8 s失衡、平衡现象与控制根因均有帧和评分器字段(fell=1、123步、两项误差超阈值)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": "2026-10-30T03:22:32Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--v5f2nhs/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard--v5f2nhs/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "d226dd10381509e08149f6f7eec39875fcd75052962513477efe90c50b1a18b1",
+   "video_sha256": "17ca1c88f853ff082f5bf7c7c100343afc7cdfe20b98dfe239f6f258bc2d055b",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard-realtime--3jnmac4.json

@@ -0,0 +1,187 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard-realtime--3jnmac4",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "n_actions": 0,
+   "sim_steps": 0,
+   "calls": 13,
+   "rows": 0,
+   "resets": 0,
+   "reset_budget": 10,
+   "sim_startups": 1,
+   "elapsed_s": 261.0,
+   "video_rendered": 1,
+   "realtime": 1,
+   "slowdown": 1.0,
+   "clock_s": 0.0,
+   "max_lag_s": 0.0,
+   "stalls": 0,
+   "stalled_s": 0.0,
+   "stalled_share": 0.0,
+   "held_steps": 0,
+   "late_steps": 0,
+   "requests_refused": 0,
+   "request_busy_s": 0.009,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.175,
+   "peak_processes": 6,
+   "peak_disk_gb": 0.249,
+   "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/mujoco_playground_barkour_joystick-standard-realtime--3jnmac4/replay.mp4",
+  "duration_s": 0.04,
+  "fps": 25.0,
+  "frames": 1,
+  "sha256": "db7c3bad0cdaa29b631ba80a4dc0951a25e850225f7b2dea17374a001258cc6c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：本次视频全长0.04 s、只有视频0.00 s的一帧；左侧跟随视角显示Barkour处于初始站姿，右侧头部视角显示棋盘格地面，没有可供观察的迈步、平移或转向过程。"
+     },
+     {
+      "t": null,
+      "log_step": 14,
+      "text_zh": "观察：log step 14中机器人服务返回running: False；该次调用为get_obs(seconds=0.0)。log step 3的服务状态为control_steps=0、planned_steps=0、running=false；评分器最终记录sim_steps=0、n_actions=0、rows=0、clock_s=0.0。"
+     },
+     {
+      "t": null,
+      "log_step": 15,
+      "text_zh": "推断：log step 15记录Gemini过滤器拦截，此前全部机器人请求仅查询状态或零时长观测，未提交动作；结合视频0.00 s初始画面及评分器sim_steps=0，可确定完整1000控制步指令日程未执行，节点从未达到。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "不适用",
+      "how": "不适用",
+      "note_zh": "观察：评分器记录n_actions=0、rows=0、sim_steps=0，log step 14仍为running: False；仅有视频0.00 s初始画面，没有交付行进轨迹，按无交付规则标不适用。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "unknown",
+   "t": null,
+   "log_step": 15,
+   "where": "首步前",
+   "evidence_zh": "观察：log step 15的最后消息为“This request was blocked by Gemini's filters”，未附动作调用；log step 14的机器人服务仍返回running: False，评分器记录n_actions=0、sim_steps=0。推断：模型提供方过滤器在回合启动和轨迹交付之前截断会话，无法定位机器人执行中的失败步骤，按外部中断规则标unknown／首步前。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [],
+  "root_causes": [],
+  "gap_zh": "观察：log step 15记录Gemini提供方过滤器拦截，log step 14仍未启动回合；inputs/replay.mp4只有视频0.00 s一帧，无状态日志，评分器未报告worst_lin_err和worst_yaw_err。推断：没有已执行的行进尝试可供判定物理失败现象或机器人根因，按CASES.md FAIL-5作为外部中断留空两层，不能标为没尝试或控制失败；目录中的success=0保留为原始评分结果，不将本次纳入代理失败归因。"
+ },
+ "summary_zh": "观察：本次仅有视频0.00 s初始站姿，评分器记录零动作、零仿真步，“按指令行进”未达到，最终节点0/1。观察：log step 14的服务仍未运行，log step 15随即记录Gemini过滤器拦截；推断：这是首步前的外部中断，断点为unknown，现象与根因留空，不作机器人失败归因。观察：无状态日志且两项速��误差均未报告，无法评价行进控制质量。",
+ "conf": 10,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 9,
+    "summary_zh": "标注符合FAIL-5：Gemini过滤器在未提交任何动作时截断会话，节点未达到、断点unknown/首步前、两层为空并附gap_zh，视频仅一帧初始站姿，与评分器n_actions=0、sim_steps=0一致。",
+    "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": "2026-10-30T03:22:37Z",
+  "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/mujoco_playground_barkour_joystick-standard-realtime--3jnmac4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--3jnmac4/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "3318f2942f87e21b02ecd16e53f96d978f2a1bc048a07cb78e5801f127e0a3da",
+   "video_sha256": "db7c3bad0cdaa29b631ba80a4dc0951a25e850225f7b2dea17374a001258cc6c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard-realtime--5jjbxvh.json

@@ -0,0 +1,225 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard-realtime--5jjbxvh",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "n_actions": 578,
+   "sim_steps": 7053,
+   "calls": 9544,
+   "rows": 57622,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_startups": 1,
+   "elapsed_s": 439.5,
+   "video_rendered": 1,
+   "realtime": 1,
+   "slowdown": 1.0,
+   "clock_s": 70.44,
+   "max_lag_s": 0.0035,
+   "stalls": 0,
+   "stalled_s": 0.0,
+   "stalled_share": 0.0,
+   "held_steps": 0,
+   "late_steps": 5,
+   "requests_refused": 10,
+   "request_busy_s": 3.981,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.148,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.24,
+   "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/mujoco_playground_barkour_joystick-standard-realtime--5jjbxvh/replay.mp4",
+  "duration_s": 5.8,
+  "fps": 25.0,
+  "frames": 145,
+  "sha256": "71d598d6c7b3b7e59d8116cd6aa6c0cb3cfc4b40cc828cc086982dcaa40ea844"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.8,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–5.0 s，Barkour四足机器人交替抬腿、落足并在棋盘格地面移动；视频 2.0 s 机身近侧向，视频 2.8–3.0 s 转成较正对跟随相机，视频 4.0 s 又回到近侧向，行进朝向明显摆动。"
+     },
+     {
+      "t": 5.76,
+      "log_step": 27,
+      "text_zh": "观察：视频 4.80–5.76 s 逐帧仍可见机器人迈步及腿部支撑，录像随即结束，未展示完整10秒日程；log step 27 中仿真服务状态为episode=11、control_steps=578、done=True、reason=failed，且与评分器n_actions=578一致。"
+     },
+     {
+      "t": 5.76,
+      "log_step": 27,
+      "text_zh": "推断：log step 2 的服务规格给出 control_dt=0.01 s，log step 27 最后第11回合的578个控制步对应仿真5.78 s，与 inputs/video.json 的视频总长5.8 s 仅在时长上吻合；视频中看不到可辨认的指令切换事件，因此没有独立可见事件核对控制步到视频时间的映射，不据此给出精确的视频换指令时刻。该回合未运行满1000步，故完整日程节点从未成立。成功参考视频后半程持续侧移、转向并运行到参考视频约10秒，仅用来比较完成外观。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：本次视频 0.0–5.76 s，Barkour的四条腿足端交替落在棋盘格地面上，承重并推进躯干；视频末帧仍可见腿部接地支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 5.76,
+   "log_step": 27,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 27 中机器人服务的 final 状态给出 episode=11、control_steps=578、done=True、reason=failed，与评分器 n_actions=578 一致；视频5.0–5.76 s 仍可见机器人迈步且机身由腿部支起，随后录像结束，未见躯干翻倒。推断：以已检查的视频末帧5.76 s 作为“按指令行进”失败终止附近的粗略导航锚点，不再用视频2.8 s 的朝向摆动定位终止；按服务 control_dt=0.01 s，终止为仿真5.78 s，较指令日程在仿真5.01 s 的切换晚约0.77 s，但视频时间映射仅有时长吻合，具体终止触发条件无法确定。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 5.76,
+    "log_step": 27,
+    "evidence_zh": "观察：视频5.0–5.76 s，Barkour仍迈步并由腿部支起，录像在完整日程完成前结束；log step 27 的服务状态确认第11回合在578个控制步以 reason=failed 终止。视频2.0 s 机器人近侧向，视频2.8–3.0 s 转为较正对跟随相机，视频4.0 s 又转回���这是较早的方向摆动观察，不作为终止时刻。推断：按可见行进及未完成整段日程标为全身·步态；现有证据不能将其具体化为翻倒或失去支撑。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "反馈控制",
+    "log_step": 27,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 5 的控制器代码以周期腿部目标叠加机体线速度和偏航角速度反馈，log step 17 增加关节跟踪补偿；log step 26–27 修改并启用横滚补偿，log step 27 的最终参数相较 log step 21 降低了部分反馈增益。log step 28 中 agent 自述这些最终改动未经测试，此自述只说明其调试判断，不证明具体参数错误。视频2.0–4.0 s 可见行进朝向摆动；log step 27 服务记录的最终回合在578步失败，按控制周期和指令日程推得终止在仿真5.01 s 换指令后约0.77 s。结合控制器结构、可见摆动与换指令后的提前终止，将反馈控制未稳定完成跟踪作为根因推断；不采用 agent 脚本的误差统计，也不能确定某个增益或横滚补偿符号导致了终止。"
+   }
+  ],
+  "gap_zh": "观察：inputs/没有状态日志；run.json 的 reward 未报告 worst_lin_err、worst_yaw_err 或具体终止触发条件，故两项 native_metrics 值保留 null，不能确定是否触发关节越界、躯干过低或翻倒阈值。agent 在 log step 28 自述“倒地”，但视频最后一秒至视频5.76 s 仍可见机身由腿部支起，画面不足以确认摔倒，不采用该自述作为平衡现象。推断：578步按控制周期对应仿真5.78 s，与视频总长5.8 s 仅在时长上吻合；视频没有可辨认的指令切换事件，故没有独立事件验证其时间映射。反馈控制归因有控制器代码和可见摆动支持，但换指令后的具体终止机制、各参数和推扰的影响均无法隔离。",
+  "unmet": {
+   "joystick_schedule_followed": false
+  }
+ },
+ "summary_zh": "评分采用最后第11回合：机器人以足端支撑迈步移动，但未完成完整“按指令行进”日程，节点最终为0/1；断点粗锚在视频5.76 s 末段，服务记录578步后以 failed 终止，按控制周期及日程推得终止在仿真5.01 s 换指令后约0.77 s。视频2.0–4.0 s 的朝向摆动只作为次要观察，结合控制器代码推断反馈跟踪未稳，不能确认具体不当参数。因无状态日志、原生误差和具体终止字段，视频未证实倒地且时钟映射仅有时长吻合，关键时刻对应关系及根因仍含推断，置信度降为5。",
+ "conf": 5,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点值（未达成）、contact_actual 和根因方向基本可接受；但断点与现象锚在视频 2.8 s 的朝向摆动，而服务读数显示回合在换指令（约 5.01 s）后约 0.77 s、视频约 5.76 s 以 reason=failed 终止，这才是任务真正失败的位置，需要改锚点并澄清证据。",
+    "n_issues": 3,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "第二轮修改后节点未达成、断点锚在服务读数的 578 步终止（视频约 5.76 s）、时钟映射标为推断、根因证据不再依赖脚本统计，均与视频和 log step 27 一致；未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:26:20Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--5jjbxvh/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--5jjbxvh/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--5jjbxvh/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--5jjbxvh/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "96e9a3e8b4e964673dcceb3c9d893bc14b694cb5f4c52c702bfa9ef66f571553",
+   "video_sha256": "71d598d6c7b3b7e59d8116cd6aa6c0cb3cfc4b40cc828cc086982dcaa40ea844",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard-realtime--bjj83yu.json

@@ -0,0 +1,202 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard-realtime--bjj83yu",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "n_actions": 127,
+   "sim_steps": 7188,
+   "calls": 3715,
+   "rows": 3693,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_startups": 1,
+   "elapsed_s": 468.2,
+   "video_rendered": 1,
+   "realtime": 1,
+   "slowdown": 1.0,
+   "clock_s": 71.808,
+   "max_lag_s": 0.0027,
+   "stalls": 0,
+   "stalled_s": 0.0,
+   "stalled_share": 0.0,
+   "held_steps": 3510,
+   "late_steps": 0,
+   "requests_refused": 4,
+   "request_busy_s": 1.312,
+   "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/mujoco_playground_barkour_joystick-standard-realtime--bjj83yu/replay.mp4",
+  "duration_s": 2.56,
+  "fps": 25.0,
+  "frames": 64,
+  "sha256": "f80beb053ccdd7d19cc17222eddca61bb5390e38ab2d9fd8881973a4a2820c30"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.8,
+      "log_step": 55,
+      "text_zh": "观察：视频 0.00–0.80 s，Barkour机器人由足端支撑并交替迈腿，机身随步态起伏；log step 55显示最后一次重置进入episode 11，控制器开始执行本回合。"
+     },
+     {
+      "t": 2.52,
+      "log_step": 56,
+      "text_zh": "观察：视频 2.08–2.52 s，机器人机身持续倾斜、一侧降低，部分腿抬高或外伸，未恢复起始的腿部支撑姿态；log step 56的机器人服务状态为episode 11、control_steps=127、done=True、reason=failed，评分记录n_actions=127。"
+     },
+     {
+      "t": 2.52,
+      "log_step": 56,
+      "text_zh": "推断：最后回合仅执行127个控制步即终止，未完成节点要求的1000步及后半段指令，因此全程没有满足完整日程节点；会话中另一次1000步测试（log step 28–32）不属于这段评分视频。视频锚点按帧号/25取得，没有把控制步或agent工作时间换算成视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–2.00 s，Barkour机器人的足端交替落在棋盘格地面上，支撑机身并迈步；视频 2.08–2.52 s机身倾斜降低，足端支撑未能维持起始姿态。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 2.08,
+   "log_step": 56,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 2.08–2.52 s，正在迈步的Barkour机器人机身逐渐侧倾并降低，末帧仍未恢复；log step 56的机器人服务确认最后回合episode 11仅到control_steps=127便以failed终止，未进入第二段指令，更未完成完整行进日程。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 2.08,
+    "log_step": 56,
+    "evidence_zh": "观察：视频 2.08–2.52 s，机器人机身明显侧倾、靠近地面，一侧腿抬起，未重新建立起始的支撑姿态；log step 56打印的最后回合控制步125重力方向为[-0.483,0.413,-0.772]，随后服务在控制步127判为failed。这里标记可见的失衡，不把它当作已证实超过90度翻倒。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：agent的关节步态控制器未能在跟随首段速度指令时维持机体稳定。log step 55–56的最后回合输出中，指令始终为[0.459,0.221,0.016]，实际前向速度却从控制步75的-0.448变到控制步100的0.263、控制步125的-0.537，重力方向也从[-0.218,0.076,-0.973]发展为[-0.483,0.413,-0.772]；视频 2.08–2.52 s可见机���倾斜降低。日志支持步态控制失稳，但未给出完整控制器代码及逐步目标，不能确定是接触布局、反馈律还是某个参数造成。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志，最后回合的服务只报告failed，未给出触发终止的具体关节、躯干高度或翻倒谓词，故不判定是哪条约束触发。评分字段未报告worst_lin_err与worst_yaw_err，均保留null；会话测试的误差不能替代最后回合的原生指标。控制周期为0.01 s、最后回合127步，而视频长2.56 s，二者没有已核实的逐步映射，视频锚点均直接取自所检查帧。控制根因仅能确定到大类。"
+ },
+ "summary_zh": "观察：评分视频中的Barkour机器人已开始迈步，但视频 2.08–2.52 s机身明显倾斜降低，log step 56确认最后回合只执行127个控制步便失败，“按指令行进”节点最终为0/1。推断：关节步态控制未能维持机体稳定，导致完整指令日程未完成；先前的1000步测试不计入本回合。没有状态日志或明确的终止谓词，无法确认具体触发约束与控制子类，两项原生误差指标也未报告。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标签与证据一致：评分视频（2.56 s）中机器人迈步后机身侧倾降低，log step 56 服务报告 episode 11 在 control_steps=127 以 failed 终止，节点 joystick_schedule_followed 为 false；断点、平衡现象和停在“控制”大类的根因均有证据支持，时钟未做换算。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:24:01Z",
+  "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/mujoco_playground_barkour_joystick-standard-realtime--bjj83yu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--bjj83yu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "3ef3ffd43af5719f9de53a462b388a18682d12662a95be3a7a4180f11984673f",
+   "video_sha256": "f80beb053ccdd7d19cc17222eddca61bb5390e38ab2d9fd8881973a4a2820c30",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard-realtime--hfgptfp.json

@@ -0,0 +1,204 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard-realtime--hfgptfp",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "n_actions": 80,
+   "sim_steps": 3662,
+   "calls": 326,
+   "rows": 2913,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_startups": 1,
+   "elapsed_s": 714.6,
+   "video_rendered": 1,
+   "realtime": 1,
+   "slowdown": 1.0,
+   "clock_s": 36.526,
+   "max_lag_s": 0.0032,
+   "stalls": 0,
+   "stalled_s": 0.0,
+   "stalled_share": 0.0,
+   "held_steps": 2802,
+   "late_steps": 0,
+   "requests_refused": 28552,
+   "request_busy_s": 1.55,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.2,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.273,
+   "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/mujoco_playground_barkour_joystick-standard-realtime--hfgptfp/replay.mp4",
+  "duration_s": 1.64,
+  "fps": 25.0,
+  "frames": 41,
+  "sha256": "a039d7d420a3a745b1e607407a10322d4445bacd34ccc84bad050adda643b558"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s，Barkour四足机器人由足端支起，尚未完成行进日程；完整日程是该节点的验收条件，开局站立不计为到达。"
+     },
+     {
+      "t": 1.2,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–1.60 s（全部41帧），机器人主要在原处调整腿部；视频 0.72、1.20、1.56 s 可见腿部逐渐弯折，机身降低，末段未重新撑高，也未出现完整平移和后半段明显转向。"
+     },
+     {
+      "t": null,
+      "log_step": 45,
+      "text_zh": "观察：log step 44 返回最后一次重置、resets_left=0；log step 45 的机器人服务状态返回 episode=11、control_steps=80、done=True、reason=failed，log step 47 再次确认相同终态。服务状态是现场回合读数；该回合未运行满1000个控制步，不能通过完整日程验收。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–1.60 s，Barkour足端接触棋盘格地面承重，腿关节发生调整并逐渐弯折；末段机身降低。可见的是足端与地面的支撑交互，未据阴影判定躯干触地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 1.2,
+   "log_step": 45,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.72–1.56 s，机器人腿部逐渐弯折、机身降低，未恢复到开局支撑高度；行进与转向日程未完成。log step 45 的现场服务记录最后回合 episode=11 在80个控制步终止为 failed，log step 47 确认未恢复。视频 1.20 s 为可见支撑塌低的粗略导航锚点，未由日志工作时间换算。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.2,
+    "log_step": null,
+    "evidence_zh": "观察：对比视频 0.00、0.72、1.20、1.56 s，机器人承重腿越来越弯，足端附近的腿段接近地面，机身明显降低；末段未重新撑起。推断：机器人未保持足够的腿部承重支撑，属于支撑塌低的平衡失败；画面未显示机身翻转超过90度，不将其描述为侧翻。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": null,
+    "log_step": 42,
+    "evidence_zh": "观察：log step 42–43 的控制脚本通过 joint_position 提交腿关节目标，含外展正反探测及后续步态、姿态修正逻辑；视频 0.72–1.56 s 显示腿部调整后支撑高度持续降低。log step 45 确认最后现场回合仅运行80个控制步即失败；log step 46 的 EpisodeOver 异常明确发生于服务已判 failed 后。推断：该回合的腿部控制未能维持承重姿态，根因停留在控制层；不能把回合结束后的异常当成引发塌低的原因，也不能据此前调试回合或agent最终自述确定某一增益、抬腿方向或反馈项有误。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "观察：inputs/提供了评分视频、完整会话记录及成功参考视频，但没有状态日志；reward未报告worst_lin_err或worst_yaw_err，故两项原生指标均为null。最后现场回合的服务状态只给reason=failed，没有明确是哪条终止条件触发，也没有该回合逐步高度、实际关节角或速度记录。log step 3 的control_dt=0.01 s与log step 45 的80个控制步对应仿真 0.80 s，而评分视频长度为1.64 s；未建立经可见事件核对的两者映射，因此所有视频锚点直接取自帧，未将仿真时间套入视频。log step 46 的异常位于探测提交处，不能确认最终脚本已进入主步态循环；log step 53 所称摆腿俯仰原因仍是agent自述，缺少最后回合数据验证。",
+  "unmet": {
+   "joystick_schedule_followed": false
+  }
+ },
+ "summary_zh": "最后评分回合中，Barkour足端接地并调整腿部，但随后腿部弯折、机身塌低，未完成“按指令行进”，节点最终0/1。断点在该步骤操作中，现象为平衡失败，根因仅支持宽泛的控制问题；服务在log step 45确认最后回合80/1000控制步即失败。没有状态日志或最后回合的精确终止读数，两项速度误差指标未报告，控制步与视频时间的映射也未核实。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点0/1、断点（操作中）、平衡现象与停在“控制”的根因均与视频和服务读数一致；视频 0.00→1.56 s 可见腿弯折、机身降低，log step 45/47 的服务读数确认80步failed。未发现需要修改的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:23:58Z",
+  "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/mujoco_playground_barkour_joystick-standard-realtime--hfgptfp/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--hfgptfp/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "2a06b143ff638b265d892efdb7f09e63402a5c52dcb306567972761c64b80b6a",
+   "video_sha256": "a039d7d420a3a745b1e607407a10322d4445bacd34ccc84bad050adda643b558",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard-realtime--kcephp9.json

@@ -0,0 +1,202 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard-realtime--kcephp9",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "n_actions": 324,
+   "sim_steps": 3834,
+   "calls": 12135,
+   "rows": 8913,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_startups": 1,
+   "elapsed_s": 1102.1,
+   "video_rendered": 1,
+   "realtime": 1,
+   "slowdown": 1.0,
+   "clock_s": 38.244,
+   "max_lag_s": 0.0033,
+   "stalls": 0,
+   "stalled_s": 0.0,
+   "stalled_share": 0.0,
+   "held_steps": 1034,
+   "late_steps": 0,
+   "requests_refused": 32,
+   "request_busy_s": 3.254,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.388,
+   "peak_processes": 10,
+   "peak_disk_gb": 1.059,
+   "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/mujoco_playground_barkour_joystick-standard-realtime--kcephp9/replay.mp4",
+  "duration_s": 4.36,
+  "fps": 25.0,
+  "frames": 109,
+  "sha256": "a94769a7d04ec5284c712ea18c8d4e51fc3db9e1b43a3df2ea31484b9dafa983"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.8,
+      "log_step": null,
+      "text_zh": "观察：视频0.0–1.6 s，跟随视角中Barkour机器人反复迈腿、足端接触棋盘格地面并移动，已开始行进；起始站立和短时迈步不满足完整日程的节点判据。"
+     },
+     {
+      "t": 3.4,
+      "log_step": null,
+      "text_zh": "观察：视频3.2–3.76 s，机器人机身明显倾斜；视频3.8–4.32 s，机身继续下沉，末帧呈低伏姿态，没有恢复正常站立行进。"
+     },
+     {
+      "t": null,
+      "log_step": 113,
+      "text_zh": "观察：log step 113的机器人服务status返回episode=11、control_steps=324、done=true、reason=failed、resets_left=0；log step 115及120再次确认同一终止状态。该评分回合未运行到要求的1000个控制步，故完整日程节点从未成立；此前回合的测试结果不计入。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "视频0.0–3.2 s可见Barkour四条腿反复屈伸，足端交替落地承重并推动机身行进；视频3.4–4.32 s机身倾斜下沉，正常腿部支撑未维持。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 3.4,
+   "log_step": 111,
+   "where": "操作中",
+   "evidence_zh": "观察：视频3.2–3.76 s，正在迈步的Barkour机器人出现明显机身倾斜，随后在视频3.8–4.32 s下沉至低伏姿态，未恢复正常行进。log step 111启动最后一次步态控制后收到episode finished、reason failed；log step 113确认该回合为episode 11，止于control_steps=324，未完成1000步日程。"
+  },
+  "unmet": {
+   "joystick_schedule_followed": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 3.4,
+    "log_step": null,
+    "evidence_zh": "观察：视频3.2–3.76 s，机器人机身明显倾斜，腿部未能使机身保持平稳；视频3.8–4.32 s，机身逐渐下沉，最后低伏在地面附近。这一失稳在视频结束前没有修复。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 97展示agent的步态脚本以控制步更新相位、由速度误差调整步幅，并以IMU重力及角速度修正膝关节目标；log step 111修改俯仰反馈项后运行最后一回合，仍收到失败终止。结合视频3.2–4.32 s的倾斜和下沉，实际步态控制未能维持机身稳定，归为控制；没有足够证据将最终失稳唯一归于某个增益、符号或支撑布局。"
+   }
+  ],
+  "gap_zh": "没有状态日志。log step 113的服务状态只报告reason=failed，未给出��后回合的关节角、躯干高度、向上轴分量或终止子原因，不能确认具体是哪条终止规则触发；run.json的reward也没有worst_lin_err或worst_yaw_err，二者保留null。服务control_steps=324按0.01秒控制周期对应仿真3.24 s，但视频时长为4.36 s，最后可见帧为视频4.32 s，未提供经核实的时钟映射；所有视频锚点均直接取自画面，未把仿真3.24 s当作视频时间。早先回合的数值与agent关于失稳机理的自述不足以确定最终回合的具体控制子类。"
+ },
+ "summary_zh": "最后评分回合已开始迈步行进，但在视频约3.4 s出现明显倾斜，随后下沉，按指令行进节点未达成（0/1）；log step 113确认episode 11在控制步324失败，未运行完1000步日程。步态控制未能维持机身平衡，具体控制子类证据不足；参考成功视频可见机器人持续支撑行进并在后半段转向，本次未完成该过程。没有状态日志或两项速度误差评分值，且视频时长与控制步对应的仿真时长不同，无法确定精确终止条件和时钟映射。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点、平衡现象和停在“控制”的根因均有视频与服务读数支持；时钟未把控制步换算成视频时间，未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:22:38Z",
+  "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/mujoco_playground_barkour_joystick-standard-realtime--kcephp9/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--kcephp9/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7aa559700ec4e92cd5395177ed4d2ff9739784b25fc52f9cdefb573378b6e537",
+   "video_sha256": "a94769a7d04ec5284c712ea18c8d4e51fc3db9e1b43a3df2ea31484b9dafa983",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard-realtime--tebn89a.json

@@ -0,0 +1,183 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard-realtime--tebn89a",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "n_actions": 1000,
+   "sim_steps": 5131,
+   "calls": 10565,
+   "rows": 10556,
+   "resets": 6,
+   "reset_budget": 10,
+   "sim_startups": 1,
+   "elapsed_s": 1010.0,
+   "video_rendered": 1,
+   "realtime": 1,
+   "slowdown": 1.0,
+   "clock_s": 51.255,
+   "max_lag_s": 0.0075,
+   "stalls": 0,
+   "stalled_s": 0.0,
+   "stalled_share": 0.0,
+   "held_steps": 159,
+   "late_steps": 0,
+   "requests_refused": 7,
+   "request_busy_s": 3.527,
+   "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/mujoco_playground_barkour_joystick-standard-realtime--tebn89a/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "2d816ba833ffd899746638326d80211ff927554bcf235d370f1f98e10b108014"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.14,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 2.0,
+      "log_step": null,
+      "text_zh": "观察：已查看本次视频全部143帧；视频0.00–5.00 s，Barkour四足机器人持续交替迈腿，棋盘格地面相对机器人移动，躯干由腿支起。"
+     },
+     {
+      "t": 6.5,
+      "log_step": null,
+      "text_zh": "观察：视频5.14–6.79 s，机器人继续迈步，头部和躯干的朝向逐渐改变；视频7.00–10.14 s转向持续，头部由朝画面右侧逐渐转向左前方。视频9.14–10.14 s逐帧可见机器人仍由腿支撑，未见翻倒。"
+     },
+     {
+      "t": null,
+      "log_step": 53,
+      "text_zh": "观察：log step 52中agent脚本重置并启动最后一次控制回合；log step 53中机器人仿真服务的status（现场回合状态接口）报告episode=7、control_steps=1000、done=True、reason=success。inputs/run.json的评分器字段同时记录success=live_success=replay_success=1、deterministic=1、n_actions=1000；此前调试回合的失败不属于本次评分视频。"
+     },
+     {
+      "t": 10.14,
+      "log_step": 53,
+      "text_zh": "推断：视频完整行进、后半段转向及结束时保持支撑，与最终现场回合和评分器的成功结果一致，故完整指令日程及其误差验收已通过；以已检查的视频末帧10.14 s作为节点达成的近似锚点。没有将agent工作时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.50–10.14 s，Barkour机器人四条腿的足端交替落在棋盘格地面上，支撑躯干并完成平移和转向。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "本次评分轨迹在视频约10.14 s完成“按指令行进”，最终1/1节点成立；机器人前半段平移，后半段持续转向，直到视频末帧仍由腿支撑。log step 53的服务状态确认最终第7回合运行满1000个控制步并成功，评分器的现场与回放结果均为成功。没有状态日志，评分器reward也未单独报告worst_lin_err和worst_yaw_err，因此精确原生误差值保留为null；log step 53中agent脚本自行统计的误差不替代评分器原生指标。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "评分成功（success=1，live/replay均为1），目标节点final=true、无失败层，与视频末帧机器人仍由腿支撑一致；contact_actual为腿/躯干支撑，合理。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:23:28Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--tebn89a/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--tebn89a/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "1bfd20b9311400a702b212990d19b4d81a478c7809ba65d5b0b514d51ced90ef",
+   "video_sha256": "2d816ba833ffd899746638326d80211ff927554bcf235d370f1f98e10b108014",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/barkour_joystick/trajectories/mujoco_playground_barkour_joystick-standard-realtime--vyhemke.json

@@ -0,0 +1,208 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_barkour_joystick-standard-realtime--vyhemke",
+ "taskKey": "mujoco-playground/barkour_joystick",
+ "dag_sha256": "e24d0f402dbcad765944ffad94e7b301b234603065e44e6cf6bb76b5bbdcfb00",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "n_actions": 1000,
+   "sim_steps": 10128,
+   "calls": 9156,
+   "rows": 9178,
+   "resets": 10,
+   "reset_budget": 10,
+   "sim_startups": 1,
+   "elapsed_s": 782.2,
+   "video_rendered": 1,
+   "realtime": 1,
+   "slowdown": 1.0,
+   "clock_s": 101.191,
+   "max_lag_s": 0.003,
+   "stalls": 0,
+   "stalled_s": 0.0,
+   "stalled_share": 0.0,
+   "held_steps": 950,
+   "late_steps": 0,
+   "requests_refused": 18,
+   "request_busy_s": 4.207,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.202,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.286,
+   "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/mujoco_playground_barkour_joystick-standard-realtime--vyhemke/replay.mp4",
+  "duration_s": 10.214,
+  "fps": 14.0,
+  "frames": 143,
+  "sha256": "8c8fa91a9b6a395a850d896042f6a9519f5bd087aaf2552171e2cd7e6507b3fc"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "joystick_schedule_followed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.0 s机器人以四足支撑站在棋盘格地面；节点要求完整日程及两段跟踪验收，开局站立不代表已完成。"
+     },
+     {
+      "t": 2.5,
+      "log_step": null,
+      "text_zh": "观察：视频0.5–3.5 s机器人腿部交替运动，足端接触地面支撑机身；黄色前端由画面右侧逐渐转向相机，前半段已发生明显偏航。"
+     },
+     {
+      "t": 5.5,
+      "log_step": null,
+      "text_zh": "观察：视频5.0–10.143 s机器人继续由腿支起，机身方向仅小幅变化，未出现持续的大幅转向，最后一秒及末帧也没有翻倒。成功参考视频5.5–10.0 s可见机身持续转向；该参考仅用于识别可见动作。"
+     },
+     {
+      "t": 10.143,
+      "log_step": null,
+      "text_zh": "推断：本次run.json的reward记录n_actions=1000、success=0且infra_error=0；结合视频前半段明显偏转、后半段转向不足，完整按指令行进的节点未达到。评分字段未提供两段误差均值，不能把走满日程或未倒地等同于跟踪达标，也不把前次测试的自算误差当成本次评分值。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.0–10.143 s，Barkour四条腿的足端接触棋盘格地面，腿部交替活动并支撑躯干；没有手臂或抓手参与。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "joystick_schedule_followed",
+   "t": 2.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频1.5–3.5 s，机器人在第一段指令期间已从侧向相机的姿态明显转向相机；视频5.5–10.143 s又未呈现后半段所需的持续转向。该行进步骤一直未完成，断点以视频2.0 s的明显偏转作为粗略导航锚点，不表示精确误差阈值在此刻首次被突破。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 2.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频1.5–5.0 s，机器人边动腿边明显改变朝向；视频5.5–10.143 s，机身却基本维持同一朝向，仅有小幅摆动，未像成功参考视频后半段那样持续旋转。推断：结合log step 85所打印的共同指令日程（第一段偏航指令接近零，第二段为正向持续转动），可见本次行进的转向时序与幅度不符；全程未见失去支撑或倒地，故标步态。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 109,
+    "source_zh": "日志",
+    "evidence_zh": "推断：agent在log step 109、111检查的是前一次测���的遥测，显示偏航反复正负摆动且第二段侧移和转向跟踪不足；随后在log step 112改写步态及偏航补偿控制器，log step 114启动最终回合。最终视频1.5–3.5 s仍有明显非指令转向，视频5.5–10.143 s转向不足，支持行进控制未实现所要求的速度与偏航跟踪。前次测试只作为控制问题的会话证据，不作为最终节点或评分值；具体是增益、运动学还是足地作用造成，现有证据不能细分。"
+   }
+  ],
+  "gap_zh": "无状态日志；本次reward未报告worst_lin_err、worst_yaw_err、分段误差或终止条件细项，最终回合的完整服务返回也未收录于log step 114之后，因此原生指标值留null，不能精确定位误差首次超标时刻或控制子类型。log step 109–111的遥测属于此前测试，log step 120的完成情况是agent自述，均不替代最终评分回合的观测。",
+  "unmet": {
+   "joystick_schedule_followed": false
+  }
+ },
+ "summary_zh": "评分记录显示最终回合执行1000个控制动作，视频中机器人始终由腿支撑，但前半段明显偏转、后半段持续转向不足，“按指令行进”最终为0/1。断点在该步骤操作中，现象为步态，结合会话中的偏航调试记录将根因归于控制。缺少本次分段误差及状态日志，两个原生指标值未知，无法进一步定位控制子类型。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点、步态现象与停在“控制”的根因均有视频与评分字段（success=0、n_actions=1000）支持；视频全程未倒地，前半段明显偏航，后半段转向不足，与标签一致。log step 85的日程与会话遥测被标为推断/会话证据，未当作评分值，未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:23:48Z",
+  "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/mujoco_playground_barkour_joystick-standard-realtime--vyhemke/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_barkour_joystick-standard-realtime--vyhemke/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c81e8bdf5ebe91f7159ad1e1ec4330809b6d2c8235c2df9ebda6ba135e61b863",
+   "video_sha256": "8c8fa91a9b6a395a850d896042f6a9519f5bd087aaf2552171e2cd7e6507b3fc",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-privileged--7b5rrvr.json

@@ -0,0 +1,193 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-privileged--7b5rrvr",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 0,
+   "n_actions": 341,
+   "horizon": 500,
+   "fell": 1,
+   "fell_at_step": 26.0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.0554,
+   "end_pose_angle_deg": 96.87,
+   "video_rendered": 1,
+   "agent_sessions": 53,
+   "agent_resets": 83,
+   "agent_sim_steps": 35445,
+   "agent_replays": 9,
+   "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/mujoco_playground_go1_handstand-privileged--7b5rrvr/replay.mp4",
+  "duration_s": 6.706,
+  "fps": 17.0,
+  "frames": 114,
+  "sha256": "204697cfda223d28020206276859530241a0d05a49f60e5dd1d240fa376828d3"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 的track与side并排画面显示Go1四足姿态，躯干近水平，未处于头朝下的竖直倒立。"
+     },
+     {
+      "t": 0.53,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–1.00 s，Go1短暂抬起后躯，但躯干始终明显偏离竖直；视频约 0.53 s 前腿折低、腿段落地，随后躯干下沉，未建立可持续的前脚倒立。"
+     },
+     {
+      "t": 6.65,
+      "log_step": 119,
+      "text_zh": "观察：视频 1.0–6.65 s 的检查帧中Go1一直低伏，视频 5.65–6.65 s 逐帧检查也未见恢复；log step 119 对最终交付文件的重放返回success=false、fell_at=26、pose_angle_deg=96.87、imu_height=0.0554，与评分器终态字段一致。未出现符合倒立判据的连续保持窗口。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.53 s，Go1脚部与地面接触，在抬后躯、前倾尝试中以前脚接地；随后前腿折低、腿段及躯干落到地面，视频 0.8–6.65 s 维持低伏接地。未形成持续的竖直倒立支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better",
+   "value": 96.87
+  },
+  {
+   "name": "imu_height",
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better",
+   "value": 0.0554
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 0.53,
+   "log_step": 119,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 0.53 s，Go1正在将四足姿态转为前腿倒立，前腿却折低、腿段落地，视频约 0.8 s 躯干已低伏，直到视频 6.65 s 未恢复。log step 119 的最终交付重放及评分器字段均记录fell_at=26且success=false；控制步编号仅作日志证据，视频锚点取自逐帧观察。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 0.53,
+    "log_step": 119,
+    "evidence_zh": "观察：视频 0.24–0.53 s，Go1后躯抬起但前腿支撑折低，前腿腿段落到地面；视频 0.65–1.0 s 躯干继续下沉，之后直到视频 6.65 s 都低伏在地，未站成倒立。log step 119 的最终文件重放返回fell_at=26，评分器fell=1。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "轨迹或逆解",
+    "log_step": 111,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 46 的交付生成脚本依次使用固定起跳目标、24个控制步的线性插值和恒定保持目标；log step 111 将保持目标替换为单独验证过的静态倒立目标，但重放最终文件仍返回fell_at=26、success=false，log step 119 再次得到相同终态。log step 113 的成功保持测试先直接写入倒立qpos并清零qvel，并未验证从规定初态起立。结合视频 0.24–0.8 s 的前腿折倒，可判断交付的起立过渡轨迹未能将规定初态带入稳定倒立；单独成立的静态姿态不能补足这段运动。"
+   }
+  ],
+  "gap_zh": "本次与��考运行均无状态日志；无法逐控制步核对接触传感器、IMU角度和高度，也无法确定首次跌倒具体由哪项终止传感条件触发。视频足以判断始终未形成倒立保持；根因依据已交付脚本及其重放结果推断，未进一步归因为某个增益、摩擦或力参数。"
+ },
+ "summary_zh": "Go1尝试从四足姿态抬后躯并转为前腿倒立，但视频约 0.53 s 前腿折倒，随后低伏至片尾，“前腿倒立”始终未达成（终态0/1）。交付的起立过渡轨迹未将机器人带入稳定倒立；log step 119 的最终重放和评分器一致报告失败，终态夹角96.87度、IMU高度0.0554米，单独写入倒立状态的保持测试不属于本次评分轨迹。缺少状态日志，无法确认首次失败的具体接触传感条件。",
+ "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（0/1）、断点（front_leg_handstand，约0.53 s，对应fell_at=26×0.02 s）、平衡现象与轨迹/逆解根因均与视频和log step 119重放结果（success=false, fell_at=26, 96.87°, 0.0554 m）一致；视频约0.5 s后躯干低伏直到片尾。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:04:18Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--7b5rrvr/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--7b5rrvr/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "cbe8ec98f6cfc7057fc7d6a3d51a12fd24b532453f58babe0450191abb91903f",
+   "video_sha256": "204697cfda223d28020206276859530241a0d05a49f60e5dd1d240fa376828d3",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-privileged--7bm4aag.json

@@ -0,0 +1,173 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-privileged--7bm4aag",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 1,
+   "n_actions": 135,
+   "horizon": 500,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.5442,
+   "end_pose_angle_deg": 10.24,
+   "video_rendered": 1,
+   "agent_sessions": 2,
+   "agent_resets": 4,
+   "agent_sim_steps": 1350,
+   "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/mujoco_playground_go1_handstand-privileged--7bm4aag/replay.mp4",
+  "duration_s": 2.72,
+  "fps": 25.0,
+  "frames": 68,
+  "sha256": "188eb95a52ecb3bbe903201601e19ecb88e7a3f179942990afd12272ef16ac65"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [
+     [
+      2.68,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s的并排视图中，Go1躯干近水平、四脚位于地面，尚未进入前腿倒立。逐帧检查视频0.04–0.68 s可见机器人屈腿、腾起并将后躯向上转，随后躯干转为近竖直、头端朝下。"
+     },
+     {
+      "t": 0.88,
+      "log_step": null,
+      "text_zh": "观察：视频0.68–0.88 s机器人由腾起后的近竖直姿态逐渐伸出前腿，视频0.88 s前脚已到地面支撑位置、后脚高于躯干；逐帧检查视频0.92–2.68 s，机器人保持头朝下、躯干近竖直和双后脚离地，前脚在地面附近作小幅支撑调整，未见倒地。"
+     },
+     {
+      "t": 2.68,
+      "log_step": 25,
+      "text_zh": "观察：log step 25记录agent修正生成脚本后重新保存的最终交付文件为135行动作，Session重放和独立CLI重放均报告success=true、steps=135、time_s=2.7、fell_at=null、pose_angle_deg=10.24、imu_height=0.5442；inputs/run.json评分器字段同样为success=1、replay_success=1、deterministic=1、fell=0。视频2.64 s和2.68 s均显示前腿支撑倒立姿态。"
+     },
+     {
+      "t": 2.68,
+      "log_step": 25,
+      "text_zh": "推断：依本次评分规则，成功结果确认末尾连续100个控制步满足夹角、高度和后脚离地条件。节点包含完整保持窗口，故将完成锚点近似放在视频2.68 s末帧，而非首次转为近竖直的时刻；log step 25的135个控制步×0.02 s=仿真2.70 s与视频2.72 s总长及视频末尾倒立相符，仅用于核对交付片段。没有逐步状态记录，无法精确确定最早合格采样和最早完成时刻。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：Go1以脚与地面作用起立，转起过程中有短暂腾空；视频0.88–2.68 s前脚处于地面支撑位置、双后脚悬空，前腿与躯干协同保持倒立。并排视图中两前脚有重叠，不能逐步量化各脚承重。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better",
+   "value": 10.24
+  },
+  {
+   "name": "imu_height",
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better",
+   "value": 0.5442
+  }
+ ],
+ "failure": null,
+ "summary_zh": "Go1由四足姿态转起，最终完成前腿倒立节点（1/1）；视频末尾保持头朝下、躯干近竖直、后脚离地，前脚位于地面支撑位置。最终交付的135步轨迹在log step 25经重新生成和独立重放成功，评分器确认末尾连续100步合格、无历史翻倒，终态夹角10.24°、IMU高度0.5442米。无逐步状态日志，精确阈值和连续计数依据评分结果，完成时刻近似标在视频2.68 s；oracle元数据的runId与本次相同，未作为独立参考证据。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "视频显示Go1由四足起立至头朝下倒立并保持到末尾，评分器成功、无翻倒，节点final=true、时间锚点2.68 s（含100步保持窗口）合理，接触标注为腿/躯干·支撑，无失败层，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": "2026-10-30T03:09:19Z",
+  "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/mujoco_playground_go1_handstand-privileged--7bm4aag/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--7bm4aag/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "a51668b0ba8538cab1aa9384f8926bdd61981672f386add1b7a2aaeaac8bda7a",
+   "video_sha256": "188eb95a52ecb3bbe903201601e19ecb88e7a3f179942990afd12272ef16ac65",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-privileged--efkdv8w.json

@@ -0,0 +1,192 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-privileged--efkdv8w",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 0,
+   "n_actions": 500,
+   "horizon": 500,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.2465,
+   "end_pose_angle_deg": 113.91,
+   "video_rendered": 1,
+   "agent_sessions": 666,
+   "agent_resets": 1335,
+   "agent_sim_steps": 35180,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 11.554,
+   "peak_processes": 12,
+   "peak_disk_gb": 0.241,
+   "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/mujoco_playground_go1_handstand-privileged--efkdv8w/replay.mp4",
+  "duration_s": 10.5,
+  "fps": 12.0,
+  "frames": 126,
+  "sha256": "83d7953eb59b17a73010df5d629127364bbfc986012bfdead185c54094b487b5"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 的track与side并排视图中，Go1以四足姿态接地，躯干近水平；视频约 0.7–1.2 s 腿部调整后，头端抬高、后部下沉，未转成头朝下的近竖直姿态。"
+     },
+     {
+      "t": 1.0,
+      "log_step": null,
+      "text_zh": "观察：视频 1.0–10.0 s 的逐段画面及视频 9.5–10.42 s 的逐帧末段中，Go1持续保持头高尾低的倾斜姿态，后脚仍在地面，未见后腿抬至身体上方；成功参考视频约 0.6–2.68 s 中近竖直躯干、头朝下和后腿高悬的可见特征在本次均未出现。"
+     },
+     {
+      "t": 10.4167,
+      "log_step": 109,
+      "text_zh": "观察：评分器字段end_pose_angle_deg=113.91、end_imu_height=0.2465，分别不满足不超过30°和至少0.40米的要求；log step 108保存500行动行后，log step 109及113对该交付文件重放，均返回success=false、pose_angle_deg=113.91、imu_height=0.2465、fell_at=null，与视频末态一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–1.2 s，Go1前后脚与地面接触支承机体，前腿调整、后腿弯曲使躯干头端抬高；视频 1.2–10.42 s 仍以前后脚接地支撑，未形成后脚悬空的前脚倒立支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 113.91,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.2465,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 1.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 0.7–1.2 s，Go1调整腿部时头端上扬、后部下沉，随后停留在该倾斜姿态直到视频 10.42 s；本次没有建立头朝下、后腿在上方的倒立状态。评分器记录fell=0，失败末态角度113.91°、IMU高度0.2465米。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 1.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频约 1.0 s 起，Go1头端高于后部、躯干仍接近水平而略倾斜，后腿折在地面附近；视频 9.5–10.42 s 仍是此姿态，没有转成头朝下、躯干近竖直、后腿高悬的目标倒立。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：agent在log step 108用俯仰路径及腿部关节控制器生成并保存500行轨迹，目标俯仰参数thd=78°，但该文件在log step 109及113的两次重放均停在pose_angle_deg=113.91、imu_height=0.2465、success=false；与视频约 1.0–10.42 s 的头高尾低停滞一致，说明交付的关节运动未实现起倒立的姿态转换。现有记录不足以将问题细分到某个增益、逆解或反馈环节。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有逐控制步的状态日志，无法精确复核IMU角度、高度、脚触地传感器和连续100步窗��；视频中的持续姿态偏差及评分器末态字段已足以判节点未达。控制失效的具体机制未被独立验证；agent在log step 121自述的会话试验过度翻转和制动不足不作为本次交付轨迹发生翻倒的证据。"
+ },
+ "summary_zh": "Go1尝试调整腿部进入前腿倒立，但在视频约 1.0 s 形成头高尾低、后脚仍接地的倾斜姿态，并持续至视频 10.42 s，前腿倒立节点始终未达（0/1）。交付控制轨迹未实现目标姿态转换；评分器记录未翻倒，末态角度113.91°、IMU高度0.2465米均不合格。缺少逐控制步状态日志，无法进一步定位控制机制或复核传感器连续窗口。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达（0/1）、断点 front_leg_handstand、位姿不准 + 控制（未细分）均有视频、评分器字段（end_pose_angle_deg=113.91、end_imu_height=0.2465）和 log step 108/109/113 支持，未发现需改动的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:10:53Z",
+  "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/mujoco_playground_go1_handstand-privileged--efkdv8w/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--efkdv8w/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "ec285becb26769726f8cbab3d17b5308be4795bf93a478db55b627daf9667f4f",
+   "video_sha256": "83d7953eb59b17a73010df5d629127364bbfc986012bfdead185c54094b487b5",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-privileged--etmewbj.json

@@ -0,0 +1,178 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-privileged--etmewbj",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 1,
+   "n_actions": 108,
+   "horizon": 500,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.5046,
+   "end_pose_angle_deg": 17.3,
+   "video_rendered": 1,
+   "agent_sessions": 21,
+   "agent_resets": 544,
+   "agent_sim_steps": 51130,
+   "agent_replays": 6,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.937,
+   "peak_processes": 8,
+   "peak_disk_gb": 0.252,
+   "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/mujoco_playground_go1_handstand-privileged--etmewbj/replay.mp4",
+  "duration_s": 2.2,
+  "fps": 25.0,
+  "frames": 55,
+  "sha256": "ff2ace2b61cf756c36503e7f86540b73089f35e29a832cdafd49dd56d8192d9d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [
+     [
+      0.18,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.04 s 的两路画面中，Go1躯干近水平、四脚位于地面，尚未进入倒立；视频 0.08–0.16 s 机器人以前脚支撑向头端转起，后脚逐渐离地。"
+     },
+     {
+      "t": 0.2,
+      "log_step": null,
+      "text_zh": "观察：视频 0.20 s 起，Go1头端朝下、躯干接近竖直、后腿位于身体上方，前脚位于机体下方支撑；已逐帧检查视频 0.20–2.16 s，包括末秒，机器人保持此姿态至末帧，未见倒地或后脚落地。"
+     },
+     {
+      "t": 0.18,
+      "log_step": null,
+      "text_zh": "推断：依据视频 0.16 s 与视频 0.20 s 的起立过渡帧，将持续合格姿态窗口起点近似标为视频 0.18 s（误差约一个视频帧）；按state节点持续窗口的区间惯例回记起点，完整100个控制步的验收在视频末尾完成。评分器reward给出success=1、replay_success=1、deterministic=1、n_actions=108、fell=0，支持该窗口连续合格；无逐步状态日志，不能由画面单独精确测量阈值。"
+     },
+     {
+      "t": 2.16,
+      "log_step": null,
+      "text_zh": "观察：视频 2.16 s 末帧中，Go1仍以前脚支撑倒立，后脚悬空；inputs/run.json的评分器字段end_pose_angle_deg=17.3、end_imu_height=0.5046，分别满足不超过30度和至少0.40米的终态要求。"
+     },
+     {
+      "t": null,
+      "log_step": 75,
+      "text_zh": "观察：log step 75 的最终生成脚本重置后执行起立及保持动作，首次check返回成功便结束动作数组，并调用replay检查后保存；log step 76 调用该脚本和回放。较早log step 65所报告的115行动测试不是评分器记录的108行动交付，未据此前次测试标注本视频节点。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00 s 起立前Go1四脚支承于地面；视频约 0.20–2.16 s，机器人以前脚与地面的接触支撑头朝下的机体，双后脚抬离地面。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better",
+   "value": 17.3
+  },
+  {
+   "name": "imu_height",
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better",
+   "value": 0.5046
+  }
+ ],
+ "failure": null,
+ "summary_zh": "Go1从四足站立转为前脚支撑倒立，视频约 0.18 s 起的姿态窗口保持至视频 2.16 s 末帧，前腿倒立节点终态完成（1/1）。评分器确认连续100个控制步合格、无历史翻倒，末态前向夹角17.3度、IMU高度0.5046米。未提供逐步状态日志，姿态窗口起点为视觉近似，精确角度、高度及连续窗口验收依据评分器记录。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "前腿倒立节点终态完成、区间起点约0.18 s、contact_actual（腿/躯干 支撑）与帧一致；无失败层，符合���功运行。时钟均为视频时间，validator通过。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:02:19Z",
+  "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/mujoco_playground_go1_handstand-privileged--etmewbj/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--etmewbj/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "de30815becac1744ade0cefb80032e79a9987466d1f51e319e86fc7dd10adf28",
+   "video_sha256": "ff2ace2b61cf756c36503e7f86540b73089f35e29a832cdafd49dd56d8192d9d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-privileged--kyknckd.json

@@ -0,0 +1,173 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-privileged--kyknckd",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 1,
+   "n_actions": 116,
+   "horizon": 500,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.5252,
+   "end_pose_angle_deg": 1.42,
+   "video_rendered": 1,
+   "agent_sessions": 607,
+   "agent_resets": 613,
+   "agent_sim_steps": 71615,
+   "agent_replays": 5,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 18.916,
+   "peak_processes": 7,
+   "peak_disk_gb": 0.255,
+   "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/mujoco_playground_go1_handstand-privileged--kyknckd/replay.mp4",
+  "duration_s": 2.36,
+  "fps": 25.0,
+  "frames": 59,
+  "sha256": "1eaec060d5bf2595bfbe5a76563c7b2459b83009180db123beefa5f5713e0f1a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [
+     [
+      2.32,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s中Go1四足着地、躯干近水平，尚未满足倒立；视频0.08–0.32 s中机器人后腿离地，躯干向头端下方转起。"
+     },
+     {
+      "t": 0.48,
+      "log_step": null,
+      "text_zh": "观察：视频0.36–0.48 s的track与side并排画面中，Go1头端朝下、躯干近竖直，双前脚落在地面支撑，后腿举向身体上方；视频0.60–2.32 s持续保持该姿态，已逐帧检查视频1.32–2.32 s，未见倒下或后脚重新落地。"
+     },
+     {
+      "t": 2.32,
+      "log_step": 157,
+      "text_zh": "观察：评分器reward报告success=1、replay_success=1、deterministic=1、n_actions=116、fell=0，终态前向轴夹角1.42度、IMU高度0.5252米；log step 157重新生成最终交付文件后，基准回放工具同样返回116步、仿真2.32 s、success=true，终态哈希与log step 140、141一致。"
+     },
+     {
+      "t": 2.32,
+      "log_step": 139,
+      "text_zh": "推断：log step 138的最终生成脚本在连续100个合格控制步后停止，log step 139报告保存116步，且log step 157对该交付文件的回放成功。按DAG控制周期0.02秒，116×0.02对应视频2.32 s；已在视频2.32 s末帧的竖直倒立处核对，故将包含持续窗口的节点完成时刻标为视频2.32 s，而非刚转起时。无逐步状态日志，连续窗口由成功验收与视频保持过程共同支持。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–0.04 s中Go1四脚接触地面；转起时后脚离地，视频0.40–2.32 s中左右前脚接触地面并支撑竖直躯干，后腿保持抬起。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 1.42,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.5252,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "Go1从四足站立转起，约在视频0.36 s形成近竖直的头朝下姿态，之后以前脚支撑并保持至视频2.32 s；包含连续100控制步保持要求的前腿倒立节点在视频2.32 s完成，最终1/1。评分器确认确定性回放成功、无翻倒，终态夹角1.42度、IMU高度0.5252米；缺少逐步状态日志，精确连续窗口依赖最终交付文件的成功验收，视频提供姿态与接触支持。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点front_leg_handstand达成、final=1与评分器success/fell=0一致；视频中姿态与oracle一致（头朝下、前脚支撑、后腿抬起），2.32 s按116步×0.02 s换算并在末帧核对，无失败层，未见错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:07:24Z",
+  "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/mujoco_playground_go1_handstand-privileged--kyknckd/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--kyknckd/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "05f6fbf1c883227c478fbb9e63e518997c2e1519471b3c755cdf4abdfde44d2b",
+   "video_sha256": "1eaec060d5bf2595bfbe5a76563c7b2459b83009180db123beefa5f5713e0f1a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-privileged--p6ijlye.json

@@ -0,0 +1,173 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-privileged--p6ijlye",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 1,
+   "n_actions": 116,
+   "horizon": 500,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.5338,
+   "end_pose_angle_deg": 12.71,
+   "video_rendered": 1,
+   "agent_sessions": 16,
+   "agent_resets": 19,
+   "agent_sim_steps": 7840,
+   "agent_replays": 3,
+   "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/mujoco_playground_go1_handstand-privileged--p6ijlye/replay.mp4",
+  "duration_s": 2.36,
+  "fps": 25.0,
+  "frames": 59,
+  "sha256": "94b4948be309cbb3b352a5ddc6b162f7da15afddc45d64894600c4579d9f13b3"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [
+     [
+      0.34,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s的track与side并排画面中，Go1四足站立、躯干近水平，尚未倒立；视频0.12–0.32 s机器人将后腿抬起，躯干逐渐转向头朝下。"
+     },
+     {
+      "t": 0.36,
+      "log_step": null,
+      "text_zh": "观察：逐帧查看本次视频0.00–2.32 s；视频0.36 s机器人头端朝下、躯干近竖直，前脚位于地面支撑位置，后脚悬空。视频0.40–2.32 s机器人继续用前腿调整支撑，后脚未见落地，末秒及视频2.32 s末帧仍保持倒立。成功参考视频0.48 s及2.40–2.68 s呈现相同的头朝下、后腿悬空姿态，仅用于辨识目标外观。"
+     },
+     {
+      "t": 2.32,
+      "log_step": 17,
+      "text_zh": "观察：log step 17中agent重新生成并保存交付轨迹后，评分器replay命令报告steps=116、time_s=2.32、success=true、fell_at=null、pose_angle_deg=12.71、imu_height=0.5338；inputs/run.json的reward同时给出n_actions=116、success=1、replay_success=1、deterministic=1、fell=0及相同终态角度和高度。log step 16的213步重放属于另一会话测试，不用作本次节点进度。"
+     },
+     {
+      "t": 0.34,
+      "log_step": 17,
+      "text_zh": "推断：DAG要求末尾100个连续合格控制步，交付116步且成功，故至少第17–116个控制步满足姿态条件；按每控制步0.02秒，第17步为仿真0.34 s。log step 17的116步对应仿真2.32 s，与视频2.32 s末帧核对一致，视频0.32–0.36 s也正呈现转入近竖直倒立，因此以视频约0.34 s作为合格保持区间起点（位于相邻帧之间，非传感器实测边界）。该区间在末尾经连续窗口验收后回溯标出；无逐步状态日志，不能独立核验每个0.02秒采样。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00 s Go1四脚接地；转起后，视频0.16–2.32 s前脚在地面承担支撑、前腿屈伸调整机体姿态，双后脚抬离地面。双前脚并非每帧均能分辨同时接地，未据此声称持续双脚承重或给出接触力。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better",
+   "value": 12.71
+  },
+  {
+   "name": "imu_height",
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better",
+   "value": 0.5338
+  }
+ ],
+ "failure": null,
+ "summary_zh": "Go1从四足站立转为以前脚支撑、头朝下且后腿悬空的倒立，视频约0.34 s起进入最终合格保持区间，并持续至视频2.32 s末帧，唯一节点最终完成。评分器确认116步交付成功且无历史翻倒，末态前向夹角12.71°、IMU高度0.5338米，均满足阈值；本次无失败断点。缺少逐步状态日志，连续100步的精确角度、高度与接触判据由成功评分确认，视频用于核对姿态和接触外观，区间起点为近似推断。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "唯一节点front_leg_handstand最终完成、无失败断点，与评分器116步成功、视频末帧倒立一致；0.34 s起点已明确标为由100步窗口推断。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:02:57Z",
+  "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/mujoco_playground_go1_handstand-privileged--p6ijlye/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--p6ijlye/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "dae45ebd330974e46e8e037f4ccef2d6496e6f4c405b36e633abfd0335eaaf3f",
+   "video_sha256": "94b4948be309cbb3b352a5ddc6b162f7da15afddc45d64894600c4579d9f13b3",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-privileged--trejvr7.json

@@ -0,0 +1,187 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-privileged--trejvr7",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 64,
+   "horizon": 500,
+   "fell": 1,
+   "fell_at_step": 64.0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.4344,
+   "end_pose_angle_deg": 22.51,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 3066,
+   "agent_resets": 3080,
+   "agent_sim_steps": 483135,
+   "agent_replays": 4,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 1,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 9.883,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.831,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 7,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "over_oom": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/mujoco_playground_go1_handstand-privileged--trejvr7/replay.mp4",
+  "duration_s": 1.32,
+  "fps": 25.0,
+  "frames": 33,
+  "sha256": "8c659eb60755e5752e19822fb946266b1e6caee342b70fa5ab643f81ab26f54c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s的track与side并排画面中，Go1为四足站姿；视频0.08–0.52 s机器人蹬地腾起并将头端转向下方，视频0.60–1.04 s可见前脚落地支撑、后腿悬空，躯干接近竖直。"
+     },
+     {
+      "t": 1.28,
+      "log_step": 71,
+      "text_zh": "推断：逐帧检查视频0.00–1.28 s，整段运动也不足节点要求的100个控制步（首末跨度1.98 s），不能把短暂倒立记为完成。log step 71对已保存交付文件的重放报告steps=64、time_s=1.28、success=false、fell_at=64；评分器同样记录n_actions=64、fell=1、deterministic=1。视频1.12–1.28 s躯干继续转动、头端下降，未恢复稳定保持。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–0.08 s Go1脚部接触地面蹬起；腾空后，视频约0.60–1.28 s左右前脚在地面支撑机体，后脚悬空；左右前脚承重分配及前腿非脚部是否触地无传感记录，不能精确区分。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 22.51,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.4344,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 1.28,
+   "log_step": 71,
+   "where": "操作中",
+   "evidence_zh": "观察：视频1.08–1.28 s Go1在前脚撑地、后脚悬空时躯干继续向另一侧倾转，头端逐渐下降，末帧尚未完全躺倒；log step 71交付文件重放在控制步64报告fell_at=64、success=false，评分器记录fell=1。推断：该控制步按任务给定0.02 s步长对应视频1.28 s，已与末段姿态变化及末帧核对；未完成保持时已触发历史翻倒判据，断点属于前腿倒立操作中。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.28,
+    "log_step": 71,
+    "evidence_zh": "观察：视频1.04 s后Go1的后腿再次摆动，视频1.12–1.28 s躯干越过近竖直姿态继续倾转、头端降低，前脚支撑未使姿态稳定；log step 71对交付文件的重放报告fell_at=64，评分器记录fell=1。此处为倒立过程失衡，画面结束前未显示完全倒地。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频 / 评分器字段",
+    "evidence_zh": "推断：log step 70列出的交付动作和log step 71重建脚本一致：机器人先执行20个控制步的蹬起目标，再固定前腿目标、分段改变后腿目标，最后12个控制步重复同一组目标。视频约0.60–1.04 s前脚支撑时躯干近竖直，随后在视频1.12–1.28 s继续倾转；log step 71重放与评分器均确认控制步64翻倒。该开环关节目标序列未能稳定转起后的机体姿态，支持控制层归因；不足以证明特定增益、接触几何或逆解问题。"
+   }
+  ],
+  "gap_zh": "无逐步状态日志，无法确定控制���64的fell由躯干上轴低于阈值还是前腿非脚部触地触发，也不能精确恢复角度、足部接触和高度的逐步变化；因此不细分控制子类。视频仍足以确认实际尝试和短暂前脚支撑，64步交付长度及重放记录足以否定100步连续保持。"
+ },
+ "summary_zh": "Go1完成蹬地转起，并在视频约0.60–1.04 s以前脚支撑、后腿悬空呈近竖直姿态，但未达到前腿倒立的100步连续保持，节点最终0/1。视频1.12–1.28 s机体继续倾转，交付文件重放在控制步64记录翻倒；开环关节目标序列未能稳定姿态，根因标为控制。缺少逐步状态与接触传感记录，无法确定翻倒的具体触发条件或控制子类。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点（前腿倒立，操作中）、平衡/控制归因与证据均与视频和评分器字段一致，未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:09:39Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--trejvr7/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-privileged--trejvr7/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f53d65d008cb4379b4e018c163d5133378b607645ada4734d1505550ebbd6b3b",
+   "video_sha256": "8c659eb60755e5752e19822fb946266b1e6caee342b70fa5ab643f81ab26f54c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-standard--emkqvnw.json

@@ -0,0 +1,204 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-standard--emkqvnw",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 31,
+   "sim_steps": 31,
+   "physics_steps": 155,
+   "steps": 54,
+   "elapsed_s": 846.0,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.28,
+   "end_pose_angle_deg": 15.97,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "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/mujoco_playground_go1_handstand-standard--emkqvnw/replay.mp4",
+  "duration_s": 0.64,
+  "fps": 25.0,
+  "frames": 16,
+  "sha256": "81c93d5bdd3b363a55550af6866eaee3d61fc49360a9551784216b33a8fe5c64"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 74,
+      "text_zh": "观察：log step 74重置到episode 11、resets_left=0；任务文本说明只评分最后一局。视频0.00–0.60 s全部16帧为同一段从四足站立向头朝下俯转的运动，没有重置；评分器记录episodes=11、n_actions=31、live_matches_replay=1。"
+     },
+     {
+      "t": 0.6,
+      "log_step": 81,
+      "text_zh": "观察：视频0.36–0.60 s机器人后躯升起、头端持续降低，前腿仍折在机体下方，末帧头端和前腿关节已贴近地面；没有形成参考视频0.60–2.60 s所示的机体抬高、前腿伸出支撑的倒立。评分器报告end_pose_angle_deg=15.97、end_imu_height=0.28 m、fell=1；角度虽进入30度阈值，高度未达0.40 m。"
+     },
+     {
+      "t": 0.6,
+      "log_step": 81,
+      "text_zh": "推断：整段视频仅0.64 s，且初态为四足站立，不能完成连续100个控制步的保持窗口；结合末态高度不足与服务在log step 81返回done=True、reason=fell，前腿倒立节点从未达到。log step 83–85才留下的未测试策略不属于这段评分运动。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–0.32 s，Go1的前后脚与地面接触支承，机器人屈伸后腿并向头端俯转；视频0.40–0.60 s后脚抬离地面、前脚在低位接触地面，前腿保持折叠，未建立能抬高机体的稳定支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 15.97,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.28,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 0.56,
+   "log_step": 81,
+   "where": "操作中",
+   "evidence_zh": "观察：视频0.48–0.60 s机器人在前腿仍折叠的情况下继续向头端俯转，头端降至地面附近，未能以前脚将机体撑高；log step 81服务在control_steps=31返回done=True、reason=fell，此后没有评分运动。推断：断点是正在执行的前腿倒立。control period=0.02 s，31步约为仿真0.62 s，与视频0.60 s末帧和0.64 s片长相符，时间锚点取自所见帧。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 0.56,
+    "log_step": 81,
+    "evidence_zh": "观察：视频0.48–0.60 s，Go1头端快速下沉，前腿折在低位，机体未被前脚稳定托起，末帧头端与前腿关节贴近地面；同局服务在log step 81返回fell，评分器fell=1、末态IMU高度仅0.28 m。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：前腿展开与机体俯转的控制未协调好，未能在头端降低前建立足够高的支撑。依据是log step 76–77的脚本连续提交前大腿1.3 rad、前膝-1.8 rad目标，log step 79改为前大腿-0.3 rad、前膝-2.7 rad时服务报告俯仰角速度8.23 rad/s；log step 81再将前膝目标改为-1.8 rad，但终止时实际两前膝仍为-2.81与-2.45 rad、俯仰角速度7.94 rad/s。视频0.40–0.60 s也显示机体继续俯转而前腿未撑开；这些证据支持控制层归因，但不足以细分接触设计、轨迹或反馈问题。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志或逐控制步的前腿非脚部接触传感器记录；末帧前腿与头端局部重叠，无法精确确定首先触地的腿段，也不能进一步区分控制根因的子类。视频已覆盖全段，节点未完成和失衡终止有画面、服务结果与评分器字段共同支持。"
+ },
+ "summary_zh": "评分采用第11局：Go1从四足站立转为头朝下，后腿抬起，但前腿未撑高机体，前腿倒立节点未达到（0/1）。视频约0.56 s失衡下沉，服务随后以fell终止；末态角度15.97度符合角度阈值，但IMU高度仅0.28 m，且未形成连续保持窗口，根因可归于控制协调失败。缺少逐步状态与腿段接触记录，无法确定首先触地的部位或细分控制原因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标签与视频、日志一致：评分局为第11局，视频0.64 s全程俯转，末帧头端贴地、机体近竖直但高度仅0.28 m，服务返回fell；节点未达到、断点在front_leg_handstand、平衡现象与控制根因均有证据支持，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": "2026-10-30T03:20:42Z",
+  "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/mujoco_playground_go1_handstand-standard--emkqvnw/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--emkqvnw/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "14bae1e4099a260cc775c05477e164c0b596721644a29d23d79599f06f6eed2e",
+   "video_sha256": "81c93d5bdd3b363a55550af6866eaee3d61fc49360a9551784216b33a8fe5c64",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-standard--fpg6bcp.json

@@ -0,0 +1,206 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-standard--fpg6bcp",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 76,
+   "sim_steps": 76,
+   "physics_steps": 380,
+   "steps": 441,
+   "elapsed_s": 1766.5,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.1678,
+   "end_pose_angle_deg": 49.14,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 6,
+   "reset_budget": 10,
+   "episodes": 7,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.434,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.273,
+   "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/mujoco_playground_go1_handstand-standard--fpg6bcp/replay.mp4",
+  "duration_s": 1.56,
+  "fps": 25.0,
+  "frames": 39,
+  "sha256": "2e5a6c3ca32f839671685a0f9158520399bdb957036efc49d005ce5f0edaec8e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.64 s，Go1从四足站立开始，躯干接近水平并逐渐前倾，未处于头朝下、躯干近竖直的倒立姿态。"
+     },
+     {
+      "t": 1.12,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.68–1.52 s，Go1反复抬脚和俯仰，视频 1.12–1.20 s虽后腿收起、后端抬高，但躯干仍明显倾斜；视频 1.28–1.52 s前腿折低、机体下坠，始终未出现持续合格倒立。成功参考视频 0.80–2.60 s中近竖直的高位姿态仅用于对照，非本次进度证据。"
+     },
+     {
+      "t": 1.52,
+      "log_step": 34,
+      "text_zh": "观察：log step 34在最后一次重置后执行第7个episode，服务返回control_steps=76、done=True、reason=fell；评分器同样记录n_actions=76、fell=1、live_matches_replay=1，终态IMU高度0.1678 m、前向轴与竖直向下夹角49.14°，分别未达0.40 m和30°要求；与视频末帧低位折倒相符。log step 35只修改脚本，未再执行机器人动作。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.64 s，Go1四脚接触地面支撑；视频 0.68–1.20 s转起过程中前后脚交替抬起和触地，前脚曾短暂承重，后脚未能持续悬空；视频 1.28–1.52 s腿部折低，未形成稳定的前腿倒立支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better",
+   "value": 49.14
+  },
+  {
+   "name": "imu_height",
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better",
+   "value": 0.1678
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 1.28,
+   "log_step": 34,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.12–1.20 s，Go1尝试收后腿并以前脚支撑抬高后端；视频 1.28 s开始明显回落，视频 1.36–1.52 s前腿折低、头端降到地面附近，未重新建立倒立。log step 34的最后episode随后以fell结束，评分器fell=1。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.28,
+    "log_step": 34,
+    "evidence_zh": "观察：视频 0.68–1.20 s，Go1支撑脚不断变化、躯干前后俯仰；视频 1.28–1.52 s支撑高度塌落、前腿折叠，机体向地面下坠，未稳住。log step 34服务终止原因为fell。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "反馈控制",
+    "t": 1.28,
+    "log_step": 34,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 34的受评控制脚本用IMU俯仰和角速度反馈，每5个控制步更新前脚位置，并依据后脚触地读数切换后腿目标；同一日志显示角速度w由1.04变为-3.54、再至-6.54 rad/s，前脚位置修正dx达到-0.120 m限幅，随后fell。结合视频 0.80–1.52 s的俯仰反复与支撑塌落，表明这套转起反馈未能稳定姿态和角速度；不能据此确定某个增益或逆解公式是唯一原因。"
+   }
+  ],
+  "gap_zh": "无状态日志，无法逐控制步核验IMU角度、高度与全部接触，也无法区分最终fell具体由前腿非脚部触地还是躯干过度翻倒触发。控制根因来自log step 34的脚本、稀疏遥测与视频联合推断，缺少逐步目标与实际关节轨迹，不能进一步确定失稳参数或模型误差。"
+ },
+ "summary_zh": "受评轨迹是log step 34的最后一次episode；Go1启动了转起尝试，但反复俯仰、换脚支撑，前腿倒立节点始终未完成，终态为0/1。视频 1.28–1.52 s显示前腿折低、机体下坠，服务以fell结束；脚本与遥测支持转起反馈未能稳住姿态和角速度。无状态日志，具体触发终止的接触或姿态边界及失稳参数无法确定。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "前腿倒立节点未达成、终态0/1、断点与平衡/反馈控制标签与视频帧和评分器字段（fell=1、高度0.1678、角度49.14°）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": "2026-10-30T03:05:48Z",
+  "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/mujoco_playground_go1_handstand-standard--fpg6bcp/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--fpg6bcp/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "95dd6b61c839aabcaf46d085a94f737047ef6874449ff1a1c806a5609ff07f67",
+   "video_sha256": "2e5a6c3ca32f839671685a0f9158520399bdb957036efc49d005ce5f0edaec8e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-standard--hwfgywj.json

@@ -0,0 +1,190 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-standard--hwfgywj",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "sim_error": 0,
+   "grader_error": 0,
+   "infra_error": 0,
+   "n_actions": 25,
+   "sim_steps": 25,
+   "physics_steps": 125,
+   "steps": 25,
+   "elapsed_s": 1834.5,
+   "video_rendered": 1,
+   "fell": 0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.255,
+   "end_pose_angle_deg": 91.88,
+   "instance_mismatch": 0,
+   "no_verdict": 0,
+   "gave_up": 0,
+   "sim_crashed_on_call": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 0,
+   "reset_budget": 10,
+   "episodes": 1,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.246,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.239,
+   "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/mujoco_playground_go1_handstand-standard--hwfgywj/replay.mp4",
+  "duration_s": 0.52,
+  "fps": 25.0,
+  "frames": 13,
+  "sha256": "3568125d9f16ec6ba8ce7dcbeb20ecafaac6eb224c205547fad5db86080f3637"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：已逐帧查看本次视频0.00–0.48 s的全部13帧（间隔0.04 s）；track与side两视图中Go1躯干始终近水平，前后脚处于地面支撑位置，仅有轻微起伏，没有头朝下、躯干近竖直和后腿抬至身体上方的倒立动作。"
+     },
+     {
+      "t": 0.48,
+      "log_step": 6,
+      "text_zh": "观察：log step 6中agent脚本只向实时机器人服务提交25行相同的初始站立关节目标，dt=0.02 s，用于延迟测试；服务返回executed=25、control_steps=25，末尾四脚接触读数均为1。评分器记录单一episode、无reset/resume、总计25个控制步，终态前向夹角91.88度、IMU高度0.255 m；视频0.48 s末帧仍为四足站姿。"
+     },
+     {
+      "t": null,
+      "log_step": null,
+      "text_zh": "观察：成功参考视频0.40–2.60 s中可见Go1头朝下、躯干近竖直、后腿处于上方的明确倒立外观；本次视频所有帧均没有这一外观。推断：本次前腿倒立节点从未成立，并非仅由success=0推定。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 91.88,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.255,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 0.0,
+   "log_step": 6,
+   "where": "首步前",
+   "evidence_zh": "观察：视频0.00–0.48 s中Go1一直保持近水平四足站姿；log step 6唯一执行的控制是重复初始站姿25步的延迟测试，log step 7只查询正向/逆向运动学，未提交倒立动作。推断：评分轨迹停留在初始站姿保持，前腿倒立步骤未启动；视频0.00 s仅作首步前的导航锚点。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.0,
+    "log_step": 6,
+    "evidence_zh": "观察：视频0.00–0.48 s全部13帧均未见Go1转起或抬后腿进入倒立；log step 6提交的25行目标均为qhome初始站姿，评分器n_actions=25且episodes=1、resets=0，与这段站姿保持相符。推断：有评分轨迹，但其中没有开始倒立步骤。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "观察：inputs内没有状态日志；log step 6和log step 7的候选倒立脚位逆解查询返回ok=false，但均未执行相应动作。log step 8与log step 10没有调用、输出或说明，完整日志也无最终自述。推断：这些查询不足以区分候选目标或种子不当、求解限制与后续未执行的具体原因，故根因留空；不能由未尝试或预算耗尽反推规划或控制根因。倒立未启动，未将初始四足站立的背景支撑接触记作该步骤的contact_actual。"
+ },
+ "summary_zh": "评分轨迹仅包含初始四足站姿保持，前腿倒立��未达到，最终完成0/1节点；视频0.00–0.48 s全部帧均无转起或后腿上举，评分器终态夹角91.88度、IMU高度0.255 m也未达标。断点为前腿倒立的首步前，现象为没尝试；日志中的逆解查询失败未形成实际倒立动作，缺少足够证据确定未执行的根因。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "标注与证据一致：视频全部帧为四足站姿，log step 6 仅提交25步初始站姿延迟测试，评分器25步、91.88度/0.255 m；节点未达成、断点首步前、没尝试、根因留空并有gap_zh，均成立。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:21:32Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--hwfgywj/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--hwfgywj/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "619e33df09ae19e5c906f6e668679a62cdebde61820206dcd61f804646811dfc",
+   "video_sha256": "3568125d9f16ec6ba8ce7dcbeb20ecafaac6eb224c205547fad5db86080f3637",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-standard--mexor4y.json

@@ -0,0 +1,190 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-standard--mexor4y",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 129,
+   "sim_steps": 129,
+   "physics_steps": 645,
+   "steps": 129,
+   "elapsed_s": 1823.9,
+   "video_rendered": 1,
+   "fell": 0,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.5525,
+   "end_pose_angle_deg": 9.15,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 0,
+   "reset_budget": 10,
+   "episodes": 1,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.85,
+   "peak_processes": 60,
+   "peak_disk_gb": 0.384,
+   "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/mujoco_playground_go1_handstand-standard--mexor4y/replay.mp4",
+  "duration_s": 2.6,
+  "fps": 25.0,
+  "frames": 65,
+  "sha256": "17cbf22a45075d67f7a088787f881499a2398569c0578d9838feaeaeba2a5e11"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [
+     [
+      2.56,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00 s的track与side并排画面中，Go1躯干近水平、四脚在地面支撑，尚未形成前腿倒立。"
+     },
+     {
+      "t": 0.6,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频0.00–0.96 s，Go1先屈伸腿部，随后头端向下、后躯抬起；视频约0.60 s躯干已近竖直，前脚处于下方支撑位置，双后脚位于身体上方并离地。此时只是开始维持姿态，尚不能计为完成连续保持窗口。"
+     },
+     {
+      "t": 2.56,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频1.00–2.56 s，尤其视频1.60–2.56 s末段，Go1保持头端向下、躯干近竖直，左右前脚接地支撑，后脚悬空；视频2.56 s末帧仍为该姿态。成功参考视频末段具有相同的头朝下、后脚离地特征，仅用来识别姿态，不作为本次完成证据。"
+     },
+     {
+      "t": 2.56,
+      "log_step": 79,
+      "text_zh": "观察：log step 79中agent脚本向机器人实时服务提交joint_position动作，dt=0.02；该步未记录返回结果。inputs/run.json评分字段另记录n_actions=129、sim_steps=129、live_success=1、replay_success=1、deterministic=1、live_matches_replay=1、fell=0，终态end_pose_angle_deg=9.15、end_imu_height=0.5525。"
+     },
+     {
+      "t": 2.56,
+      "log_step": null,
+      "text_zh": "推断：本次评分器的成功结果确认末尾100个控制步满足倒立验收，且视频末段姿态与此一致，因此final为true；节点含1.98 s保持窗口，不在初次转起时计为完成，而把窗口完成时刻近似锚定在视频末帧2.56 s。无状态日志，无法从25 fps画面精确重建50 Hz控制步的首次合格时刻或逐步传感器值，时间存在约一帧至数帧的不确定性。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00 s起Go1以脚接触地面；转起后左右前脚接地支撑机体，前腿屈伸调整躯干姿态，双后脚抬起并在倒立保持阶段悬空，末帧仍由前脚支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 9.15,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.5525,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "观察：Go1从四足站立转起，视频约0.60 s形成前脚支撑、头朝下且后脚悬空的倒立姿态，并保持至视频2.56 s末帧。推断：结合评分器live_success=1、replay_success=1及终态角度9.15°、IMU高度0.5525米，唯一节点“前腿倒立”在末尾连续保持窗口完成后达成，终态为1/1，无失败断点。观察：本次没有状态日志，无法独立逐控制步核验100步窗口及精确���触阈值，窗口完成时间近似标在视频末端；log step 78的本地测试后续翻倒不属于本次评分视频。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "视频帧显示Go1由四足站立转起，0.6 s起头朝下、躯干近竖直、后脚离地并保持至2.56 s末帧；评分器live_success=1、fell=0，节点final=true、无失败断点与标注一致，时钟均为视频时间。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:11:56Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--mexor4y/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--mexor4y/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "64ea4a9224396eb4dc5f426c63bda80e1c9a7bf00473d398ee90de3916fe3914",
+   "video_sha256": "17cbf22a45075d67f7a088787f881499a2398569c0578d9838feaeaeba2a5e11",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-standard--pjmrq6k.json

@@ -0,0 +1,209 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-standard--pjmrq6k",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 188,
+   "sim_steps": 188,
+   "physics_steps": 940,
+   "steps": 891,
+   "elapsed_s": 1317.7,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.2298,
+   "end_pose_angle_deg": 22.17,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.57,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.833,
+   "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/mujoco_playground_go1_handstand-standard--pjmrq6k/replay.mp4",
+  "duration_s": 3.8,
+  "fps": 25.0,
+  "frames": 95,
+  "sha256": "f66f7e0cf552ef4e2e03cfe8c22a8c973528cf3d4e7eafad5d0bffe6090e835e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 的两个视角均显示Go1四足站立，躯干近水平，初始并未倒立。"
+     },
+     {
+      "t": 3.2,
+      "log_step": 79,
+      "text_zh": "观察：视频 0.20–3.40 s 中Go1逐渐调整腿部、前移机体，躯干一直近水平，视频 3.20 s 头端还略高于后端；没有出现躯干近竖直、后腿位于身体上方的持续倒立。"
+     },
+     {
+      "t": 3.76,
+      "log_step": 83,
+      "text_zh": "观察：逐帧检查视频 2.60–3.76 s；视频 3.56 s 起Go1头端快速下沉、后腿离地，视频 3.68–3.76 s 前腿折低，头端与前腿段接近或落到地面，没有抬高身体并稳定保持倒立。评分器reward报告end_pose_angle_deg=22.17、end_imu_height=0.2298 m、fell=1；末态角度虽在30度内，高度低于0.40 m且已判倒地，不能完成该节点。"
+     },
+     {
+      "t": 3.6,
+      "log_step": 85,
+      "text_zh": "推断：任务文本规定只评最后一局；log step 71–83记录最后一次重置后的动作，log step 85的机器人服务状态确认episode=11、control_steps=188、done=true、reason=fell。log step 7所读服务规格为50 Hz，故log step 81的control step 180对应视频 3.60 s，已与画面头端下倾核对；log step 83的control step 188对应视频 3.76 s，吻合最后一帧。前面重置掉的试验不计入节点进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–3.40 s，Go1前后脚与地面支撑并调整机体；视频 3.56–3.76 s 后脚抬离地面、前脚继续触地，但前腿折低，末帧前腿段及头端贴近地面。实际尝试包含足地支撑，未形成成功的前腿倒立支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 22.17,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.2298,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 3.6,
+   "log_step": 81,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 3.56–3.76 s，Go1在抬后腿、向头端转起时前部快速下沉，前腿折低，未能建立高位倒立平衡；log step 81–83记录补救关节指令后仍快速俯转，最终机器人服务以fell终止。失败发生在前腿倒立的执行中，之后无恢复。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 3.6,
+    "log_step": 83,
+    "evidence_zh": "观察：视频 3.56 s 起Go1后腿抬起而头端下坠，视频 3.68–3.76 s 前腿折低至地面，未能用伸展的前腿支撑机体；log step 83的服务反馈为done=true、reason=fell，评分器fell=1。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：agent的控制没有使前腿转起与机体俯仰稳定下来。log step 77的脚本按IMU和关节反馈生成前腿位置指令；log step 78–79显示运行该控制时俯仰从接近水平偏至-13.7度，随后log step 80发出20步恢复轨迹，log step 81在control step 180读到俯仰27.7度、俯仰角速度4.07 rad/s。log step 82–83继续两次补救指令后角速度仍增至4.79、5.61 rad/s并判fell，与视频 3.60–3.76 s 前部下沉、前腿折低一致。证据支持运动控制失稳，尚不足以确定某个增益、模型参数或接触设计是具体根因。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐步状态日志或前腿各段触地传感器记录，无法独立定位最先触地的腿段及精确触发控制步；agent在log step 84自述前腿段触地，视频末段支持前腿折低，但不据此细分具体传感器。日志中的ref,F是agent脚本计算的参考量，不是实测接触力；缺少实测力及控制器验证，根因仅标控制，不进一步细分。"
+ },
+ "summary_zh": "评分的第11局完成0/1节点：Go1先以足地支撑调整机体，随后抬后腿、向头端转起，但视频 3.60–3.76 s 前部下沉、前腿折低，未连续保持倒立并被服务判fell。末态IMU高度0.2298 m低于0.40 m；补救指令未抑制俯仰转动，根因归为控制，缺少逐步触地和力数据以确定具体失稳机制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点未达成、断点在front_leg_handstand、平衡+控制的标注与视频和log一致：视频 3.48–3.76 s 机体快速前俯、头端/前腿触地，log step 81–83 俯仰27.7→46.1→67.8度并判fell；时钟换算（188步×0.02=3.76 s）与末帧吻合。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:05:47Z",
+  "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/mujoco_playground_go1_handstand-standard--pjmrq6k/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--pjmrq6k/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "44332ce1047d6d9687410599a17b8ea8997f37aa2ca136499157d9779b6a9f6d",
+   "video_sha256": "f66f7e0cf552ef4e2e03cfe8c22a8c973528cf3d4e7eafad5d0bffe6090e835e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-standard--vxa5wag.json

@@ -0,0 +1,204 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-standard--vxa5wag",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 135,
+   "sim_steps": 135,
+   "physics_steps": 675,
+   "steps": 838,
+   "elapsed_s": 1338.4,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.279,
+   "end_pose_angle_deg": 68.2,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "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/mujoco_playground_go1_handstand-standard--vxa5wag/replay.mp4",
+  "duration_s": 2.72,
+  "fps": 25.0,
+  "frames": 68,
+  "sha256": "faf6c58cd3c713df53a026e11a78cd5da1d5702b51ddcdc058108da9bd625ab4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.72 s，Go1四脚位于地面，躯干近水平并逐渐下蹲，尚未倒立。"
+     },
+     {
+      "t": 0.88,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频 0.00–2.68 s；视频约 0.80–0.88 s，Go1后躯和双后腿抬高、头端向下，之后以前脚落地与短暂腾空交替支撑；视频约 2.32–2.68 s，躯干逐渐向后倾回并下降，末帧前腿折低，未保持倒立。"
+     },
+     {
+      "t": 2.4,
+      "log_step": 69,
+      "text_zh": "推断：视频中近竖直倒立仅约从视频 0.88 s持续到视频 2.32 s，短于100个控制步所需的1.98秒采样跨度，故不记为完成区间。log step 69的服务状态明确为最后一次episode=11、control_steps=135、reason=fell，与评分器n_actions=135、fell=1及末段下落一致；评分器终态前向夹角68.2度、高度0.279米也均不达标。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.72 s，Go1四脚接地支撑并下蹲；视频约 0.76–0.88 s，后脚蹬地后离地，随后左右前脚反复落地支撑机体，双后脚悬空；视频约 2.40–2.68 s，前腿弯折、机体下降。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 68.2,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.279,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 2.4,
+   "log_step": 69,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 2.40–2.68 s，正在以前脚支撑倒立的Go1后躯向后落下，躯干由近竖直转为明显倾斜，前腿折低；此后没有恢复。log step 69的服务记录最后一次episode=11在control_steps=135以reason=fell结束；评分器也报告fell=1、live_matches_replay=1。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 2.4,
+    "log_step": 69,
+    "evidence_zh": "观察：视频约 0.88–2.32 s，Go1以前脚落地和腾空交替支撑，后脚在空中；视频约 2.40–2.68 s，后躯持续向后落下、前腿弯折，倒立支撑失稳，未恢复。log step 69的服务终止原因为fell。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "反馈控制",
+    "evidence_zh": "推断：log step 42中agent编写的控制脚本在下蹲和蹬腿后，仅以IMU俯仰角及俯仰角速度的线性反馈修正前腿等关节目标；log step 57给该脚本加入初始站立段，log step 69以俯仰增益0.15、角速度增益0.12运行最后一次episode=11。视频约 0.88–2.32 s显示机器人已抬起后躯并能短时以前脚支撑，视频约 2.40–2.68 s却持续向后倾倒，说明这套反馈未能稳定维持倒立��这里不把某个具体增益或模型误差认定为已证实原因。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有逐控制步状态日志，无法精确核验连续合格计数、IMU角度和高度的越界时刻，也无法在躯干过度翻倒与前腿非脚部触地两种终止条件间确定实际触发项。近竖直阶段的起止时间由画面近似判断；控制根因可定位到反馈失稳，具体动力学或增益缺陷未被独立验证。",
+  "unmet": {
+   "front_leg_handstand": false
+  }
+ },
+ "summary_zh": "评分采用最后一次episode=11：Go1下蹲蹬地后抬起双后腿，并短时以前脚支撑，但没有完成连续100个控制步的倒立，节点最终为0/1。视频约 2.40–2.68 s，机器人在倒立操作中向后倾倒、前腿折低，现象为平衡失稳，日志中的俯仰反馈控制未能维持姿态。缺少逐步状态数据，无法精确确认合格窗口边界及具体终止传感器。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点未达成、断点 front_leg_handstand（操作中）、现象平衡、根因反馈控制均与视频和日志一致；近竖直阶段约 0.9–2.3 s 短于 1.98 s，末段 2.5 s 后塌倒，final=0 正确。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:05:07Z",
+  "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/mujoco_playground_go1_handstand-standard--vxa5wag/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--vxa5wag/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9de7a5d7f92401b22bf21d3f588bd778fce6e254b33aaf42e05e9f9aab97c1bf",
+   "video_sha256": "faf6c58cd3c713df53a026e11a78cd5da1d5702b51ddcdc058108da9bd625ab4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_handstand/trajectories/mujoco_playground_go1_handstand-standard--wydcxp2.json

@@ -0,0 +1,206 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_handstand-standard--wydcxp2",
+ "taskKey": "mujoco-playground/go1_handstand",
+ "dag_sha256": "e5a5a01397ef8139c61c86d38bf73d3ac237709765240ddc30438b6a648fb70c",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "live_matches_replay": 1,
+   "finalized": 1,
+   "n_actions": 78,
+   "sim_steps": 78,
+   "physics_steps": 390,
+   "steps": 1576,
+   "elapsed_s": 1571.9,
+   "video_rendered": 1,
+   "fell": 1,
+   "worst_lin_err": -1.0,
+   "worst_yaw_err": -1.0,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": 0.1723,
+   "end_pose_angle_deg": 106.57,
+   "grader_error": 0,
+   "instance_mismatch": 0,
+   "infra_error": 0,
+   "no_verdict": 0,
+   "missing_log": 0,
+   "gave_up": 0,
+   "sim_error": 0,
+   "sim_startups": 1,
+   "resumes": 0,
+   "resets": 10,
+   "reset_budget": 10,
+   "episodes": 11,
+   "resume_state_mismatch": 0,
+   "resume_max_joint_dev": 0.0,
+   "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/mujoco_playground_go1_handstand-standard--wydcxp2/replay.mp4",
+  "duration_s": 1.6,
+  "fps": 25.0,
+  "frames": 40,
+  "sha256": "f6040baa4aca2b02bdeaa18d3e1e8686750a72b492176a5d736981f81e7dfa0e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "front_leg_handstand": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–0.76 s的并排视图显示Go1从四足站立开始，躯干近水平，前后脚位于地面，未处于前腿倒立姿态。"
+     },
+     {
+      "t": 1.04,
+      "log_step": 69,
+      "text_zh": "观察：视频0.80–1.28 s中Go1前腿向前伸展、后腿弯折，头端仍高于后部，躯干未转到头朝下的近竖直姿态；视频1.32–1.56 s后部继续下沉，最后前腿也折低，未出现合格倒立或保持窗口。已逐帧检查视频0.00–1.56 s的全部40帧。"
+     },
+     {
+      "t": 1.56,
+      "log_step": 71,
+      "text_zh": "观察：log step 71的机器人实时服务status返回episode=11、control_steps=78、done=true、reason=fell、resets_left=0；评分器字段live_matches_replay=1、fell=1，末态end_pose_angle_deg=106.57、end_imu_height=0.1723。推断：该最后一局即受评轨迹；78个控制步×0.02秒=仿真1.56 s，与视频1.56 s的低伏终止姿态核对一致，节点未达成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00 s起Go1前后脚接地支撑身体；视频0.80–1.28 s伸展前腿、弯折后腿时，脚部仍与地面交互支撑，末段后部下沉、前腿折低。实际没有形成双前脚承重且双后脚持续悬空的倒立支撑。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pose_angle_deg",
+   "value": 106.57,
+   "unit": "deg",
+   "threshold": 30,
+   "direction": "lower_better"
+  },
+  {
+   "name": "imu_height",
+   "value": 0.1723,
+   "unit": "m",
+   "threshold": 0.4,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "front_leg_handstand",
+   "t": 1.32,
+   "log_step": 69,
+   "where": "操作中",
+   "evidence_zh": "观察：视频1.28–1.56 s，Go1在伸前腿、收后腿的姿态转换中后部明显下沉，身体倾斜并低伏，前腿末段也折低，始终未起成倒立；log step 69返回ENDED: fell、control_steps=78，log step 71确认episode 11已经结束，之后没有恢复或新一局。"
+  },
+  "unmet": {
+   "front_leg_handstand": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.32,
+    "log_step": 69,
+    "evidence_zh": "观察：视频1.32–1.56 s中Go1后腿收折后身体后部向地面塌下，躯干倾斜，前腿随后折低，未能维持稳定的脚部支撑；同一受评局的机器人实时服务在log step 69报告fell，评分器字段fell=1。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 1.32,
+    "log_step": 69,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 69实际执行的控制脚本按阶段伸前腿、收后腿，并用IMU重力方向及速度调节前腿路径和髋外展；同一步输出的rol由-7.0度增至-16.4度，随后服务报告fell。结合视频0.96–1.56 s的身体倾斜��后部下沉及前腿折低，控制器未能在姿态转换时维持稳定支撑，支持控制层面的归因；不能据此确定某个增益或接触几何就是具体根因。"
+   }
+  ],
+  "gap_zh": "无逐控制步的状态日志或非脚部接触传感记录，无法确定触发fell的具体前腿部位，也无法精确重建支撑力和重心轨迹；因此控制根因不细分。log step 69的knee与h是agent脚本的运动学估算，未当作真实接触或IMU高度。agent在log step 77对初始冲量、触地机制和后脚离地的解释仅为其自述，未据此补记物理事实。"
+ },
+ "summary_zh": "受评的最后一局（episode 11）中，Go1从四足站立伸展前腿并收折后腿，但始终未形成头朝下的近竖直倒立，节点最终0/1。视频约1.32–1.56 s身体后部下沉、前腿折低，实时服务在78个控制步后以fell结束；控制器未能稳定完成姿态转换。缺少逐步状态与接触力数据，控制根因不细分，未将此前重置测试或agent的机制解释当作受评局事实。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "受评局为最后一局（episode 11，log step 71 的服务status显示fell、control_steps=78，78×0.02=1.56 s与视频时长吻合）；视频逐帧显示始终是四足站立后伸腿、后部下沉，从未倒立，节点false、断点、平衡现象与控制归因均有证据支持，未发现需修改之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-30T03:10:01Z",
+  "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/mujoco_playground_go1_handstand-standard--wydcxp2/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_handstand-standard--wydcxp2/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "55c2f81a997292b8b89e4687601eb102ece3ff6050bb934215cb7669f0ed8d28",
+   "video_sha256": "f6040baa4aca2b02bdeaa18d3e1e8686750a72b492176a5d736981f81e7dfa0e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/mujoco-playground/go1_joystick_flat_terrain/trajectories/mujoco_playground_go1_joystick_flat_terrain-privileged--6xynkat.json

@@ -0,0 +1,198 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "mujoco_playground_go1_joystick_flat_terrain-privileged--6xynkat",
+ "taskKey": "mujoco-playground/go1_joystick_flat_terrain",
+ "dag_sha256": "869ae70c21c400f0d0532c63bf5be4417cd9459d397a35e1ff6684ee2b3521cc",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "deterministic": 1,
+   "grader_error": 0,
+   "infra_error": 0,
+   "replay_success": 0,
+   "n_actions": 1000,
+   "horizon": 1000,
+   "fell": 0,
+   "fell_at_step": -1.0,
+   "worst_lin_err": 0.7456,
+   "worst_yaw_err": 0.4176,
+   "end_tilt_deg": -1.0,
+   "end_imu_height": -1.0,
+   "end_pose_angle_deg": -1.0,
+   "video_rendered": 1,
+   "agent_sessions": 1823,
+   "agent_resets": 1829,
+   "agent_sim_steps": 501825,
+   "agent_replays": 6,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 1,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 36.874,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.255,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "over_memory_gb": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/mujoco_playground_go1_joystick_flat_terrain-privileged--6xynkat/replay.mp4",
+  "duration_s": 20.429,
+  "fps": 7.0,
+  "frames": 143,
+  "sha256": "f76b69e177f04b54ffe5f46291200e3dbc347ab51e992ae27bb827534f08b26e"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "command_schedule_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.714,
+      "log_step": null,
+      "text_zh": "观察：视频0.000–2.000 s逐帧检查可见Go1四条腿交替抬落，足端接触方格地面并推动机体移动和转向，已经开始按指令行走的尝试；起始站立不代表完成整个时间表。"
+     },
+     {
+      "t": 6.0,
+      "log_step": null,
+      "text_zh": "观察：视频4.571–7.000 s密集帧及视频8.0–19.0 s抽帧中，Go1持续屈伸腿、交替落足，躯干保持朝上，但地面格线相对机体仅缓慢移动；视频19.286–20.286 s末段逐帧仍如此，未见翻倒。"
+     },
+     {
+      "t": 20.286,
+      "log_step": 180,
+      "text_zh": "观察：log step 179重新生成最终轨迹，log step 180对该交付文件的官方回放报告steps=horizon=1000、fell_at=null、tracking_ok=false、success=false；四段(lin_err m/s, yaw_err rad/s)依次为(0.2680,0.3013)、(0.6557,0.4176)、(0.6247,0.3604)、(0.7456,0.3330)，每段均超过0.2 m/s和0.3 rad/s阈值；最差误差与本次reward字段完全一致。"
+     },
+     {
+      "t": 20.286,
+      "log_step": 180,
+      "text_zh": "推断：完整步数和未翻倒已经满足，但各段速度误差不达标，故唯一的完整时间表节点从未成立，最终为false；此结论来自最终交付的回放结果和本次视频，不将会话内其他试验的结果计入进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.000–2.000 s、4.571–7.000 s及19.286–20.286 s，Go1四条腿的足端交替着地、承重和推进，实际支撑对象为方格平地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "worst_lin_err",
+   "unit": "m/s",
+   "threshold": 0.2,
+   "direction": "lower_better",
+   "value": 0.7456
+  },
+  {
+   "name": "worst_yaw_err",
+   "unit": "rad/s",
+   "threshold": 0.3,
+   "direction": "lower_better",
+   "value": 0.4176
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "command_schedule_completed",
+   "t": 0.714,
+   "log_step": 180,
+   "where": "操作中",
+   "evidence_zh": "观察：视频0.571–0.857 s机器人已在交替踏步移动；log step 180最终交付文件的官方回放表明，首段平均平面速度误差0.2680 m/s及偏航误差0.3013 rad/s已经超过阈值，后续三段也全部超限。推断：失败发生在按指令行走的执行过程中，首段的整段验收失败未被后续运动修复；视频0.714 s仅为首段运动附近的粗略导航锚点，不是由画面测得的瞬时越阈时刻。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 6.0,
+    "log_step": 180,
+    "evidence_zh": "观察：视频4.571–7.000 s以及视频8.0–20.286 s中，Go1不断交替抬腿落足，机体相对地面格线的平移很慢，后段没有形成成功参考视频6.0–20.0 s所示的持续斜后行位移；末秒逐帧未见翻倒。log step 180确认全程无翻倒但速度跟踪不合格。推断：这是踏步产生的行进速度和方向未满足指令的步态失败。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "观察：log step 146的最终生成器用固定周期N=10的关节正弦目标驱动对角腿，并仅以pitch、roll修正关节目标；代码未将实测平面速度或偏航角速度与指令之差用于反馈。后两段沿用前段参数并将转向偏置置零，最后一段直接复制上一段参数；log step 179重新运行该生成器，log step 180官方重放得到与评分器相同的超限误差。推断：交付的关节运动控制没有实现所需的速度跟踪，解释了视频中的缓慢行进；证据不足以进一步确定为单一参数、逆解或接触几何问题，停在控制大项。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有状态日志或最终交付回放的逐控制步速度序列，视频也不显示速度指令或IMU读数，无法独立重算误差均值或精确定位首个瞬时偏差；分段误差与完整1000步以log step 180最终官方回放及reward为依据。控制根因仅能确定到大项，不能从现有证据锁定某个错误参数或具体接触力问题。",
+  "unmet": {
+   "command_schedule_completed": false
+  }
+ },
+ "summary_zh": "Go1实际执行了全部1000步且未翻倒，但四段平面速度与偏航速度误差均超限，唯一节点“按指令行走”从未达到，最终0/1。失败在行走操作中，视频后段持续踏步而平移缓慢，最终生成器和官方回放支持控制能力不足的归因。没有逐控制步状态日志，精确误差依赖最终官方回放，断点视频时间仅为首段的粗略定位。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点false、断点ref、步态现象与控制根因均有log step 179/180官方回放和视频支持，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": "2026-10-30T03:08:19Z",
+  "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/mujoco_playground_go1_joystick_flat_terrain-privileged--6xynkat/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/mujoco_playground_go1_joystick_flat_terrain-privileged--6xynkat/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "fba5cbafe58be084d4adc55778482bf20a642bab9e909ca47dd924450e536d19",
+   "video_sha256": "f76b69e177f04b54ffe5f46291200e3dbc347ab51e992ae27bb827534f08b26e",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}