Add 118 trajectories (phase 2, remaining runs; batch S41)

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-privileged--26tx94n.json

@@ -0,0 +1,186 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-privileged--26tx94n",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 2000,
+   "sim_steps": 413,
+   "horizon": 2000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 4,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 3.42,
+   "bad_trajectory": 0,
+   "metric_air_s": 0.429,
+   "metric_turn_deg": 315.3,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 18,
+   "agent_resets": 20,
+   "agent_replays": 13,
+   "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": 4.427,
+   "peak_processes": 29,
+   "peak_disk_gb": 1.0,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-privileged--26tx94n/replay.mp4",
+  "duration_s": 4.16,
+  "fps": 25.0,
+  "frames": 104,
+  "sha256": "eb4173e821cf560d79cd695ff94dc1e2fd059d134325d05d7937dfb75f9baa43"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.12,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.64,
+      "log_step": null,
+      "text_zh": "观察：视频 0.32–0.64 s，机器人在蓝色场地上屈腿准备后伸腿、摆臂；视频 0.64 s 脚仍支撑地面，开局站立本身尚未完成后空翻节点。"
+     },
+     {
+      "t": 0.84,
+      "log_step": null,
+      "text_zh": "观察：视频 0.72–1.04 s，机器人全身离地，收腿并向后翻转；视频 0.80–0.88 s 可见头在下、脚翻到上方，随后身体继续后转并展开准备落地。"
+     },
+     {
+      "t": 1.08,
+      "log_step": null,
+      "text_zh": "观察：视频 1.04 s 机器人脚接近地面，视频 1.08–1.12 s 可见脚先承重、双膝屈曲缓冲，双手在身体上方；视频 1.16–2.12 s 机器人通过小幅踏步与身体调整恢复直立，未见手、膝或躯干撑地。"
+     },
+     {
+      "t": 4.12,
+      "log_step": null,
+      "text_zh": "观察：视频 2.16–4.12 s 机器人保持基本直立、脚支撑地面；逐帧检查视频 3.04–4.12 s，直至末帧仍站立。inputs/run.json 的评分器 reward 记录 success=1、replay_success=1、deterministic=1、stages_reached=4/4、fell=0，metric_turn_deg=315.3、metric_air_s=0.429。"
+     },
+     {
+      "t": 4.12,
+      "log_step": 321,
+      "text_zh": "观察：log step 321 中，agent 脚本重新生成最终交付的 2000 行轨迹并核对一致；官方回放器执行 413 行后报告 success=true，四阶段 take-off、full turn、landing、stand 全部完成，转角 315.3°、腾空 0.429 s。log step 329 记录三次新进程回放均成功且终态哈希一致；这些结果与本次评分器字段和视频动作一致。"
+     },
+     {
+      "t": 4.12,
+      "log_step": null,
+      "text_zh": "推断：结合评分器成功和视频末帧站立，将包含合法后翻、脚先落地及连续站稳验收的整个节点标为视频末尾约 4.12 s 完成，时间为末帧导航锚点，非独立测得的精确验收采样时刻；没有状态日志，25 fps 录像无法逐采样核验累计转角、速度及连续 200 次站立条件，数值边界依评分器结果确认。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.40–0.68 s 机器人以脚对蓝色场地地面的支撑发力起跳；视频 0.72–1.04 s 全身腾空，随后视频约 1.08 s 起脚接地承重并屈腿缓冲，调整后继续用脚支撑站立至视频 4.12 s。未见手、膝或躯干撑地；双臂摆动协调，未见其接触场地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": 315.3
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": 0.429
+  }
+ ],
+ "failure": null,
+ "summary_zh": "本次交付轨迹完成唯一节点“后空翻并落地站稳”：机器人从脚支撑起跳、空中后翻、脚先落地，调整后站立至视频 4.12 s 末帧，评分器确认四阶段全部完成且回放一致。评分器记��空中后转 315.3°、腾空 0.429 s，未摔倒，无失败断点。缺少状态日志，具体数值边界及连续站稳采样依评分器确认，节点完成时间取视频末尾约 4.12 s；oracle 元数据所指 runId 与本次相同，没有独立参考视频。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点完成、末帧站立、评分器 success=1/fell=0、无断点，标注与视频和评分器字段一致；时间取末帧锚点并已说明为推断。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:32:47Z",
+  "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/embodied-olympics-fun-backflip-i00-privileged--26tx94n/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--26tx94n/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "e65d5f838874d4360acf308fb8b4e7eab20bd3a59ebea84cecd2aef14765acf6",
+   "video_sha256": "eb4173e821cf560d79cd695ff94dc1e2fd059d134325d05d7937dfb75f9baa43",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-privileged--btmnxme.json

@@ -0,0 +1,181 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-privileged--btmnxme",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 368,
+   "sim_steps": 368,
+   "horizon": 2000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 4,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 5.06,
+   "bad_trajectory": 0,
+   "metric_air_s": 0.682,
+   "metric_turn_deg": 313.8,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 10,
+   "agent_resets": 25,
+   "agent_replays": 4,
+   "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": 0.738,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.251,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-privileged--btmnxme/replay.mp4",
+  "duration_s": 3.72,
+  "fps": 25.0,
+  "frames": 93,
+  "sha256": "38b61a8954a2a4c9f12e0b732dd8a0e5f9fea1551089decadb2625e334f94fdb"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [
+     [
+      3.68,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.48,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–0.48 s，机器人先以双脚站在蓝色场地上，屈膝下蹲后伸腿、摆臂起跳；开局站立尚未完成后空翻。"
+     },
+     {
+      "t": 0.84,
+      "log_step": null,
+      "text_zh": "观察：视频0.52–1.08 s的逐帧画面显示机器人全身离地，头向后运动、屈髋收腿，翻过头低脚高的倒立姿态后展开双腿；机器人在这一连续腾空中完成后翻。"
+     },
+     {
+      "t": 1.16,
+      "log_step": null,
+      "text_zh": "观察：视频1.08 s机器人双脚仍在地面上方，视频1.12–1.16 s脚部先接近并接触地面承重；视频1.20–1.52 s机器人调整身体恢复直立，未见手、膝或躯干撑地。"
+     },
+     {
+      "t": 3.68,
+      "log_step": null,
+      "text_zh": "观察：视频1.68–3.68 s机器人保持双脚支撑、躯干接近直立且位置基本不变；视频2.72–3.68 s逐帧检查直至最后一帧，未见摔倒。inputs/run.json评分器字段为success=1、replay_success=1、stages_reached=4/4、fell=0、metric_turn_deg=313.8、metric_air_s=0.682。"
+     },
+     {
+      "t": 3.68,
+      "log_step": null,
+      "text_zh": "推断：视频中的后翻、脚先落地及末段站稳与评分器成功结果一致，据此判定包含连续站稳验收的完整节点达成；首次达成时间近似记在视频末帧3.68 s，未将视频约1.16 s的落地提前算作完成。没有状态日志，25帧/秒画面不能独立核验逐控制采样的接触、角度、骨盆速度及200次站稳计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–0.48 s机器人双脚在蓝色场地上支撑并蹬地起跳；视频0.52–1.08 s腾空；视频约1.12–1.16 s脚先触地，随后双脚承重站稳至视频3.68 s。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": 313.8
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": 0.682
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人从双脚支撑起跳，在空中完成后翻，双脚先落地并恢复直立，持续站稳至视频3.68 s，完成「后空翻并落地站稳」节点（1/1），无失败断点。评分器确认四个原生阶段全部通过，后转角313.8°、腾空0.682秒；无状态日志，完整验收时刻近似记在视频末帧，精确接触采样、速度与连续站稳计数依评分器结果确认。",
+ "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0–3.68 s显示起跳、空中后翻、脚先落地并直立站稳，与评分器success=1、fell=0一致；节点时间、contact_actual与无失败断点均成立。",
+    "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": "2026-10-31T00:37:29Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--btmnxme/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--btmnxme/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "735b4f0fe1d4e6eddacf3a19b88a5fa45f05eec955506518b6a68d14dd74fefd",
+   "video_sha256": "38b61a8954a2a4c9f12e0b732dd8a0e5f9fea1551089decadb2625e334f94fdb",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-privileged--l6om7hm.json

@@ -0,0 +1,181 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-privileged--l6om7hm",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 352,
+   "sim_steps": 352,
+   "horizon": 2000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 4,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 4.37,
+   "bad_trajectory": 0,
+   "metric_air_s": 0.436,
+   "metric_turn_deg": 304.2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 164,
+   "agent_resets": 26253,
+   "agent_replays": 5,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 3.289,
+   "peak_processes": 22,
+   "peak_disk_gb": 0.845,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-privileged--l6om7hm/replay.mp4",
+  "duration_s": 3.56,
+  "fps": 25.0,
+  "frames": 89,
+  "sha256": "f2bcf858a5a557aa4759a1a984e746d5cdc85684204874b04bc7713415719799"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [
+     [
+      3.52,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.56,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查本次视频 0.00–0.56 s，机器人在蓝色场地上由站立转为双腿屈曲，再伸腿发力；起跳前可见脚部承重，双手未撑地。开局站立尚未完成后空翻，不计为节点已达。"
+     },
+     {
+      "t": 0.76,
+      "log_step": null,
+      "text_zh": "观察：本次视频 0.64–0.96 s，机器人双脚离地、身体与地面分离；连续帧中机器人收腿向后翻，视频 0.76–0.84 s 可见脚高于头的倒立翻转姿态，随后展开双腿。"
+     },
+     {
+      "t": 1.04,
+      "log_step": null,
+      "text_zh": "观察：本次视频 0.96 s 机器人仍在空中，视频 1.00–1.04 s 脚部先到达蓝色地面并开始承重；随后机器人屈膝缓冲、直起躯干。视频 1.04–1.52 s 未见手、膝或躯干撑地。"
+     },
+     {
+      "t": 3.52,
+      "log_step": 163,
+      "text_zh": "推断：log step 163 的最终封装输出与 log step 167 的逐行一致复现均对应交付的352行轨迹，裁判记录控制步352完成stand并成功；每控制步0.01秒，故对应视频约3.52 s。已用裁判仿真1.008 s落地事件与视频1.00–1.04 s脚先承重核对时钟，且视频末帧为3.52 s；因此完整后翻及连续站稳验收的节点从视频约3.52 s成立，不能提前在刚落地时记完成。"
+     },
+     {
+      "t": 3.52,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查本次视频 1.52–3.52 s，机器人保持脚部支撑、躯干接近竖直，末段基本停稳；包括视频最后一秒在内均未见倒地。inputs/run.json评分器字段success=1、replay_success=1、deterministic=1、stages_reached=4/4、fell=0，metric_turn_deg=304.2、metric_air_s=0.436，与可见动作一致，支持终态完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：本次视频 0.00–0.56 s 机器人双脚接触蓝色场地承重并起跳发力；视频 0.64–0.96 s 腾空，与地面分离；视频约1.00 s起脚部落地承重，之后双脚支撑直至视频3.52 s末帧。双臂摆动协调，未见手部撑地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": 304.2
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": 0.436
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人完成「后空翻并落地站稳」节点，终态为1/1：视频约0.64–0.96 s可见腾空后翻，视频约1.00 s脚先落地，随后恢复直立并保持至视频3.52 s验收完成。评分器确认后转304.2°、腾空0.436秒、四阶段全部通过且确定性回放成功，无失败断点。缺少状态日志，无法仅凭25帧/秒录像独立复核毫秒接触及姿态、速度的精确数值边界；完成时刻结合交付轨迹控制步记录与视频事件核对确定。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "单节点后空翻落地站稳：视频约0.6–1.0 s腾空后翻、脚先落地，之后直立站稳至末帧；评分器字段success=1、fell=0支持，3.52 s完成时刻按控制步352×0.01换算并有落地事件核对，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": "2026-10-31T00:40:12Z",
+  "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/embodied-olympics-fun-backflip-i00-privileged--l6om7hm/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--l6om7hm/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "20bfbd9defceaef491a82da91e84d4a25dc9e41b1993e63167ce51c114b9eadf",
+   "video_sha256": "f2bcf858a5a557aa4759a1a984e746d5cdc85684204874b04bc7713415719799",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-privileged--njqpcrg.json

@@ -0,0 +1,197 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-privileged--njqpcrg",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.75,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 115,
+   "sim_steps": 115,
+   "horizon": 2000,
+   "episode_over": 0,
+   "progress": 0.75,
+   "stages_reached": 3,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 2.57,
+   "bad_trajectory": 0,
+   "metric_air_s": 0.59,
+   "metric_turn_deg": 310.9,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 6992,
+   "agent_resets": 77,
+   "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": 107.948,
+   "peak_processes": 17,
+   "peak_disk_gb": 0.298,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "over_memory_gb": 1,
+   "session_log_tampered": 1,
+   "session_log_from_monitor": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-privileged--njqpcrg/replay.mp4",
+  "duration_s": 1.16,
+  "fps": 25.0,
+  "frames": 29,
+  "sha256": "503d4c80c10bd8e4fd12b2bc5aa7963c1bbbd91c2007493743ca6e363713e12b"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.4,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–1.12 s；机器人先双脚站立并下蹲，在视频约 0.36–0.40 s 离地，随后收腿向后翻过倒立位置。开局站立没有后翻历史，不满足本节点。"
+     },
+     {
+      "t": 0.92,
+      "log_step": 1069,
+      "text_zh": "观察：视频 0.88 s 双脚仍在地面上方，视频约 0.92 s 双脚先落地，随后躯干前倾、双臂向前伸出；手、膝和躯干未见撑地。log step 1069 对交付文件的回放报告 stages=[take-off, full turn, landing]、turn_deg=310.9、air_s=0.59，与评分器同名指标及可见动作相符。"
+     },
+     {
+      "t": 1.08,
+      "log_step": 1069,
+      "text_zh": "观察：视频 0.96–1.12 s 机器人落地后躯干逐渐抬起，骨盆仍明显在脚后，末段向后倾；交付视频随即结束，没有连续站稳 2.0 秒的片段。log step 1069 回放交付文件仅执行 115 个控制步，success=false、done=false、failed=null、episode_over=false；评分器同样记录 stages_reached=3/4、fell=0。推断：完整的“后空翻并落地站稳”从未完成，原生三阶段进度不等于这个合并节点已达到。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–约 0.36 s 机器人双脚支承蓝色场地，屈腿后蹬地起跳；腾空期间脚离地，视频约 0.92 s 起双脚先接触场地并承重，末段脚掌抬起、主要在脚跟附近支承。双臂摆动与前伸，未见手臂触地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "value": 310.9,
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better"
+  },
+  {
+   "name": "air_s",
+   "value": 0.59,
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 1.08,
+   "log_step": 1069,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 0.92 s 机器人脚先落地；视频 1.04–1.12 s 双脚留在前方、骨盆后移，机器人以脚跟附近为支点向后倾，交付轨迹在落地恢复过程中结束。log step 1069 对该文件的回放与评分器均只有起跳、完整转体和落地三阶段，尚无站稳成功；该连续后翻动作未被修复完成。"
+  },
+  "unmet": {
+   "backflip_landed_stable": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.08,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.96–1.12 s 机器人落地后双脚基本留在前方，骨盆持续退到脚后，脚掌翘起、围绕脚跟附近后仰；视频 1.08–1.12 s 躯干抬起时整个人继续向后转动。推断：这是落地后的后向失衡，未形成稳定站姿；交付片段中尚未出现倒地。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "log_step": 1075,
+    "evidence_zh": "观察：log step 1026 的交付生成脚本用固定的关节目标与分段增���控制落地，仅运行 22 个落地后控制步即保存；log step 1069 已验证该文件为 115 步、三阶段、尚未成功。log step 1075 使用同一组起跳、收腿和落地控制参数做延长测试，站立计数仅短暂从 1 增至 3，接着归零；log step 1088 的同参数测试中躯干由前倾经过直立后继续后转。推断：落地控制未能制止后向俯仰并维持承重平衡，解释了交付末段的后倾与未站稳；延长测试只作为控制根因证据，不计作交付片段内的后续运动。具体是轨迹、增益还是反馈律不足，证据不足以细分。"
+   }
+  ],
+  "gap_zh": "无模拟器状态日志，25 帧/秒录像不能精确确定瞬时接触及姿态、速度阈值。交付视频仅 1.16 秒、最后可见帧为视频 1.12 s，落地后的完整站稳窗口未交付；末段可见后倾失衡，但没有看到倒地，且评分器 fell=0。log step 1075、1088 的延长测试属于会话诊断，不能把其中超出交付长度的运动补入本次轨迹；控制根因子类保持未细分。"
+ },
+ "summary_zh": "机器人完成脚支撑起跳、空中后翻和脚先落地，评分器记录转角 310.9°、腾空 0.59 秒及原生阶段 3/4，但合并节点“后空翻并落地站稳”最终为 0/1。视频约 0.92 s 落地后机器人向后倾，落地控制没有建立持续稳定站姿，交付文件仅含 115 个控制步，未完成站稳验收。无状态日志，交付末段尚未倒地；会话延长测试仅用于支持控制根因，不作为本次轨迹的后续进度或摔倒证据。",
+ "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（final=false）、0.4 s 起跳、0.92 s 脚先落地、末段后倾，均与视频及 log step 1069 回放（3/4 阶段、310.9°、0.59 s、115 步）相符；断点、平衡现象和控制根因有证据支持，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": "2026-10-31T00:34:23Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--njqpcrg/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--njqpcrg/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f5476814692387cb0dc50623a200a7f38c9a5dbd5f202c1c5a30caacfbeab3ff",
+   "video_sha256": "503d4c80c10bd8e4fd12b2bc5aa7963c1bbbd91c2007493743ca6e363713e12b",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-privileged--o3j4nwb.json

@@ -0,0 +1,179 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-privileged--o3j4nwb",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 2000,
+   "sim_steps": 477,
+   "horizon": 2000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 4,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 3.7,
+   "bad_trajectory": 0,
+   "metric_air_s": 0.917,
+   "metric_turn_deg": 330.2,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 125,
+   "agent_resets": 16780,
+   "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/embodied-olympics-fun-backflip-i00-privileged--o3j4nwb/replay.mp4",
+  "duration_s": 4.8,
+  "fps": 25.0,
+  "frames": 120,
+  "sha256": "6410c317d20d7bb748f1fcab5e338a4b4490efdd6224bcfa2e75e7f93fe67e67"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.76,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.68,
+      "log_step": null,
+      "text_zh": "观察：视频0.56–0.68 s，机器人从深蹲伸腿、上摆双臂，双脚为地面支撑；视频0.72 s身体已与蓝色地面分离，视频0.80–1.28 s头向后、脚向前翻过倒立位置。"
+     },
+     {
+      "t": 1.6,
+      "log_step": null,
+      "text_zh": "观察：视频1.52 s机器人仍在空中，视频1.56–1.60 s双脚首先落到蓝色地面，随后屈膝缓冲；视频1.68–1.72 s深蹲时未见手、膝或躯干撑地，视频1.84–2.32 s又有回弹和姿态调整，不能在首次落脚时就计作完成站稳验收。"
+     },
+     {
+      "t": 4.76,
+      "log_step": null,
+      "text_zh": "观察：视频2.72–4.76 s机器人已直身，以脚承重，位置变化很小；逐帧检查视频3.76–4.76 s，末尾仍直立，未见倒地。inputs/run.json的reward记录success=1、replay_success=1、deterministic=1、stages_reached=4/4、fell=0，原生指标metric_turn_deg=330.2、metric_air_s=0.917。"
+     },
+     {
+      "t": 4.76,
+      "log_step": 87,
+      "text_zh": "观察：log step 77将选定参数生成最终trajectory.npz并回放成功；log step 87重新生成的文件逐字节一致，对此文件的两次独立验证回放均输出success且终态哈希相同，指标为后转330.2°、腾空0.917秒。后续log step 88–93是额外扰动搜索与说明更新，未见替换已验证的交付轨迹。"
+     },
+     {
+      "t": 4.76,
+      "log_step": null,
+      "text_zh": "推断：视频中的完整后翻、脚先落地和末尾持续站立与本次评分器四阶段成功一致，节点最终成立；将视频4.76 s末帧作为完成时间的近似锚点，误差约0.2秒。录像没有裁判连续站稳计数或骨盆速度，不能从画面独立精确定位200次采样验收的首个时刻；该数值验收由本次评分器成功结果支持，未把首次落地或开局站立当作节点完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：机器人双脚在视频0.56–0.68 s抵住蓝色地面起跳发力，视频0.72–1.52 s后翻腾空，视频1.56–1.60 s脚先落地承重，随后以双脚支撑恢复直立并站稳；未见双手撑地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": 330.2
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": 0.917
+  }
+ ],
+ "failure": null,
+ "summary_zh": "本次评分轨迹完成唯一节点“后空翻并落地站稳”：视频约0.72 s离地后翻、约1.56–1.60 s脚先落地，经回弹调整后直立至视频4.76 s末帧，最终进度1/1。评分器记录后转330.2°、腾空0.917秒、四阶段全部通过且未摔倒，最终文件在log step 87的两次验证回放均成功且终态一致。缺少状态日志和逐帧裁判计数，速度、倾角和连续站稳窗口由成功判定支持，节点完成时间以视频末帧近似标注。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "单节点成功轨迹：视频0.6–1.3 s起跳后翻、1.56 s脚先落地、1.7 s深蹲、2.4–4.76 s直立，与评分器success=1、四阶段4/4、fell=0一致；节点值、时间锚点、contact_actual（腿/躯干 支撑）和无失败层均成立。",
+    "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": "2026-10-31T00:32:23Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--o3j4nwb/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--o3j4nwb/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "4501bed864a98761237c0e013e41f4028950bfe15c885d8734929f76cdbc2167",
+   "video_sha256": "6410c317d20d7bb748f1fcab5e338a4b4490efdd6224bcfa2e75e7f93fe67e67",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-privileged--prucqee.json

@@ -0,0 +1,228 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-privileged--prucqee",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.25,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 432,
+   "sim_steps": 52,
+   "horizon": 2000,
+   "episode_over": 1,
+   "progress": 0.25,
+   "stages_reached": 1,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 2.64,
+   "bad_trajectory": 0,
+   "metric_air_s": 0.271,
+   "metric_turn_deg": -101.9,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 201,
+   "agent_resets": 7945,
+   "agent_replays": 0,
+   "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": 3.07,
+   "peak_processes": 19,
+   "peak_disk_gb": 0.24,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-privileged--prucqee/replay.mp4",
+  "duration_s": 0.56,
+  "fps": 25.0,
+  "frames": 14,
+  "sha256": "d0c47c93ac9ef36a4e15538fa7d3892465b117bc800f8250b21020501f7503b6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.08 s，机器人双脚站在蓝色场地上，尚未执行空翻；初始站立不满足后空翻并落地站稳的完整条件。"
+     },
+     {
+      "t": 0.4,
+      "log_step": null,
+      "text_zh": "观察：逐帧查看视频 0.00–0.52 s 的全部14帧及末段裁剪图；机器人先屈腿、躯干前倾，再短暂离地，视频 0.36–0.52 s 头和躯干向前俯冲，手臂前伸，最终身体接近水平并贴近蓝色场地，未出现完整后翻和落地直身站稳。"
+     },
+     {
+      "t": 0.52,
+      "log_step": null,
+      "text_zh": "观察：inputs/run.json 的评分器字段为 stages_reached=1/4、metric_air_s=0.271、metric_turn_deg=-101.9、success=0、fell=0、deterministic=1；仅起跳阶段得到确认，负转角及视频中向前俯下的动作均不满足累计后转300°。"
+     },
+     {
+      "t": null,
+      "log_step": 183,
+      "text_zh": "观察：agent 在 log step 183 自述交付轨迹约后转100°且未成功；但这只是会话总结，评分器实际报告转角为−101.9°，视频也未显示后翻。节点按实际评分回放标为未完成，不采用该自述的旋转方向。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.28 s 机器人双脚接触蓝色场地并承重，随后短暂腾空；此条记录脚与地面的起跳支撑。"
+     },
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频 0.44–0.52 s 机器人向前翻倒，手臂前伸，躯干及上肢贴近蓝色场地，末段并非稳定的脚支撑落地。推断：手臂或躯干可能触地；无状态接触日志，25帧/秒画面不能确认末段首次触地的部位及承重方式，故此段接触标未知。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": -101.9
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": 0.271
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 0.48,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.40–0.52 s，机器人在本次跳跃中向前俯冲并翻倒，身体最终接近水平、手臂前伸贴近地面；评分器记录腾空0.271秒、转角−101.9°且只到达起跳阶段，后空翻并落地站稳这一步未完成。"
+  },
+  "unmet": {
+   "backflip_landed_stable": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 0.48,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.36–0.52 s 机器人头和躯干向前俯冲，手臂向前伸出，最终身体接近水平并趴向蓝色场地，未维持脚支撑站姿，也未完成后翻；这是可见的向前翻倒。inputs/run.json 的评分器记录 metric_turn_deg=-101.9、fell=0；fell=0 是裁判记录，不能据此把画面中的翻倒改记为无失衡的跳跃不足，末段首次���地部位仍无法精确确认。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 158,
+    "evidence_zh": "推断：agent 的控制搜索评价在 log step 158 用 attempt=dict(turn=abs(t), ...) 将正负转角都计为正收益；log step 163 随后以自定义评价 max_turn_deg=100.0 保存432行轨迹，log step 176 再次生成同样行数的交付文件。实际评分器 metric_turn_deg=-101.9，与视频 0.40–0.52 s 的前俯动作一致，支持控制轨迹及转角评价实现不当；log step 173 将交付自述为“后转100°”并不能消除该符号错误。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "观察：本任务没有提供状态日志；已检查 inputs/ 中的完整日志、全部14帧评分视频及成功参考视频。25帧/秒画面不足以精确判断末段首次触地部位、关节力矩和角动量，不能将 agent 在 log step 183 自述的起跳角动量数值当作评分轨迹事实；自定义测试的阶段结果（log step 163 全部为假）也与实际评分器确认起跳不同，不能用于节点时间。控制根因保留在大类，不进一步指定反馈、逆解或某个参数。"
+ },
+ "summary_zh": "机器人进入起跳与短暂腾空，随后向前俯冲翻倒，没有完成后空翻并落地站稳，语义节点最终为0/1；评分器仅确认原生起跳阶段，腾空0.271秒、转角−101.9°、fell=0。断点在空翻操作中，现象为全身失衡，日志中的绝对值转角评价支持控制实现不当；agent 自述的“后转约100°”不符合实际评分回放。缺少状态日志，无法精确判定末段首次触地部位及力矩原因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 7,
+    "summary_zh": "节点0/1、断点 backflip_landed_stable、控制根因基本有据；但现象与接触读错：视频 0.40–0.52 s 机器人向前扑倒，躯干/手臂伸向前方、整体趴到场地上，不是“屈腿低姿态回落”，应记 全身·平衡 而非 步态，并补充手/躯干触地的 contact_actual。",
+    "n_issues": 2,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点0/1、断点 backflip_landed_stable（操作中）、全身·平衡现象与控制根因均有帧和评分字段（turn=-101.9°、stages 1/4）支持；时钟均为视频时间，contact_actual 对末段如实标未知。无需修改。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:48:24Z",
+  "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/embodied-olympics-fun-backflip-i00-privileged--prucqee/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--prucqee/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--prucqee/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--prucqee/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9c4a1f38314cb40c3951bc120c384edfec88b757f4e9339a772c12d5a91c248a",
+   "video_sha256": "d0c47c93ac9ef36a4e15538fa7d3892465b117bc800f8250b21020501f7503b6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-privileged--zgfqyyv.json

@@ -0,0 +1,186 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-privileged--zgfqyyv",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 900,
+   "sim_steps": 400,
+   "horizon": 2000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 4,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 5.07,
+   "bad_trajectory": 0,
+   "metric_air_s": 0.609,
+   "metric_turn_deg": 319.5,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 2,
+   "agent_resets": 5,
+   "agent_replays": 14,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 3.223,
+   "peak_processes": 23,
+   "peak_disk_gb": 0.009,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-privileged--zgfqyyv/replay.mp4",
+  "duration_s": 4.04,
+  "fps": 25.0,
+  "frames": 101,
+  "sha256": "c6cdc0d1b6512aab31a3e99843b185af97c885d875d44fa2d0408023317a0d2d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.0,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00 s 机器人双脚站在蓝色场地上，尚未完成后翻；视频 0.28–0.60 s 机器人屈膝下蹲、向后倾身并蹬地。开局站立不满足包含后翻历史和落地验收的完整节点。"
+     },
+     {
+      "t": 0.64,
+      "log_step": null,
+      "text_zh": "观察：视频 0.64–1.16 s 机器人身体与地面分离，头向后、双脚向前翻转；视频约 0.88–0.96 s 机器人经过头朝下的倒立姿态，随后展开身体使脚朝向地面。"
+     },
+     {
+      "t": 1.2,
+      "log_step": null,
+      "text_zh": "观察：视频 1.16 s 机器人双脚仍在地面上方，视频 1.20–1.32 s 可见脚先承重、双臂高举，手、膝和躯干均离地；视频 1.40–1.80 s 机器人屈膝缓冲并重新伸直身体。首次落地约在视频 1.20 s，按录像帧率估计。"
+     },
+     {
+      "t": 4.0,
+      "log_step": null,
+      "text_zh": "观察：视频约 2.00 s 起机器人恢复近直立，双脚支撑且身体基本停稳；逐帧检查视频 2.96–4.00 s，机器人保持该姿态至片尾，没有可见倒地或手、膝撑地。"
+     },
+     {
+      "t": null,
+      "log_step": 123,
+      "text_zh": "观察：log step 113 保存最终候选并回放，log step 119、123 再次核验同一交付轨迹，输出 success=True、四阶段 take-off/full turn/landing/stand 全部完成、两次回放终态哈希相同。inputs/run.json 的评分器字段 success=1、replay_success=1、deterministic=1、stages_reached=4/4、fell=0，metric_turn_deg=319.5、metric_air_s=0.609。"
+     },
+     {
+      "t": 4.0,
+      "log_step": 124,
+      "text_zh": "推断：log step 124 渲染最终交付轨迹，记录 executed_rows=400、success=true 和 101 帧，与本次视频一致；log step 36 的生成器控制间隔为 0.01 s，因此第 400 个控制步对应视频约 4.00 s。已用视频约 0.60–1.20 s 的腾空及视频 4.00 s 的直立终帧核对：前者与评分器记录的 0.609 s 腾空相符，后者与回放结束相符。完整节点需计入连续站稳验收，首次达成锚定视频约 4.00 s，非首次落脚时刻；数值阈值及连续采样通过由评分器成功结果确认。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.28–0.60 s 机器人双脚在蓝色场地上支撑蹬地；视频 0.64–1.16 s 腾空时与地面分离；视频约 1.20 s 双脚先落地承重，随后双脚支撑屈膝缓冲、起身并站立至视频 4.00 s。双臂摆动协调，未见手部撑地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "value": 319.5,
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better"
+  },
+  {
+   "name": "air_s",
+   "value": 0.609,
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "观察：本次评分录像显示机器人从双脚支撑起跳、空中后翻经过倒立，视频约 1.20 s 脚先落地，随后缓冲��身并保持近直立至视频 4.00 s；评分器记录后转 319.5°、腾空 0.609 s、四阶段全达成和回放成功。推断：结合最终交付轨迹的 400 个控制步及片尾姿态，“后空翻并落地站稳”节点在视频约 4.00 s 完成，最终为 1/1，无失败断点。观察：本次没有状态日志，25 帧/秒录像不能独立核验毫秒级触地先后、精确速度和连续站稳计数，这些数值验收依评分器结果确认。",
+ "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：起跳、倒立、脚先落地、站稳至片尾；成功节点在约4.00 s（400控制步=结束）完成，接触与总结无误，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": "2026-10-31T00:42:15Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--zgfqyyv/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-privileged--zgfqyyv/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "cfdf841758b25a746963ebfa64fe5e21cccd88d742da713a064ddfbfd579a199",
+   "video_sha256": "c6cdc0d1b6512aab31a3e99843b185af97c885d875d44fa2d0408023317a0d2d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-standard--7u2lrch.json

@@ -0,0 +1,182 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-standard--7u2lrch",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "live_success": 1,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 361,
+   "sim_steps": 1131,
+   "calls": 1142,
+   "rows": 1131,
+   "resets": 4,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 4755.8,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 4.394,
+   "peak_processes": 22,
+   "peak_disk_gb": 0.332,
+   "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/embodied-olympics-fun-backflip-i00-standard--7u2lrch/replay.mp4",
+  "duration_s": 3.64,
+  "fps": 25.0,
+  "frames": 91,
+  "sha256": "5acfa900a1e624b244db37d4ce7f7f57d5ba61cc8b0a1f029e04cfbbad493a47"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [
+     [
+      3.6,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频0.00–0.60 s，机器人从双脚站立转为屈膝下蹲、摆臂蹬地；初始站立尚未包含合法后翻及落地历史，因此节点不从开局成立。"
+     },
+     {
+      "t": 0.84,
+      "log_step": null,
+      "text_zh": "观察：视频0.68–1.12 s，机器人全身与蓝色场地分离，收腿向后翻转，视频0.76–0.84 s经过倒立姿态；视频1.16 s双脚已落地承重，随后屈膝吸收冲击，未见手、膝或躯干撑地。"
+     },
+     {
+      "t": 3.6,
+      "log_step": null,
+      "text_zh": "观察：视频1.60–3.60 s，机器人仅由脚支撑，双臂前伸、躯干近竖直，位置基本稳定；逐帧检查视频2.60–3.60 s，直到末帧仍站立，没有翻倒或其他身体部位触地。"
+     },
+     {
+      "t": 3.6,
+      "log_step": 116,
+      "text_zh": "观察：log step 116的第5次实时尝试返回“DONE 360 success success 4 / 4”，服务状态为episode=5、control_steps=361、done=true、reason=success；log step 118再次查询得到success、stages_reached=4。inputs/run.json评分字段success=1、live_success=1、graded_by_live=1、finalized=1，确认评分的是这次成功尝试。"
+     },
+     {
+      "t": 3.6,
+      "log_step": 118,
+      "text_zh": "推断：结合视频中后翻、脚先落地及持续站立，与第5次尝试的原生成功终态，将包含连续站稳窗口的完整节点首次达成锚定在视频约3.60 s末帧，而非首次落地时。没有状态日志，精确转角、速度、倾角及连续200次采样由该次原生裁判成功结果确认，无法从25帧/秒录像独立复算。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频0.00–0.64 s，机器人双脚在蓝色场地上支撑、屈膝蹬地；视频0.68–1.12 s腾空无地面接触；视频1.16–3.60 s双脚落地承重并保持站立。双臂摆动协调，未见上肢接触场地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": null,
+ "summary_zh": "评分尝试为log step 116执行并由log step 118确认的第5次实时尝试：机器人在视频约0.68 s离地后翻、约1.16 s脚先落地，随后站稳至视频约3.60 s，完整节点完成1/1，原生裁判成功且无失败断点。会话中此前的失败测试不属于这段评分视频；成功参考录像也呈现后翻、脚落地再站稳的顺序。没有状态日志，视频无法独立核实精确转角、速度及采样窗口，相关数值条件依据本次原生成功结果确认；评分字段未报告turn_deg、air_s，二者保留null。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "评分视频为第5次实时尝试（361控制步≈3.64 s）：0.68 s起离地、视频约0.8 s倒立、约1.16 s脚先落地并站稳至末帧，节点在约3.60 s达成、无失败断点、contact_actual（腿/躯干·支撑）与log step 116/118的服务读数一致；未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:39:10Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--7u2lrch/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--7u2lrch/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "a990f725f29ff2de7421f7820f8fc128cb1afebfa631625655a1be4877c36090",
+   "video_sha256": "5acfa900a1e624b244db37d4ce7f7f57d5ba61cc8b0a1f029e04cfbbad493a47",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-standard--ccgnbbs.json

@@ -0,0 +1,190 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-standard--ccgnbbs",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 123,
+   "sim_steps": 184,
+   "calls": 27,
+   "rows": 184,
+   "resets": 1,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 1975.5,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 3.326,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.318,
+   "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/embodied-olympics-fun-backflip-i00-standard--ccgnbbs/replay.mp4",
+  "duration_s": 1.24,
+  "fps": 25.0,
+  "frames": 31,
+  "sha256": "71d31056cd1ef8025bc1bfeecb2196177e74f91684107459a97d7daf3a6e1713"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.6,
+      "log_step": 67,
+      "text_zh": "观察：视频 0.00–0.52 s 机器人由双脚站立下蹲并摆臂，视频 0.60–1.12 s 离地后向后翻过倒立位置；视频 1.16 s 双脚落地，视频 1.20 s 躯干已明显前倾、双手下探，没有落地后连续站稳。"
+     },
+     {
+      "t": 1.2,
+      "log_step": 71,
+      "text_zh": "观察：log step 71 的机器人实时控制服务（执行关节命令并返回原生裁判状态的接口）报告 episode=2、done=True、outcome=fell、stages_reached=3、stages_total=4；该局前三阶段通过，站稳阶段未通过。log step 67 重置后执行119个10毫秒控制步，log step 71 再执行4步结束，共123步，与这段视频的下蹲、翻转、落地前倾及约1.24秒总长一致；这些是最终评分局的服务结果，而非本地模拟结果。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.56 s 机器人双脚支撑蓝色场地以完成下蹲与起跳；腾空后，视频约 1.16–1.20 s 双脚再次接地承重，但躯干前倾，未稳住。最后画面双手接近地面，不能确认具体非脚触地部位，未另记手部接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 1.16,
+   "log_step": 71,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.12 s 机器人仍在下降，视频 1.16 s 脚着地时膝屈、身体后仰，至视频 1.20 s 躯干迅速向前倒、双臂下垂，未恢复直立；log step 71 的最终实时裁判结果为 fell、3/4阶段。失败发生在后空翻动作的落地稳定过程中。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.2,
+    "log_step": 71,
+    "evidence_zh": "观察：视频 1.16–1.20 s 机器人脚着地后躯干由后仰快速转为近水平前倾，头与双手向地面下降；log step 71 原生裁判报告 outcome=fell，尚未完成站稳验收。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 67,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 67 的 agent 脚本依据本地模拟器状态一次性生成并向实时机器人提交119行 lowcmd，实际执行期间没有根据服务返回的机器人姿态更新落地命令；log step 71 随后只提交固定站姿关节目标，执行4个控制步即得到 fell。结合视频 1.16–1.20 s 落地后快速前倾，说明该动作序列及落地接续控制未能消除落地姿态与运动偏差；证据不足以区分轨迹、反馈或参数的具体问题，保留上位标签。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 中没有状态日志；视频最后一帧为视频 1.20 s，原生裁判在 log step 71 报告仿真 1.23 s 摔倒，故不能从画面独立确认最终是哪一非脚部位先触地。评分字段和最终服务结果未报告 turn_deg、air_s 数值，均留空，不把本地模拟器的角度当作本次真实指标；控制根因的具体子类无法确定。log step 80 的模型过滤发生在 log step 71 实时局已经结束判负之后，不改变对该物理失败的标注。",
+  "unmet": {
+   "backflip_landed_stable": false
+  }
+ },
+ "summary_zh": "评分局完成起跳、空中后转和脚先落地三个原生阶段，但未完成落地站稳，唯一节点始终未达成，最终为0/1。视频 1.16–1.20 s 机器人落地后迅速前倾，log step 71 裁判判定摔倒；提交的动作序列与固定站姿接续未能稳住机器人，根因归为控制。没有状态日志，具体非脚触地部位及转角、腾空时长数值无法核实；log step 80 的过滤发生在本局判负之后。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标签与视频和日志一致：评分局视频 1.16–1.20 s 脚落地后躯干前倾，log step 71 服务报告 outcome=fell、3/4 阶段，节点未达成、断点和平衡/控制标签有证据支持，未发现需改动之处。",
+    "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": "2026-10-31T00:41:56Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--ccgnbbs/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--ccgnbbs/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "fa6ed3f42508da19d33b17dcf5d155201cff2b671ae661af38b7ef521da78d97",
+   "video_sha256": "71d31056cd1ef8025bc1bfeecb2196177e74f91684107459a97d7daf3a6e1713",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-standard--dv6r3pr.json

@@ -0,0 +1,196 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-standard--dv6r3pr",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 127,
+   "sim_steps": 5916,
+   "calls": 1075,
+   "rows": 5916,
+   "resets": 50,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 5040.1,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.261,
+   "peak_processes": 15,
+   "peak_disk_gb": 0.383,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-standard--dv6r3pr/replay.mp4",
+  "duration_s": 1.28,
+  "fps": 25.0,
+  "frames": 32,
+  "sha256": "cf5b8abad0d9bc56cc45374caa31b22f2250e439bbe50e5da2cec724db68058d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.72,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查本次视频 0.00–1.24 s 的全部32帧；机器人先以双脚支撑下蹲，视频约 0.72 s 完全离地，随后头向后、脚向前翻过倒立位置，视频 1.16–1.20 s 脚部先接地。"
+     },
+     {
+      "t": 1.24,
+      "log_step": 224,
+      "text_zh": "观察：视频 1.20–1.24 s 机器人落地后未直身停稳，骨盆继续下降、身体向后倾倒；log step 224 的最后一次实时尝试在 reset 后返回 resets_left=0、done=True、reason=failed、outcome=fell、stages=3/4。该日志结果与视频动作一致，未出现后翻后连续站稳的验收阶段。"
+     },
+     {
+      "t": 1.24,
+      "log_step": 237,
+      "text_zh": "推断：log step 237 的 agent 最终报告明确称最后一次尝试为评分局，与 log step 224 的服务结果及本次短视频一致，因此此前50次重置中的测试不计入节点进度。成功参考 inputs/oracle/replay.mp4 在参考视频约 1.12 s 脚落地后恢复直立，并在参考视频约 2.08–4.12 s 持续站立；本次视频没有出现这一完成表现，节点全程未满足。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.68 s 机器人双脚与蓝色地面接触，承重下蹲并蹬地起跳；视频约 0.72–1.12 s 腾空，无地面接触；视频 1.16–1.20 s 脚部重新接地承重，随后身体下沉失衡。未将末帧难以分辨的非脚部接触记作确定接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better"
+  },
+  {
+   "name": "air_s",
+   "value": null,
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 1.2,
+   "log_step": 224,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 1.16–1.20 s 机器人后翻后脚先落地，至视频 1.24 s 身体后倾、骨盆下沉，未恢复站稳；log step 224 的最后一次实时尝试由基准机器人服务返回 outcome=fell、stages=3/4，后续日志没有新的实时尝试。失败发生在后空翻的落地承重阶段。"
+  },
+  "unmet": {
+   "backflip_landed_stable": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.2,
+    "log_step": 224,
+    "evidence_zh": "观察：视频 1.16–1.24 s 机器人从空中后翻转为脚部承重，但身体继续后倾、骨盆快速下沉，未能建立站立姿态；同一最终尝试的服务结果在 log step 224 为 fell、stages=3/4。故为落地失衡，不另重复记录摔倒约束。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 224 显示 agent 的控制脚本以转角跟踪阈值220触发伸腿，并将伸腿目标调整为髋关节-0.60、膝关节0.80；服务确认该最终尝试达到3/4阶段后摔倒。视频 1.16–1.24 s 同时显示脚先接地后未控制住身体下沉和后倾，说明后翻控制未能完成落地稳定；只能归到控���，不能凭先前测试的分析确定本次的关节动量、增益或反馈子类。",
+    "source_zh": "日志 / 视频",
+    "log_step": 224,
+    "t": 1.2
+   }
+  ],
+  "gap_zh": "没有状态日志，视频为25帧/秒且在视频 1.24 s 的最后一帧仍为下沉姿态，不能独立区分服务判定 fell 的具体触发项（骨盆低于阈值或非脚部触地），也不能精确核验空中转角、速度和接触采样。最后一次尝试未输出完整落地本体传感器轨迹；log step 221 的详细轨迹属于此前测试，故不据此量化本次失衡机制。评分字段未报告 turn_deg、air_s，两个原生指标保留null。"
+ },
+ "summary_zh": "评分局完成腾空后翻并脚先接地，但在视频约 1.20–1.24 s 落地后失衡下沉；log step 224 的服务结果为 fell、stages=3/4，因此唯一节点“后空翻并落地站稳”始终未完成，最终0/1。根因按可见落地失稳及控制日志归为控制；无状态日志和最终尝试的完整本体传感器轨迹，无法进一步确定具体失稳机制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "唯一节点未完成、断点在落地稳定、平衡现象与控制根因均有视频和服务结果（log step 224：fell、stages=3/4）支持；未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:41:16Z",
+  "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/embodied-olympics-fun-backflip-i00-standard--dv6r3pr/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--dv6r3pr/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "5244f6d32cbf7a3d8f69a4947eece56f7d930ee06a13570b960bbb85e3d79bf6",
+   "video_sha256": "cf5b8abad0d9bc56cc45374caa31b22f2250e439bbe50e5da2cec724db68058d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-standard--gqxoge4.json

@@ -0,0 +1,211 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-standard--gqxoge4",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 84,
+   "sim_steps": 4180,
+   "calls": 156,
+   "rows": 4180,
+   "resets": 50,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 2744.2,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.504,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.834,
+   "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/embodied-olympics-fun-backflip-i00-standard--gqxoge4/replay.mp4",
+  "duration_s": 0.88,
+  "fps": 25.0,
+  "frames": 22,
+  "sha256": "939a5cd8e6782f9720258cfccce71137b9e922c8c983f591992a8eb5f54f2163"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.56,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频 0.00–0.84 s；机器人先以双脚支撑蹲身、后仰和摆臂，视频 0.56 s 双脚已离开地面，随后收腿向后翻转。开局站立不满足完成后空翻的历史条件。"
+     },
+     {
+      "t": 0.84,
+      "log_step": null,
+      "text_zh": "观察：视频 0.72–0.84 s 机器人仍蜷曲翻转并下降；视频 0.84 s 末帧双脚仍抬起，未见脚先落地或直立站稳，随后录像结束。末帧不足以确认是否已触地及首次触地部位。评分字段 success=0、live_success=0，与未完成节点的录像表现一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.52 s 机器人双脚接触蓝色场地，蹲身并伸腿起跳；视频 0.56–0.80 s 可见身体腾空，视频 0.84 s 仍蜷曲且双脚抬起，未见双脚落地承重；末帧是否已触地及首次触地部位未知。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 0.84,
+   "log_step": 168,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.56–0.84 s 机器人起跳后蜷身后翻并下降，末帧双脚仍抬高，未见脚先落地和站稳，录像在此结束；末帧不能确认触地部位。log step 168 是最后一次重置并执行控制器；agent 在 log step 169、171 自述最后一次尝试未完成后转并摔倒，评分字段 success=0、live_success=0 确认该尝试失败，具体摔倒时刻仍仅有自述。"
+  },
+  "unmet": {
+   "backflip_landed_stable": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 0.84,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.76–0.84 s 机器人蜷曲下降，末帧双脚仍抬起，后翻尚未形成脚朝下的落地姿态，没有出现落地后的脚支撑或站稳，随后录像结束；是否已经发生非脚部触地不能由末帧确认。推断：结合评分失败与这段下降姿态，保留全身平衡失败，不能表述为已落地后才失衡。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 169 展示最终控制器按固定时长切换蹲身、伸腿及收腿关节目标，以 lowcmd PD 执行；视频 0.56–0.84 s 显示机器人起跳后蜷身后翻，录像结束时仍未形成脚朝下的落地姿态，也未站稳。该关节控制序列未产生完成翻转并安全落地的全身运动；现有证据不足以进一步区分轨迹、跟踪或具体参数问题。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志；最终执行调用及状态查询的完整服务返回未保留在 run.json 中。agent 在 log step 169–171 自述约253°转角、约0.42秒腾空及摔倒和阶段结果，不能当作原生评分数值；reward 未报告 turn_deg 或 air_s，故两项为 null。25帧/秒的侧视录像显示末帧仍未建立双脚落地支撑或站稳，但不能确认末帧是否已经触地及首次触地部位，不能精确核验接触的控制采样或累计骨盆转角，也不足以确定控制根因的子类。"
+ },
+ "summary_zh": "最后一次评分尝试完成蹲身和起跳并在空中蜷身后翻，但未见脚先落地；视频 0.84 s 末帧双脚仍抬起、未站稳，后空翻并落地站稳节点始终未达成，终态为0/1。失败发生在翻腾操作中，保留全身平衡现象及控制根因，但不能确定控制子类或称为落地后失衡。末帧触地部位未知，无状态日志或原生转角与腾空时长；agent 在 log step 169–171 自述约253°、未达300°及摔倒，仅作自述保留。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "单节点未达成、断点 backflip_landed_stable、终态 0/1 均与录像和评分字段一致；但「0.84 s 上半身触地」不是帧上可见的观察，且摘要把空中后翻写成已完成，与 agent 自述转角约253°<300° 的失败原因不符。",
+    "n_issues": 2,
+    "n_major": 0
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点未达成、断点 backflip_landed_stable、平衡现象与控制根因均与录像一致；末帧触地部位已标为未知，自述数值已标为自述，未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:50:39Z",
+  "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/embodied-olympics-fun-backflip-i00-standard--gqxoge4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--gqxoge4/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--gqxoge4/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--gqxoge4/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "e6717635d13da1f8bdbafe1afce4ef4f0dace8e11ada718c82b5cb017ff97181",
+   "video_sha256": "939a5cd8e6782f9720258cfccce71137b9e922c8c983f591992a8eb5f54f2163",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-standard--jceja9k.json

@@ -0,0 +1,202 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-standard--jceja9k",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 332,
+   "sim_steps": 10084,
+   "calls": 1670,
+   "rows": 10064,
+   "resets": 50,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 3946.5,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 1,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 12.314,
+   "peak_processes": 41,
+   "peak_disk_gb": 1.101,
+   "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/embodied-olympics-fun-backflip-i00-standard--jceja9k/replay.mp4",
+  "duration_s": 3.36,
+  "fps": 25.0,
+  "frames": 84,
+  "sha256": "f04646f59803a7d5d0d0537abb4f18ccd1aa8eab5888a75d4182b76b87926a82"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.52,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.48 s 机器人以脚支撑下蹲、蹬地，视频 0.52–1.04 s 全身离地并向后翻过倒立姿态；视频 1.08 s 双脚先到达地面，手和膝仍离地。"
+     },
+     {
+      "t": 3.0,
+      "log_step": 437,
+      "text_zh": "观察：视频 1.20–2.88 s 机器人落地后有反弹、前倾和屈膝，没有形成持续直立停稳；视频 3.00–3.32 s 骨盆向脚后方下沉并向后坐倒，末帧可见近侧手部触地，完整的后翻并落地站稳条件从未成立。log step 437 的最后一次服务尝试返回 OUTCOME fell、stages 3/4、episode 51、control_steps 332；log step 448 再次查询仍是该终结状态，与 reward 的 n_actions=332、resets=50 一致。"
+     },
+     {
+      "t": 3.32,
+      "log_step": 437,
+      "text_zh": "推断：以 log step 8 的 control_dt=0.01 s 和 log step 437 的最后一局 control_steps=332 核对，控制步末为视频 3.32 s，正好对应所见向后坐倒的末帧；只将最后一局用于节点判断，不把其他重置局的阶段或本地模拟成功计入本次进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.48 s 机器人双脚接触蓝色地面承重并蹬地，视频 0.52–1.04 s 腾空无地面接触，视频 1.08 s 脚先落地，之后双脚承重但身体逐渐失衡。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "支撑",
+      "note_zh": "观察：视频 3.32 s 向后坐倒的末帧中，机器人近侧手部的固定拳垫触地撑住身体；远侧手被身体遮挡，不能确认是否同时触地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better"
+  },
+  {
+   "name": "air_s",
+   "value": null,
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 3.0,
+   "log_step": 437,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.08 s 机器人脚先落地，此后仍在完成同一次后翻的落地稳定阶段；视频 3.00–3.32 s 机器人明显向后坐倒、近侧手部触地，未恢复站立。log step 437 的最后一局服务裁判给出 fell、stages 3/4，未通过站稳阶段。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 3.0,
+    "log_step": 437,
+    "evidence_zh": "观察：视频 2.88–3.32 s 机器人保持屈膝，随后骨盆向脚后方移动并下降，身体向后坐倒，末帧近侧手部撑地；log step 437 的服务裁判终结原因为 fell。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "反馈控制",
+    "t": 3.0,
+    "log_step": 437,
+    "evidence_zh": "推断：log step 265 的控制器修改使落地恢复后保持设定的 hold 关节姿态；log step 428 记录的全身俯仰修正仅在 bal_k 非零时生效，而最后一局 log step 437 明确设置 bal_k=0、吸收和保持姿态相同。该局日志中控制器已处于 stand 阶段，关节位置基本维持不变，机器人却在视频 2.88–3.32 s 持续向后倾倒，说明落地后的全身姿态稳定控制未能抑制俯仰漂移并收敛到站稳状态。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志，无法独立核验毫秒级首次接触、骨盆累计转角、三维速度和连续站稳计数；reward 未报告 turn_deg、air_s 的数值，原生指标保留 null。log step 449 中 agent 自述“向前漂移”的总结涵盖多次会话尝试，而最后评分视频 3.00–3.32 s 显示向后坐倒，具体方向以视频为准；不能据此进一步确定某个增益或目标角度的数值错误。",
+  "unmet": {
+   "backflip_landed_stable": false
+  }
+ },
+ "summary_zh": "机器人完成可见的脚支撑起跳、空中后翻和脚先落地，最后一局服务裁判记录 stages 3/4；落地后未持续站稳，视频 3.00–3.32 s 向后坐倒并用近侧手部撑地，唯一节点未完成（0/1）。落地保持控制未抑制俯仰漂移，断点为“后空翻并落地站稳”的操作中。无状态日志且原生转角、腾空时长数值未报告，精确阈值依服务裁判结果辅助判断；agent 的会话总结不能替代最后一局视频。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "唯一节点未完成、断点在落地站稳、控制步计数 332 对应视频 3.32 s 的映射均与视频和 log step 437/448 一致；现象“平衡”与根因“反馈控制”（bal_k=0、落地后保持姿态）有日志支撑，未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:38:48Z",
+  "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/embodied-olympics-fun-backflip-i00-standard--jceja9k/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--jceja9k/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "30348a1bd6bdc577ed124e9c34349caad93fec1ecdc7042da2afafe9e044ab3d",
+   "video_sha256": "f04646f59803a7d5d0d0537abb4f18ccd1aa8eab5888a75d4182b76b87926a82",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-standard--oexfg2a.json

@@ -0,0 +1,182 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-standard--oexfg2a",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 20,
+   "sim_steps": 1794,
+   "calls": 40,
+   "rows": 1794,
+   "resets": 12,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 7145.6,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 3.07,
+   "peak_processes": 13,
+   "peak_disk_gb": 0.389,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-standard--oexfg2a/replay.mp4",
+  "duration_s": 0.24,
+  "fps": 25.0,
+  "frames": 6,
+  "sha256": "fa6db33957a15962340634eaaac2b5de7566950e1db0502adb559be1b50eaee0"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "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全部6帧（视频0.00、0.04、0.08、0.12、0.16、0.20 s）；机器人始终双脚站在蓝色场地上，没有可见的下蹲起跳、腾空、后转或翻后落地。与成功参考视频中清晰可见的腾空后翻及落地站稳相比，本次没有启动后空翻，开局站立不满足本节点。"
+     },
+     {
+      "t": 0.2,
+      "log_step": 1052,
+      "text_zh": "观察：log step 1052中agent脚本先调用实时机器人服务的reset，再提交20行stand动作（dt=0.01）；服务返回episode=13、control_steps=20、sim_time_s=0.2、done=False、calls=40、rows=1794。评分器reward同为calls=40、rows=1794、n_actions=20、live_success=0且graded_by_live=1，视频末帧仍为站立；本节点未完成。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better",
+   "value": null
+  },
+  {
+   "name": "air_s",
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 0.0,
+   "log_step": 1052,
+   "where": "首步前",
+   "evidence_zh": "观察：本次视频0.00–0.20 s全部帧只显示机器人站立，未出现后空翻的起跳动作。log step 1052的最后实时局由agent脚本重置后仅执行20行stand；后续log step 1053–1060没有再次向实时服务提交动作。该局停在后空翻启动前。"
+  },
+  "unmet": {
+   "backflip_landed_stable": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.0,
+    "log_step": 1052,
+    "evidence_zh": "观察：视频0.00、0.04、0.08、0.12、0.16及0.20 s中机器人保持原地双脚站立，未起跳或翻腾；log step 1052明确只提交stand动作。较早测试局的翻转和摔倒不在本次评分视频中。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "交付",
+    "item": "交付操作出错",
+    "source_zh": "日志 / 评分器字段 / 视频",
+    "evidence_zh": "推断：agent把实时评分局重置为站立诊断局后，没有再执行后空翻控制器，最终交付了该诊断局。依据：log step 1047曾向实时服务运行后空翻控制器，但该测试局返回fell；log step 1051及1052随后调用实时reset，log step 1052仅提交20行stand。此后log step 1053–1060只做本地模拟分析或读取spec；log step 1060的SUCCESS来自BackflipRunner本地模拟，并未提交实时服务。评分器graded_by_live=1、n_actions=20、calls=40、rows=1794与log step 1052的服务回执完全一致，视频0.00–0.20 s也只显示站立，证实最终受评的是站立测试局。"
+   }
+  ],
+  "gap_zh": "没有状态日志，不能精确复核足底接触、空中转角和速度等数值；但本次视频全部帧可见机器人始终站立，足以判定后空翻未启动。评分器没有报告turn_deg和air_s，原生指标留null；会话中其他局及本地模拟的指标不能替代本次评分局。"
+ },
+ "summary_zh": "本次评分视频全部6帧仅显示机器人原地站立，后空翻并落地站稳节点未达到，最终���度0/1。最后实时操作在log step 1052重置后只执行20行站立动作，后续本地模拟成功未提交实时服务，最终评分停留在站立诊断局，属于交付操作出错。无状态日志且评分器未报告转角和腾空时长，但不影响对未启动后空翻的判断。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "标签与证据一致：评分视频6帧全程站立，节点未达成，没尝试/首步前，交付操作出错有log step 1047、1052及评分器字段支撑。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:40:22Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--oexfg2a/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--oexfg2a/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "a2c85a5c8d8942d262bac3c2845785ae6f4784dd61bacb29d29eafeaf7fa3c40",
+   "video_sha256": "fa6db33957a15962340634eaaac2b5de7566950e1db0502adb559be1b50eaee0",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/backflip/trajectories/embodied-olympics-fun-backflip-i00-standard--pu6ny3q.json

@@ -0,0 +1,198 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-backflip-i00-standard--pu6ny3q",
+ "taskKey": "embodied-olympics-fun/backflip",
+ "dag_sha256": "a2dce89dd7f0391d67651e778f82fd48e6fb3121b94be3c140496cb25fc4241d",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 121,
+   "sim_steps": 720,
+   "calls": 13,
+   "rows": 720,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 6063.9,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.972,
+   "peak_processes": 25,
+   "peak_disk_gb": 0.324,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-backflip-i00-standard--pu6ny3q/replay.mp4",
+  "duration_s": 1.24,
+  "fps": 25.0,
+  "frames": 31,
+  "sha256": "b4060780855a98c493676052a6f51b2d8caaf6b06ee813ac27f9f09468e61e82"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "backflip_landed_stable": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.68,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.60 s，机器人从双脚支撑下蹲、蹬地并向后仰；视频 0.68–1.16 s 可见机器人离地、收腿后翻并展开，视频 1.20 s 脚部触地、膝部仍弯曲。整段视频到 1.24 s 结束，没有落地后连续站稳的过程。"
+     },
+     {
+      "t": 1.2,
+      "log_step": 136,
+      "text_zh": "观察：log step 136 中，agent 脚本重置后提交第 6 局动作，机器人服务返回 outcome=fell、stages_reached=3/4、done=True、reason=failed、control_steps=121；这些是服务裁判结果，表示已通过合法腾空、后转及脚先落地阶段，未通过站稳阶段。log step 139、141 的服务状态仍为 episode=6、control_steps=121、failed，第 7 局重置被拒绝，因此第 6 局是最终评分尝试。"
+     },
+     {
+      "t": 1.2,
+      "log_step": 136,
+      "text_zh": "推断：依据视频 1.20 s 的落地画面及 log step 136 的第 6 局裁判结果，机器人在落地时失衡判负，后空翻并落地站稳节点从未成立。起始直立未包含后翻历史，不计为本节点已完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.60 s，机器人双脚接触蓝色场地地面，承担下蹲与蹬地支撑；视频 0.68–1.16 s 为腾空阶段，视频 1.20 s 脚部重新接地承重。未观察到手部撑地，不据裁判的 fell 结果虚构具体非脚接触部位。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 300.0,
+   "direction": "higher_better"
+  },
+  {
+   "name": "air_s",
+   "value": null,
+   "unit": "s",
+   "threshold": 0.2,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "backflip_landed_stable",
+   "t": 1.2,
+   "log_step": 136,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.12–1.20 s，机器人翻身后展开双腿下降，脚部在视频 1.20 s 接地，未进入持续站稳。log step 136 的本局服务裁判返回 fell、3/4 阶段和 failed；log step 139、141 确认此后没有新局，故断点为后空翻动作的落地站稳阶段。"
+  },
+  "unmet": {
+   "backflip_landed_stable": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.2,
+    "log_step": 136,
+    "evidence_zh": "观察：视频 1.04–1.20 s 中机器人仍在快速翻身、展开和屈膝接地，末帧没有稳定直立停留；log step 136 的第 6 局服务裁判明确报告 outcome=fell、stages_reached=3/4，并结束本局。失衡由本局裁判结果确认；视频止于落地附近，没有拍到完整倒地过程。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 1.2,
+    "log_step": 137,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 31 的 agent 控制脚本预生成整段关节目标、增益和力矩序列并一次性提交 lowcmd；log step 136 用该脚本提交第 6 局，未根据本局落地姿态在线修正动作。log step 137 对本局返回的本��传感数据作诊断，在仿真 1.20 s 和仿真 1.21 s 打印骨盆 gy 分别为 -8.53 和 -12.15 rad/s，说明落地附近仍有较大角速度；结合视频 1.12–1.20 s 的展开接地及 log step 136 的 fell，支持所执行控制未能在落地时消除旋转并稳定承重。现有证据不足以单独确定某个增益、开腿时刻或控制子类。"
+   }
+  ],
+  "gap_zh": "没有状态日志，25 fps 视频末帧为视频 1.20 s，未覆盖服务报告摔倒的全部细节，无法辨认最终触发判负的具体身体接触部位或独立核验骨盆高度、速度等数值。log step 137 是 agent 对返回传感数据的诊断而非完整原始状态，不能精确重建落地接触与控制误差链；agent 在 log step 144 自述空中转角约 321°，不作为原生评分数值。run.json 的评分字段未报告 turn_deg、air_s，故两项 native_metrics 的 value 均为 null。"
+ },
+ "summary_zh": "最终评分为第 6 局：机器人确实起跳并完成空中后翻，服务裁判确认通过前三个原生阶段，但在视频约 1.20 s 落地处判定摔倒，后空翻并落地站稳节点始终未完成，最终为 0/1。日志中的开环动作提交和落地附近较大骨盆角速度支持落地控制未能稳定承重；没有状态日志且视频止于落地附近，无法确认最终非脚接触部位或精确原生转角、腾空时长。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与证据一致：视频 1.24 s 为第 6 局（control_steps=121），可见起跳、后翻、约 1.20 s 脚先落地且未站稳；服务裁判 fell、3/4 阶段，节点 false、断点落地站稳、平衡现象和停在“控制”的根因均有支撑，log_step 使用 steps[].id 且存在。仅有可补充但非错误之处：log step 137 的本体读数还显示骨盆累计后转 378°/384°（落地过转），可佐证根因。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:48:14Z",
+  "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/embodied-olympics-fun-backflip-i00-standard--pu6ny3q/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-backflip-i00-standard--pu6ny3q/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0b59c9e7257c7c9425e8dfad5ed4ad4801fb90b6ef0cbe462359e59b6723a176",
+   "video_sha256": "b4060780855a98c493676052a6f51b2d8caaf6b06ee813ac27f9f09468e61e82",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--acj2kf9.json

@@ -0,0 +1,197 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-privileged--acj2kf9",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.6667,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1291,
+   "sim_steps": 1291,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.6667,
+   "stages_reached": 2,
+   "stages_total": 3,
+   "fell": 0,
+   "replay_s": 19.04,
+   "bad_trajectory": 0,
+   "metric_entry_angle_deg": -0.25,
+   "metric_entry_speed_m_s": 2.33,
+   "metric_entry_y_m": -0.11,
+   "metric_pins_down": 8.0,
+   "metric_release_speed_m_s": 3.19,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 2097,
+   "agent_resets": 2274,
+   "agent_replays": 4,
+   "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": 15.193,
+   "peak_processes": 28,
+   "peak_disk_gb": 0.983,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-bowling-i00-privileged--acj2kf9/replay.mp4",
+  "duration_s": 12.92,
+  "fps": 25.0,
+  "frames": 323,
+  "sha256": "dd95a916a50083302822cb842d713c328f3ba714ae69c72f3dc7c91a25a51f5d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.48,
+      "log_step": 221,
+      "text_zh": "观察：视频 0.00–0.44 s 机器人双掌在髋前夹住球；视频 0.48 s 双臂前送，视频 0.52–0.68 s 紫色球与双掌分离并向球道前进，双脚始终支撑在助走区。log step 221 重新生成最终交付文件，报告 delivery、pins 两个阶段。"
+     },
+     {
+      "t": 7.9,
+      "log_step": null,
+      "text_zh": "观察：视频 7.4–7.8 s 球接近远端瓶阵，视频约 7.9–9.6 s 球进入瓶区后，多只白色球瓶相继倾倒；视频 9.6–12.88 s 瓶台仍可见两根分开的直立瓶身，未呈现全中。"
+     },
+     {
+      "t": 12.88,
+      "log_step": 245,
+      "text_zh": "观察：log step 245 对最终 trajectory.npz 重新生成并回放，输出 1291 行、not a strike、pins_down=8、stages=[delivery,pins]；inputs/run.json 的评分字段 metric_pins_down=8、deterministic=1 与之相符。视频 12.00–12.88 s 仍有残瓶，机器人保持站立，因此投球全中节点未达到。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–0.48 s 左右平掌垫从两侧夹压紫色球，并共同向前送球；视频 0.52–0.68 s 球与双掌分离。候选：夹持；单掌没有闭合手指或夹爪，不符合抓握，未见明确改为从下方承托的阶段。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–12.88 s 机器人双脚接触助走区承重，送球时身体轻微摆动，随后保持原地站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": 8.0,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 9.6,
+   "log_step": 245,
+   "where": "操作中",
+   "evidence_zh": "观察：机器人在视频 0.48–0.68 s 送球，视频约 7.9 s 球开始碰瓶，视频 9.6–12.88 s 碰撞链结束后仍有两只球瓶直立。log step 245 对交付文件的回放返回 not a strike、pins_down=8；本次仅尝试这一球，未补救投球全中步骤。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "其他",
+    "t": 9.6,
+    "log_step": 245,
+    "evidence_zh": "观察：视频约 7.9–9.6 s 球进入瓶阵并带倒多瓶，视频 10.0、11.2、12.0、12.88 s 的瓶区放大画面仍见两只分开的直立白瓶，其他瓶体倒在附近；log step 245 的交付回放点数为 8。",
+    "note_zh": "候选：击瓶后残瓶直立。现有条目没有准确覆盖已命中瓶阵却只造成部分倒瓶的结果；画面不能证明具体是位姿不准或接触力异常。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 215、221 的最终生成脚本以固定双臂关节斜坡和回摆送球；log step 181 的同类送球试验改变摆臂时长即改变出球速度、入瓶偏移及倒瓶��。log step 245 对最终文件测得离手速度 3.19 m/s、入瓶速度 2.33 m/s、横向偏移 -0.11 m、入射角 -0.25°，与视频 0.48–0.68 s 出球、视频 9.6–12.88 s 留有残瓶的过程相符。所执行的送球运动未形成全中的碰撞链，归为控制；证据不能进一步区分接触几何、轨迹或参数哪一项主导。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "未提供本次轨迹的逐帧状态日志；640×360 远景及放大图可辨末态残瓶，但不能可靠测量逐瓶倾角、掌球接触力或交付投球的自旋。8 瓶精确计数依评分器及交付文件回放，不能仅凭画面独立复算；也不能由单次入瓶速度断言全中所需的普适速度阈值。agent 在 log step 256 对平掌夹持无法产生所需旋转的解释只是自述，未据此认定任务无解或细分根因。",
+  "unmet": {
+   "strike_completed": false
+  }
+ },
+ "summary_zh": "机器人在犯规线后由双掌夹压送出一球，视频约 7.9 s 碰瓶，至视频 12.88 s 仍有两只球瓶直立；评分器及最终文件回放均记录击倒 8 瓶，因此唯一节点“投球全中”未完成。断点为投球全中操作中的残瓶结果，根因保守归为送球控制，不能确定速度、旋转或入射方向哪项主导。无逐帧状态日志，精确倒瓶计数依原生评分，远景仅支持残瓶与接触方式的视觉判断。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "视频显示 0.6 s 送出球、约 8 s 起球与球瓶碰撞、12.8 s 仍有残瓶直立，评分 pins_down=8，节点未达成、断点、接触与根因保守归控制均有证据支持。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:33:28Z",
+  "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/embodied-olympics-fun-bowling-i00-privileged--acj2kf9/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--acj2kf9/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7d09e3f6540f015566c94ec29b68cf9d84a6f935fb770b029b2b57933b7f155b",
+   "video_sha256": "dd95a916a50083302822cb842d713c328f3ba714ae69c72f3dc7c91a25a51f5d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--fbpjn5r.json

@@ -0,0 +1,178 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-privileged--fbpjn5r",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 971,
+   "sim_steps": 971,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 3,
+   "stages_total": 3,
+   "fell": 0,
+   "replay_s": 11.05,
+   "bad_trajectory": 0,
+   "metric_entry_angle_deg": -0.18,
+   "metric_entry_speed_m_s": 4.25,
+   "metric_entry_y_m": -0.059,
+   "metric_pins_down": 10.0,
+   "metric_release_speed_m_s": 5.7,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 1374,
+   "agent_resets": 1106,
+   "agent_replays": 2,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 1,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 24.243,
+   "peak_processes": 20,
+   "peak_disk_gb": 0.242,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "over_memory_gb": 1,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-bowling-i00-privileged--fbpjn5r/replay.mp4",
+  "duration_s": 9.72,
+  "fps": 25.0,
+  "frames": 243,
+  "sha256": "3f0fe4763279eea7b6a352a58e84e3dc6a2ca46927f8a86aa1a4c247afae2960"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [
+     [
+      9.68,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.72,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.64 s 机器人双脚留在犯规线后的助走区，双臂共同持紫色球；视频 0.72 s 双掌位于球的下后方并向前上方摆动，视频 0.80–0.96 s 球已与两掌分离并向球道飞出。"
+     },
+     {
+      "t": 4.72,
+      "log_step": null,
+      "text_zh": "观察：视频 1.00–4.64 s 紫球沿木质球道前进，视频 4.64–4.80 s 抵达白色红颈球瓶阵列；视频 4.96–6.72 s 球瓶逐渐倾斜、散开，原有直立阵列消失。远端像素不足以逐瓶判定倾角或清点遮挡球瓶。"
+     },
+     {
+      "t": 9.68,
+      "log_step": 83,
+      "text_zh": "观察：视频 8.76–9.68 s 机器人仍双脚站在助走区，瓶区仅见低矮散落的白色瓶体；inputs/run.json 的 reward 记录 success=1、replay_success=1、deterministic=1、metric_pins_down=10、stages_reached=3。log step 83 对已保存交付文件的官方回放返回 success=true、failed=null、stages=[delivery,pins,strike]、pins_down=10；log step 84 的第二次回放返回成功且终态哈希相同。"
+     },
+     {
+      "t": 9.68,
+      "log_step": 83,
+      "text_zh": "推断：结合视频约 4.7 s 首次碰瓶、DAG 的碰瓶后 5.0 秒点数条件和交付文件的官方成功回放，投球全中在视频末端约 9.7 s 完成；区间起点取已检查的末帧视频 9.68 s 为近似导航锚点，不在球刚离手或瓶刚倒下时提前记完成。该判断采用评分器结果补足远景逐瓶不可辨的部分，未将 agent 会话时间换算为视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "托/舀",
+      "note_zh": "观察：视频 0.40–0.72 s 机器人两臂共同送球，两块平掌垫在紫球下后方承托；视频 0.80 s 球已离手。推断：结合 inputs/run.json 的开局双掌夹压说明，持球从两掌侧向夹压转为下后方承托；两手无可闭合手指，不属于抓握。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–9.68 s 机器人双脚接触木质助走区承重，双臂送球时脚仍留在原处，出手后继续站立至结束。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": 10.0,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人双脚原地支撑，以双掌承托送球，紫球约在视频 0.8 s 离手、视频 4.7 s 碰瓶，按五秒点数延迟在视频末端约 9.7 s 完成投球全中，最终节点 1/1；评分器记录十瓶全倒，交付文件两次官方回放均成功且终态哈希相同。无状态日志，远景无法逐瓶量角或辨认细微接触，全中及合法性细节由评分器成功结果佐证，完成时刻使用视频 9.68 s 末帧作近似锚点。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "单节点投球全中：视频 0.72 s 双掌承托送球、4.7 s 碰瓶、末帧约 9.68 s 记完成（碰瓶后 5 s 点数），与评分器成功结果一致；接触标签（双臂协作·托/舀 + 腿/躯干·支撑）与画面相符，无失败层合理。",
+    "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": "2026-10-31T00:32:41Z",
+  "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/embodied-olympics-fun-bowling-i00-privileged--fbpjn5r/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--fbpjn5r/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "878703737678b7c2e69f03fd75c39745ba991403fcaea8361abe3f028fa7cb16",
+   "video_sha256": "3f0fe4763279eea7b6a352a58e84e3dc6a2ca46927f8a86aa1a4c247afae2960",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--flmmlvx.json

@@ -0,0 +1,197 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-privileged--flmmlvx",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.6667,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1869,
+   "sim_steps": 1869,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.6667,
+   "stages_reached": 2,
+   "stages_total": 3,
+   "fell": 0,
+   "replay_s": 27.09,
+   "bad_trajectory": 0,
+   "metric_entry_angle_deg": 0.43,
+   "metric_entry_speed_m_s": 1.35,
+   "metric_entry_y_m": 0.088,
+   "metric_pins_down": 7.0,
+   "metric_release_speed_m_s": 2.13,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 110,
+   "agent_resets": 324,
+   "agent_replays": 3,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.099,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.84,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-bowling-i00-privileged--flmmlvx/replay.mp4",
+  "duration_s": 18.72,
+  "fps": 25.0,
+  "frames": 468,
+  "sha256": "ff20a458f7989823b85bd277440c9768382e03e7aad95eba4ea6987184bfc64c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.6,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–0.68 s，机器人双脚站在助走区，紫色球位于两掌之间；视频 0.72–0.80 s 双臂向前送球，视频 0.84–1.08 s 球与双掌分离并向球道落下。"
+     },
+     {
+      "t": 13.8,
+      "log_step": null,
+      "text_zh": "观察：视频 13.2–13.6 s 紫色球沿木质球道接近瓶阵，视频 13.8–15.4 s 球进入瓶阵并引起部分球瓶倾倒，仍有白色球瓶直立；视频 17.6–18.68 s 这些直立球瓶仍在，机器人始终双脚站立。"
+     },
+     {
+      "t": 18.68,
+      "log_step": 329,
+      "text_zh": "观察：log step 327 对最终生成的交付文件进行两次独立回放；log step 329 两次均报告 delivery、pins 两阶段已到达，pins_down=7、failed=\"not a strike\"，且终态哈希相同；run.json 的评分字段 metric_pins_down=7、deterministic=1、fell=0 与之相符。推断：结合视频 18.68 s 仍有球瓶直立，满五秒点数未全中，整个视频内本节点均未成立。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.0–0.76 s 两掌从球两侧夹压持球，并共同前送，随后球离掌；log step 195 的交付生成器同时控制左右掌前推，与画面一致。候选：夹持；平掌之间的夹压没有任何一只手闭合包围球，不能归为抓握。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–18.68 s，机器人双脚接触助走区木质地面承重，在送球和等待倒瓶期间保持站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": 7.0,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 13.8,
+   "log_step": 329,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 13.8–15.4 s，机器人送出的球进入瓶阵后只引起局部倾倒，部分球瓶一直留立至视频 18.68 s；log step 329 的最终交付回放以 not a strike 结束并点数 7 瓶。推断：失败发生于投球全中这一步，视频 13.8 s 以碰瓶阶段作粗略定位；球已送出并碰瓶，但未形成全中。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "其他",
+    "t": 14.8,
+    "log_step": 329,
+    "evidence_zh": "观察：视频 14.4–15.4 s，紫色球进入瓶阵前部后向画面左侧偏转，周围球瓶倒下，但另一些球瓶仍竖立；视频 17.6–18.68 s 球留在瓶台附近，直立球瓶未被后续碰撞带倒。log step 329 的点数为 7 瓶。",
+    "note_zh": "候选：倒瓶连锁不足。可见问题是球与球瓶碰撞后未能使剩余球瓶连锁倾倒；球确已碰瓶，且没有足够证据把该结果具体标成瞄准位姿偏差、动态时机或失去持球，因此使用其他。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 195 的交付生成器采用双掌保持朝向的短程前推，以固定方向和微小左右差速送球；log step 327 重新运行该生成器，log step 329 两次最终回放均得到离手速度 2.13 m/s、入瓶速度 1.35 m/s、入瓶横向偏移 0.088 m 和 7 瓶。结合视频 13.8–18.68 s 的局部倒瓶，交付的投球控制未产生能清倒整个瓶阵的出球运动；只能归到控制，不能确定接触设计、轨迹跟踪或参数中的具体原因。"
+   }
+  ],
+  "gap_zh": "没有逐帧状态或接触日志；远景球瓶只有少量像素，无法逐瓶核对倾角、身份和碰撞传递顺序，最终数量取评分器及交付回放的 7 瓶。视频与日志能确认未全中，但无法分离球速、旋转、入射路线和掌球传力各自的因果贡献，故现象使用候选“倒瓶连锁不足”，根因仅标控制，不指定子类。"
+ },
+ "summary_zh": "机器人从犯规线后双掌送出紫色球并有效碰瓶，但点数仅倒 7 瓶，唯一的“投球全中”节点始终未完成，最终为 0/1。断点定位于视频约 13.8 s 的碰瓶阶段：球进入瓶阵后只造成局部连锁倾倒，余瓶直立至视频 18.68 s，机器人未跌倒；最终文件的两次回放结果一致。投球控制未达到全中效果，但缺少逐帧接触和逐瓶状态，不能进一步确定球速、入射路线或传力中的具体根因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与视频、评分字段一致：球约 13.8 s 入瓶阵，仅局部倒瓶（7 瓶），节点未达成，断点、接触与根因（控制，不分子类）证据充分。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:35:30Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--flmmlvx/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--flmmlvx/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "7ed91b77ad9b11380aa5f3dbef781033bb6456f74f1d4a797dfebcfef82da3f1",
+   "video_sha256": "ff20a458f7989823b85bd277440c9768382e03e7aad95eba4ea6987184bfc64c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--pdhbyit.json

@@ -0,0 +1,206 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-privileged--pdhbyit",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.6667,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1185,
+   "sim_steps": 1185,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.6667,
+   "stages_reached": 2,
+   "stages_total": 3,
+   "fell": 0,
+   "replay_s": 21.25,
+   "bad_trajectory": 0,
+   "metric_entry_angle_deg": -0.51,
+   "metric_entry_speed_m_s": 2.5,
+   "metric_entry_y_m": -0.126,
+   "metric_pins_down": 9.0,
+   "metric_release_speed_m_s": 3.84,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 470,
+   "agent_resets": 228,
+   "agent_replays": 7,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 4.307,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.28,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 1,
+   "session_log_from_monitor": 1
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-bowling-i00-privileged--pdhbyit/replay.mp4",
+  "duration_s": 11.88,
+  "fps": 25.0,
+  "frames": 297,
+  "sha256": "e3353649328e85999efce7a29d508fb4822b3e4d412c2eff28da1271e805340a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.2,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.20 s，机器人两掌夹托紫球并向前上方送出；视频 0.28 s 球已与双掌明显分离，随后沿木质球道前进，机器人双脚留在助走区。"
+     },
+     {
+      "t": 7.0,
+      "log_step": null,
+      "text_zh": "观察：视频约 6.8–7.4 s，球到达瓶区并引起球瓶倾斜、散开；视频 8.0–11.84 s 的瓶台放大画面仍见画面右侧一只球瓶直立，机器人全程站立。"
+     },
+     {
+      "t": 11.84,
+      "log_step": null,
+      "text_zh": "观察：inputs/run.json 的评分器字段 metric_pins_down=9、stages_reached=2/3、success=0、deterministic=1、fell=0；与视频末尾仍有球瓶直立相符，因此投球全中判据从未成立。"
+     },
+     {
+      "t": null,
+      "log_step": 563,
+      "text_zh": "观察：log step 563 的调试会话结果为 8 瓶、1201 控制步，不能当作本次评分结果；评分器实际回放为 9 瓶、1185 控制步。log step 592–593 又创建并运行允许大于 8 瓶时保存轨迹的脚本，但没有返回结果，无法从这段日志唯一确认最终保存的参数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–0.16 s 左右平掌从两侧共同夹压紫球；候选：夹持。单掌没有闭合包围球体，现有抓握标签不适用。"
+     },
+     {
+      "who": "双臂协作",
+      "how": "托/舀",
+      "note_zh": "观察：视频约 0.16–0.24 s，双掌由夹球转为从球的下后方共同托送，之后球离开双掌；这是同一次出球中依次发生的接触。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–11.84 s，机器人双脚接触木质助走区承重，出球及等待球瓶运动期间保持站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": 9.0,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 7.0,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 7.0 s 紫球抵达瓶区并开始引起倒瓶，但视频 8.0–11.84 s 仍有一只球瓶直立，评分器 metric_pins_down=9，success=0。推断：这次投球未形成覆盖全部球瓶的碰撞连锁，投球全中这一步未完成；视频时间为碰瓶附近的粗略导航锚点。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 7.0,
+    "log_step": null,
+    "evidence_zh": "观察：视频约 6.8–8.0 s，紫球进入瓶阵并撞散多数球瓶，画面右侧的一只球瓶未随其余球瓶倒下，至视频 11.84 s 仍直立；评分器记录 9 瓶倒下。推断：球与倒瓶的运动路径没有击倒剩余静止目标，归为投球位姿不准。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：视频 0.16–0.28 s 显示机器人完成双掌前送与释放，视频约 7.0–11.84 s 显示球实际撞瓶后留下一只直立球瓶；评分器记录离手速度 3.84 m/s、入瓶速度 2.50 m/s、入瓶横向位置 -0.126 m、入射角 -0.51°。结合 log step 592 中通过肩、肘、腕与躯干角度设定投球的脚本，问题可归于出球运动控制尚未形成全中所需的球路；没有证据确定某一参数错误，因此不细分控制子类。",
+    "source_zh": "视频 / 评分器字段 / 日志"
+   }
+  ],
+  "gap_zh": "无状态日志或逐瓶接触记录；远景虽可辨一只球瓶始终直立，不能可靠确定其编号、精确倾角及每次碰撞。日志中 log step 563 的 8 瓶会话与评分回放的 9 瓶不同，log step 592–593 的末次脚本运行又未返回保存结果，故不把调试结果或其控制步时间映射为本视频事实。无法进一步区分出球速度、横向落点、旋转和入射角对漏瓶的贡献。",
+  "unmet": {
+   "strike_completed": false
+  }
+ },
+ "summary_zh": "机器人完成双掌送球、离手和有效碰瓶，保持站立，但点数只击倒 9/10 瓶；唯一节点“投球全中”始终未达到，最终为 0/1。视频约 7.0 s 的碰瓶是断点锚点，之后仍有一瓶直立，现象为位姿不准，根因保守归于投球控制。缺少状态及逐瓶接触记录，且会话调试结果与最终回放不同，无法确定漏瓶的具体控制参数原因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "唯一节点未达成（9/10 瓶，评分器 metric_pins_down=9）、断点、接触、现象与保守根因均有视频与评分器字段支持，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": "2026-10-31T00:33:13Z",
+  "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/embodied-olympics-fun-bowling-i00-privileged--pdhbyit/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--pdhbyit/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "5d4ef448d9adee89f0c0f32636afa2041ea30c44fdaee9ecf2b0d209ea58ef2a",
+   "video_sha256": "e3353649328e85999efce7a29d508fb4822b3e4d412c2eff28da1271e805340a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--svlk5hb.json

@@ -0,0 +1,213 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-privileged--svlk5hb",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.6667,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1179,
+   "sim_steps": 1179,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.6667,
+   "stages_reached": 2,
+   "stages_total": 3,
+   "fell": 0,
+   "replay_s": 12.33,
+   "bad_trajectory": 0,
+   "metric_entry_angle_deg": -0.24,
+   "metric_entry_speed_m_s": 2.81,
+   "metric_entry_y_m": -0.058,
+   "metric_pins_down": 8.0,
+   "metric_release_speed_m_s": 3.26,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 68,
+   "agent_resets": 1464,
+   "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/embodied-olympics-fun-bowling-i00-privileged--svlk5hb/replay.mp4",
+  "duration_s": 11.8,
+  "fps": 25.0,
+  "frames": 295,
+  "sha256": "90b39193cffa7b0ede05ec89ab379b9492c4295fa1065336608266d77a3b1e2f"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.48 s，机器人在助走区站立，紫球起初位于两掌之间，双臂随后向前摆送；视频 0.52–0.80 s 可见球与双掌分离并向球道前进。"
+     },
+     {
+      "t": 6.8,
+      "log_step": 94,
+      "text_zh": "观察：视频 6.20–7.70 s 的瓶区放大帧显示紫球到达球瓶阵列，随后部分球瓶倾倒，仍有球瓶直立；视频 10.80–11.76 s 仍可见直立瓶。log step 94 对已交付文件的独立回放报告 stages=[delivery,pins]、failed=not a strike、pins_down=8、standing=[7,8]；评分器 reward.metric_pins_down=8、stages_reached=2/3、fell=0、deterministic=1，与视频一致。"
+     },
+     {
+      "t": 11.76,
+      "log_step": 95,
+      "text_zh": "推断：本节点要求首次有效碰瓶后点数时十瓶全倒；视频终态仍有直立瓶，且 log step 94–95 对交付文件的回放均只判八瓶，故节点从未成立，final=false。原生 delivery、pins 两个阶段不是本 DAG 的已完成语义节点，不能折算为本节点达到。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00 s 紫球位于机器人两只平掌垫之间，随后两臂共同送球。候选：夹持；初始双掌从两侧夹压球，各掌没有能闭合包围球的手指，不符合抓握。"
+     },
+     {
+      "who": "双臂协作",
+      "how": "托/舀",
+      "note_zh": "推断：视频 0.12–0.48 s 双臂保持紫球并向前送出，随后双掌抬起、球离手；结合 log step 91 生成脚本中的双臂 scoop_pose 与摆送段，以及 log step 93 生成动作与交付文件完全相同的检查，初始夹压之后转为两只平掌垫从下后方共同承托送球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.80 s 双脚接触助走区支撑送球，视频 11.76 s 机器人仍双脚站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": 8.0,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 7.6,
+   "log_step": 94,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 6.8 s 球到达瓶区，视频 7.4–7.7 s 已有多瓶倒下但仍有直立瓶，直到视频 11.76 s 也未全部倾倒。log step 90 保存的 1179 行投球文件在 log step 94–95 独立回放均达到 delivery、pins，最终以 not a strike 结束，裁判记录仅八瓶倒下；本次投球的连锁击倒没有完成投球全中节点。"
+  },
+  "phenomena": [],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 90、94–95 对实际交付动作记录离手速度 3.26 m/s、入瓶速度 2.81 m/s、入瓶横向位置 -0.058 m、入瓶角度 -0.24°，多次确定性回放均留下两瓶。会话调试 log step 80 改变夹压与摆臂加速度后出球速度、方向发生变化，log step 89 改变躯干朝向产生八瓶或入沟的不同结果，说明待解决的是送球动作对球运动状态的控制；结合视频 6.8–11.76 s 的局部倒瓶，归为控制层面。现有证据不足以进一步区分速度、旋转、方向或具体接触设计的贡献。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "视频 6.8–7.6 s 可见球进入瓶阵后部分球瓶倾倒，视频 10.8–11.76 s 仍有直立瓶；log step 94–95 的交付文件回放记录 pins_down=8、failed=not a strike，其中 log step 94 记录 standing=[7,8]。这些证据确认结果未全中，但不能单独指认运动或接触缺陷；“连锁击倒不足”只是复述未达标指标，故按 docs/03_trajectory.md Step 3.1、DECISIONS.md N10 与 CASES.md FAIL-15 将现象层留空。没有逐帧状态日志，远景视频中的掌垫接触面和球瓶倾角分辨率有限；精确倒瓶数及剩余瓶编号采用实际交付文件的回放裁判记录，并由末段仍有直立瓶的画面支持。缺少本次动作的逐时掌球接触力、跟踪误差及足以区分速度、旋转和瞄准影响的对照，因此控制根因不细分；入瓶速度 2.81 m/s 本身没有已验证的成功阈值，不能单凭该值标注出球偏慢。agent 在 log step 106 自述的“速度上限使全中不可达”未被当作任务或环境问题，不能据此排除所有合法方案。",
+  "unmet": {
+   "strike_completed": false
+  }
+ },
+ "summary_zh": "机器人从助走区双掌送出一球并保持站立，球抵达瓶阵，但仅八瓶倒下；投球全中节点最终为 0/1。断点为投球全中，交付文件的多次确定性回放均判 not a strike，视频末段仍有直立瓶，精确倒瓶数由裁判记录补足。现有证据不能单独指认具体运动或接触缺陷，现象层留空；保留控制根因，但具体速度、旋转或方向因素尚不能分清。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点值、时间、断点、contact_actual 与视频和评分器数字一致（球约 6.8–7.2 s 到瓶区，终帧仍有直立瓶，pins_down=8）。唯一问题：失败现象“运动·其他：连锁击倒不足”只是复述 pins_down<10 这一指标（N10），不是可见的运动缺陷。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "节点 final=false、断点、contact_actual、控制根因与空现象层（含 gap_zh）均与视频、评分器数字（pins_down=8、stages 2/3、fell=0）一致；第 1 轮意见已正确落实，未发现新问题。",
+    "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": "2026-10-31T00:47:45Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--svlk5hb/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--svlk5hb/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--svlk5hb/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--svlk5hb/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "d97f869a42b0359231ea96510c947dd5432f8c15a29972ec668918523963d730",
+   "video_sha256": "90b39193cffa7b0ede05ec89ab379b9492c4295fa1065336608266d77a3b1e2f",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--xnyayjs.json

@@ -0,0 +1,173 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-privileged--xnyayjs",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 10,
+   "sim_steps": 10,
+   "horizon": 4000,
+   "episode_over": 0,
+   "progress": 0.0,
+   "stages_reached": 0,
+   "stages_total": 3,
+   "fell": 0,
+   "replay_s": 3.49,
+   "bad_trajectory": 0,
+   "metric_pins_down": 0.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 4,
+   "agent_resets": 1,
+   "agent_replays": 0,
+   "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": 0.824,
+   "peak_processes": 17,
+   "peak_disk_gb": 0.261,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-bowling-i00-privileged--xnyayjs/replay.mp4",
+  "duration_s": 0.12,
+  "fps": 25.0,
+  "frames": 3,
+  "sha256": "da2555daaa17e3fd54db7fe7cd5dd7e2014b06fe1153a860dc1405336e79ebea"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": 6,
+      "text_zh": "观察：视频 0.00、0.04、0.08 s（全部三帧，含末帧）中，机器人均站在犯规线后的初始位置，双臂停在腰前持球姿态，未见向前送球或球进入球道，远端球瓶仍在瓶台上；投球全中条件始终未成立。仅维持初始姿态，未开始本节点的投球动作，故不列动作接触。"
+     },
+     {
+      "t": null,
+      "log_step": 6,
+      "text_zh": "观察：log step 6 中 agent 的交付脚本调用 Session.hold()，仅执行十次相同保持命令后调用 save_trajectory()；附带说明将其称为初始占位轨迹。评分器字段 n_actions=10、sim_steps=10、stages_reached=0、metric_pins_down=0，与短视频一致。"
+     },
+     {
+      "t": null,
+      "log_step": 190,
+      "text_zh": "观察：log step 129 的会话测试曾报告 delivery 阶段，log step 188—190 仍在启动投球参数扫描并等待；这些测试之后未见再次保存交付轨迹的调用，且测试动作不出现在本次视频中，不能计入节点进度。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": 0.0,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 0.0,
+   "log_step": 6,
+   "where": "首步前",
+   "evidence_zh": "观察：视频 0.00—0.08 s 的全部三帧中，机器人维持初始站立持球姿态，未开始送球；log step 6 保存的脚本只运行十步 hold 命令。推断：评分回放停留在初始保持阶段，失败步骤为投球全中，锚点取视频 0.00 s。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.0,
+    "log_step": 6,
+    "evidence_zh": "观察：视频 0.00、0.04、0.08 s 未见机器人启动摆臂送球，球未进入球道；log step 6 的已保存轨迹只含十步保持命令，评分器 n_actions=10、episode_over=0。推断：本次交付为未启动投球的占位轨迹，而非一次投球失误。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "已检查 inputs/，没有状态日志；评分视频完整但只有三帧。log step 6 明确保存占位轨迹，log step 188—190 仍在会话测试与等待，但现有记录不足以确定未形成正式投球交付的具体能力根因；会话测试中的控制偏差不能直接作为本次静止回放的失败原因，根因层留空。"
+ },
+ "summary_zh": "评分回放仅含十步初始保持命令，视频 0.00—0.08 s 的全部三帧均未见机器人出球，投球全中未达成，最终为 0/1 节点，倒瓶指标为 0/10。断点为投球全中的首步前，现象为没尝试；log step 129 等会话投球测试未进入交付，不能计作进度。没有状态日志，且记录不足以可靠确定占位轨迹未被替换的具体能力根因，根因层留空。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "标签与证据一致：评分回放只有十步 hold（n_actions=10，pins_down=0），视频三帧均无送球；节点未达成，断点首步前/没尝试，根因层留空并有 gap_zh；会话测试未计入进度。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:32:38Z",
+  "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/embodied-olympics-fun-bowling-i00-privileged--xnyayjs/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--xnyayjs/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "1befa2c1aaa5b5e533a4a0fdd77d1d9c2dd7c8e53fe37bf8403bbc0425d083ae",
+   "video_sha256": "da2555daaa17e3fd54db7fe7cd5dd7e2014b06fe1153a860dc1405336e79ebea",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-privileged--zeqwg9e.json

@@ -0,0 +1,182 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-privileged--zeqwg9e",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 881,
+   "sim_steps": 881,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 3,
+   "stages_total": 3,
+   "fell": 0,
+   "replay_s": 11.12,
+   "bad_trajectory": 0,
+   "metric_entry_angle_deg": -0.41,
+   "metric_entry_speed_m_s": 4.66,
+   "metric_entry_y_m": -0.065,
+   "metric_pins_down": 10.0,
+   "metric_release_speed_m_s": 6.28,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 186,
+   "agent_resets": 1481,
+   "agent_replays": 0,
+   "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": 1.603,
+   "peak_processes": 15,
+   "peak_disk_gb": 0.241,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-bowling-i00-privileged--zeqwg9e/replay.mp4",
+  "duration_s": 8.84,
+  "fps": 25.0,
+  "frames": 221,
+  "sha256": "c78c9ac32dbd862c00cd7d4252c9cc3e5af315b061a43f2165338ad97c867682"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 1,
+  "final": 1,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [
+     [
+      8.8,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.24,
+      "log_step": null,
+      "text_zh": "观察：视频 0.24–0.28 s，紫球随机器人前摆的双臂向前上方移动，视频 0.32 s 球与掌垫已明显分离，随后双臂上抬；机器人双脚仍在犯规线后的助走区。"
+     },
+     {
+      "t": 3.8,
+      "log_step": null,
+      "text_zh": "观察：视频 0.36–3.76 s，已离手的紫球沿木质球道向远端白色球瓶运动；视频约 3.8–4.4 s 球进入瓶阵，瓶阵随后散开、降低。远景分辨率不足以逐瓶量角或精确区分全部十瓶。"
+     },
+     {
+      "t": 8.8,
+      "log_step": 131,
+      "text_zh": "观察：视频 7.8–8.8 s 逐帧检查中，瓶台仅见低伏的白色瓶体，机器人仍以双脚站立；inputs/run.json 的评分器 reward 记录 success=1、replay_success=1、deterministic=1、stages_reached=3/3、metric_pins_down=10、fell=0。log step 130 重新生成最终 881 行轨迹，log step 131 对该文件独立回放返回 success=True、executed_rows=881，与评分器动作数及指标相符。"
+     },
+     {
+      "t": 8.8,
+      "log_step": null,
+      "text_zh": "推断：由视频约 3.8 s 的入瓶事件、DAG 所述首次有效碰瓶后延迟五秒点数，以及评分器的全中成功结果，将投球全中首次成立标在视频约 8.8 s，时间为近似值；此前即使已倒瓶也未完成点数验收。该时刻并非由 agent 工作时间换算。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "推断：候选“夹持”。结合开局双掌夹球的已知初态、视频 0.00–0.24 s 的双臂送球动作，机器人左右平掌垫先从两侧夹压紫球；掌垫无闭合手指，故不标抓握。球和内侧掌垫在开局部分被躯干遮挡。"
+     },
+     {
+      "who": "双臂协作",
+      "how": "托/舀",
+      "note_zh": "推断：视频 0.24–0.28 s 紫球随掌垫向前上方运动，近侧掌垫位于球的下后方；结合 log step 100 的双臂镜像关节控制代码及 log step 109、130 对同一生成脚本的更新，机器人在开局夹压后转为双掌共同承托并送出紫球，视频 0.32 s 已离手。远侧掌垫受遮挡，无法逐帧确认两侧各自最后接触的时刻。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–8.80 s，机器人双脚接触犯规线后的木质助走区，承重并支撑双臂送球及之后的等待；末尾逐帧检查未见跌倒。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": 10.0,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人在犯规线后以双掌夹压、承托并前摆送出紫球，球于视频约 0.32 s 离手、约 3.8 s 到达瓶阵，机器人保持站立至结束。评分器记录十瓶全倒且三个原生阶段全部完成，投球全中节点在视频约 8.8 s 的���迟点数时刻成立，最终完成 1/1 个节点。证据缺口：没有状态日志，远端球瓶和远侧掌垫细节受分辨率及遮挡限制，逐瓶判定与合法性以评分器成功结果佐证，完成时刻为近似值。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "strike_completed 在视频约 8.8 s 成立与评分器成功（十瓶全倒、3/3 阶段、fell=0）及视频相符；球约 3.8–4.0 s 入瓶阵，+5 s 约在视频末尾，近似时间已标注。contact_actual 与画面一致，无失败层，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": "2026-10-31T00:33:00Z",
+  "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/embodied-olympics-fun-bowling-i00-privileged--zeqwg9e/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-privileged--zeqwg9e/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "001ecefd6619f56c2c1d2c69a891afa15773aa8f44f291eca98dd5132893f7ac",
+   "video_sha256": "c78c9ac32dbd862c00cd7d4252c9cc3e5af315b061a43f2165338ad97c867682",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--6x5p58r.json

@@ -0,0 +1,197 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-standard--6x5p58r",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 2282,
+   "sim_steps": 74140,
+   "calls": 200,
+   "rows": 5146,
+   "resets": 50,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 5971.9,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.739,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.348,
+   "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/embodied-olympics-fun-bowling-i00-standard--6x5p58r/replay.mp4",
+  "duration_s": 22.84,
+  "fps": 25.0,
+  "frames": 571,
+  "sha256": "b23ba32129e58ec26bdab017a0ccec2dadca16968506e0a5bb6875a3e7ac72a4"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.64,
+      "log_step": 742,
+      "text_zh": "观察：视频 0.00–0.60 s 机器人两掌夹住紫色球，约视频 0.64–0.72 s 双臂张开、球与手分离；视频 0.96–1.20 s 球在木质球道上向远端运动，机器人双脚仍在助走区。log step 742 启动最后一次在线投球，区别于此前重置后的调试局。"
+     },
+     {
+      "t": 17.8,
+      "log_step": null,
+      "text_zh": "观察：视频 16.5–17.7 s 紫色球沿画面左侧接近瓶阵，视频 17.8–18.4 s 球进入瓶区、部分白色瓶体发生倾倒和位移，但瓶阵未全部倒下。"
+     },
+     {
+      "t": 22.8,
+      "log_step": 750,
+      "text_zh": "观察：逐帧检查视频 21.80–22.80 s，瓶台仍有多只明显直立的白色球瓶；同相机的成功参考在其视频 9.20–9.68 s 已无此直立瓶阵。log step 750 的在线服务状态为 episode=51、done=True、reason=failed，run.json 的评分字段 live_success=0，与未全中画面一致。"
+     },
+     {
+      "t": 22.8,
+      "log_step": 751,
+      "text_zh": "观察：agent 在 log step 751 自述最终击倒八瓶、留下三号和十号瓶；提供的评分字段没有 pins_down 数值，远景也不足以可靠逐瓶计数，故不把该自述作为原生计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–0.60 s 左右平掌垫从球的两侧共同夹压并随双臂摆动，随后在视频约 0.64–0.72 s 张开释放。候选：夹持；两掌没有可闭合手指，且无单手包围球体，不适用抓握；远景不能进一步分辨瞬时单掌接触切换。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–1.20 s 投球时机器人双脚在犯规线后的木质助走区承重；后续保持站立至视频 22.80 s。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": null,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 17.8,
+   "log_step": 742,
+   "where": "操作中",
+   "evidence_zh": "观察：最后一次在线投球对应 log step 742；视频约 17.8 s 球从画面左侧进入瓶阵并造成部分瓶体运动，视频 21.80–22.80 s 仍有多只球瓶直立，投球全中未达成。log step 750 的服务状态确认此局已以 failed 结束。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 17.8,
+    "log_step": null,
+    "evidence_zh": "观察：视频 16.5–18.4 s 紫色球偏向画面左侧进入瓶阵，冲击后只改变部分瓶体的位置和姿态；视频 21.80–22.80 s 瓶台仍保留多只直立瓶体，未形成成功参考中球瓶全部倒下的结果。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 616 的控制脚本以预设肩、肘和躯干目标序列夹球、摆臂和开掌，log step 742 以 torso_aim=-0.01578880 启动最后一局；视频 0.64–1.20 s 球已离手沿球道前进，但视频约 17.8 s 入瓶位置偏向瓶阵一侧，之后仍留下直立球瓶。目标明确为全中，失败落在送球运动的实际效果，支持归于控制；缺乏实际关节跟踪、接触力和离手速度记录，不能确定更细的控制子类或具体错误参数。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "无状态日志，所提供视频只有场地远景；可确认仍有直立球瓶，但不能可靠识别各瓶编号、精确倾角和倒瓶总数。run.json 的评分字段未报告 pins_down，故该原生指标为 null；agent 在 log step 751 的“八瓶”说法未获逐瓶证据支持。缺少最后一局完整裁判元数据、关节跟踪和接触力记录，无法进一步区分入瓶偏差的具体控制机制。"
+ },
+ "summary_zh": "本次评分为最后一次在线投球：机器人双掌夹压后开掌送球，双脚持续支撑，球抵达瓶阵并撞动部分球瓶，但“投球全中”始终未达成，最终为 0/1 节点。断点在视频约 17.8 s 的入瓶冲击，球偏向瓶阵一侧，末秒仍有多只瓶直立，归为位姿不准及控制问题。远景和评分字段不足以核实精确倒瓶数及更细控制根因，agent 自述的八瓶未作为原生指标。",
+ "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、断点在strike_completed、双臂协作/其他与腿支撑的接触、位姿不准与控制根因均与视频一致；球在约15–18 s 缓慢接近并在约18 s 后撞瓶，末帧仍有直立瓶，未发现需改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:35:42Z",
+  "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/embodied-olympics-fun-bowling-i00-standard--6x5p58r/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--6x5p58r/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f565b0e9464ceb932b25b64e3a2e5fd6c112d614a9e8b918dfed653b532feaff",
+   "video_sha256": "b23ba32129e58ec26bdab017a0ccec2dadca16968506e0a5bb6875a3e7ac72a4",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--aydgyd7.json

@@ -0,0 +1,217 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-standard--aydgyd7",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 475,
+   "sim_steps": 20922,
+   "calls": 61,
+   "rows": 3600,
+   "resets": 17,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 7154.9,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.502,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.442,
+   "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/embodied-olympics-fun-bowling-i00-standard--aydgyd7/replay.mp4",
+  "duration_s": 4.76,
+  "fps": 25.0,
+  "frames": 119,
+  "sha256": "7d9be4ee52eece646638e9f54533e0fc098dfc5aa9f2d4a19274538252a69f51"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.32 s，机器人在犯规线后的助走区持紫色球并向前上方摆臂；视频 0.36–0.48 s 球与可见掌垫分开，随后沿球道前进。机器人确实开始了投球。"
+     },
+     {
+      "t": 3.6,
+      "log_step": null,
+      "text_zh": "观察：视频 3.60–4.72 s 逐帧检查中，紫色球沿画面左侧木质球道与灰色沟槽的交界前进，末段越来越贴近沟槽；视频 4.72 s 球仍在瓶阵前，远端白色球瓶仍成组直立，机器人保持站立。未见碰瓶及其后的点数阶段。"
+     },
+     {
+      "t": 4.72,
+      "log_step": null,
+      "text_zh": "推断：本次视频结束前没有形成十瓶全倒且完成碰瓶后点数的结果；结合评分器字段 live_success=0、graded_by_live=1，本步未完成。成功参考在参考视频约 4.5–5.0 s 球到瓶区后仍继续至约 9.7 s，本次不能把单纯送球记为全中。"
+     },
+     {
+      "t": null,
+      "log_step": 243,
+      "text_zh": "观察：log step 241–242 中 agent 将送球脚本改成多次重置的瞄准参数扫描并启动；log step 243 的最后发言只是声称已启动扫描、等待结果，没有给出末次尝试的裁判返回。日志此前各次试验的倒瓶数不能作为本次视频的节点进度。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "候选：夹持。视频 0.00–0.12 s 的初始持球阶段，结合任务给定双掌压球初态，左右平掌垫夹压紫色球；掌垫无可闭合手指，不构成抓握。部分内侧接触被躯干遮挡。"
+     },
+     {
+      "who": "手指/夹爪",
+      "how": "托/舀",
+      "note_zh": "视频 0.16–0.32 s，画面右侧的平掌垫可见位于紫球下后方，随摆臂承托并送球；视频 0.36 s 可见球与该掌分离。另一掌的脱离先后被遮挡，未据此认定全程为单掌出球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "视频 0.00–4.72 s，机器人双脚在犯规线后的助走区承重，摆臂送球后仍站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": null,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 3.6,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 3.60–4.72 s，机器人送出的球沿木质球道左缘前进；视频 4.72 s 球仍在瓶阵前，瓶阵仍直立，机器人保持站立，录像随后结束，未显示入沟、碰瓶或点数结局。inputs/run.json 的 reward.live_success=0、graded_by_live=1。推断：投球全中未完成，断点仍为该步／操作中；视频 3.60 s 仅作偏左球路的导航锚点，不能确定具体判负事件及其时刻。日志未给出末次参数扫描的逐次裁判返回，无法核实该录像对应扫描中的哪次投球；“偏向沟槽后没有修复”不能作为已观察到的失败经过。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 3.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频 1.00–3.60 s，球向远端运动并逐渐靠近画面左侧边缘；视频 3.60–4.72 s 球路偏左，视频 4.72 s 球仍贴近木质球道左缘、尚未到瓶阵，结局未见。推断：送球方向可能偏离有效击瓶球路，暂归位姿不准；这一解释不确定，画面没有显示入沟或错过瓶阵，也不能排除随后击中左侧球瓶但倒瓶不足，评分器的失败字段本身不能确认该现象。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "没有状态日志；仅有场地远景，视频在 4.72 s 的末帧仍显示球位于瓶阵前的球道左缘，缺少后续碰瓶、点数或入沟画面及对应裁判详情。inputs/run.json 的 reward.live_success=0 支持投球全中未完成，但不说明具体判负原因；log step 241–243 记录启动多次重置的参数扫描，未给出逐次裁判返回，无法把录像可靠对应到具体扫描参数或移用早期试验的入沟判词、倒瓶数。位姿不准仅为偏左球路的暂定解释，缺少其导致失败的直接证据；没有本次手球接触、离手速度或关节跟踪记录，原先送球控制失败的因果判断不能确立，根因留空。pins_down 未由本次评分器字段报告，保留 null。"
+ },
+ "summary_zh": "机器人完成持球摆臂和送球，视频 4.72 s 球仍在木质球道左缘、尚未到瓶阵，录像未显示后续结局；结合评分器 live_success=0，投球全中最终为 0/1，断点保留投球全中／操作中。偏左球路暂归位姿不准，但不能据此确认入沟、未击瓶或控制失败，根因留空。因缺少状态日志、投球结局及末次扫描的对应裁判详情，现象和失误时刻存在不确定性，置信度降为 4；先前试验的倒瓶数不计入本次进度。",
+ "conf": 4,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点 strike_completed=false、断点 strike_completed/操作中 与评分器字段（live_success=0）一致，未见 WHERE 错误。但视频（4.76 s）在球到达瓶阵之前就结束，球在 4.72 s 仍在木质球道内的左缘，并未入沟；\"球路偏向沟槽后没有修复\"及位姿不准/控制根因超出了画面所能支持的程度，conf 7 偏高。",
+    "n_issues": 3,
+    "n_major": 0
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未完成、断点（投球全中/操作中）与评分器 live_success=0 一致；视频 4.76 s 在结局前截止，文件已把球路偏左与位姿不准标为不确定推断、根因留空并说明缺口，未见需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:47:59Z",
+  "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/embodied-olympics-fun-bowling-i00-standard--aydgyd7/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--aydgyd7/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--aydgyd7/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--aydgyd7/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "f3bd3b6e513034063faa2ec854f4d14623e555d49e736047672c6c00f94e1a2a",
+   "video_sha256": "7d9be4ee52eece646638e9f54533e0fc098dfc5aa9f2d4a19274538252a69f51",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--d3sfheq.json

@@ -0,0 +1,186 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-standard--d3sfheq",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 1393,
+   "sim_steps": 57336,
+   "calls": 317,
+   "rows": 202,
+   "resets": 50,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 3976.4,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.349,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.916,
+   "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/embodied-olympics-fun-bowling-i00-standard--d3sfheq/replay.mp4",
+  "duration_s": 13.96,
+  "fps": 25.0,
+  "frames": 349,
+  "sha256": "1a6ced533abc117473b03163830686b83466a9de93bc2831485a0ef52c73ae1b"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.12,
+      "log_step": 195,
+      "text_zh": "观察：视频 0.00–0.08 s，机器人双掌在髋前夹着紫色球并向前摆臂；视频 0.12–0.20 s，球已与双掌分离，随后沿木质球道向瓶阵运动。log step 195 启动最后一次重置后的投球，log step 196 服务返回 episode 51；此前各次重置属于会话试验，不计入本段视频。"
+     },
+     {
+      "t": 9.0,
+      "log_step": null,
+      "text_zh": "观察：视频 8.80–10.00 s 的瓶区裁剪画面中，球到达瓶阵，画面右侧部分球瓶倒下或移开，左中仍有多只白色球瓶直立；视频 12.92–13.92 s 逐帧检查仍见这些直立球瓶，机器人双脚始终支撑身体。"
+     },
+     {
+      "t": 13.92,
+      "log_step": 196,
+      "text_zh": "观察：log step 196 中机器人在线服务的最终裁判状态为 episode 51、done=true、outcome=\"not a strike\"、stages_reached=2/3，表示合法送球与碰瓶已达成但全中未达成；其 control_steps=1393、sim_time_s=13.93。视频 13.92 s 的终帧仍见直立球瓶，与该裁判结果一致，因此投球全中从未成立。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–0.08 s，左右平掌共同夹压髋前紫色保龄球并向前送球，随后双掌离球。候选：夹持；两只平掌之间的夹压没有单手合拢包围或手指闭合，现有“抓握”不适用。球离手后由球与球瓶碰撞造成局部倒瓶。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–13.92 s，机器人双脚接触助走区承重，在送球和随后等待球瓶点数期间保持站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": null,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 9.0,
+   "log_step": 196,
+   "where": "操作中",
+   "evidence_zh": "观察：机器人在视频 0.12–0.20 s 已送出球，视频约 8.9–10.0 s 球进入瓶阵后仅造成局部倒瓶，左中多瓶保持直立且至视频 13.92 s 未修复；log step 196 的最终在线裁判判为“not a strike”、阶段 2/3。断点为这次投球全中步骤，视频约 9.0 s 用作入瓶结果开始显现的导航锚点。"
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "其他",
+    "t": 9.5,
+    "log_step": 196,
+    "evidence_zh": "观察：视频 8.80 s 的瓶阵与视频 9.48、10.00、12.00、13.92 s 对比，瓶阵画面右侧有球瓶倒下或离开，左中仍留有一组直立球瓶，未出现扫倒整阵的连锁碰撞。log step 196 的服务裁判同时确认已碰瓶但非全中。",
+    "note_zh": "候选：局部漏击。球确实接触并击倒部分目标，不能标“没建立接触”；远景不足以区分入瓶位姿偏差与撞击动量不足，也无力传感证据，故不强选“位姿不准”或“接触力异常”。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "观察：inputs/ 提供本次及成功参考视频、运行日志，但没有状态日志；本次视频只有 640×360 场地远景，瓶阵很小且互相遮挡，原生裁判未报告 pins_down 数值、逐瓶状态、入瓶速度或角度。推断：可以确认局部倒瓶和全中失败，不能可靠数出倒瓶总数，或确定主要原因是入瓶位置、角度、速度还是释放接触细节；根因层留空。log step 195 的参数调整意图与早先试验不能单独证明最终投球的具体控制缺陷。",
+  "unmet": {
+   "strike_completed": false
+  }
+ },
+ "summary_zh": "观察：本次评分对应最后的 episode 51，机器人从线后送出紫色球并保持站立，在线服务裁判确认合法送球与碰瓶达到 2/3 阶段；视频 13.92 s 仍有多只球瓶直立，“投球全中”节点最终为 0/1。观察：视频约 9.0–10.0 s 出现局部倒瓶而左中瓶群仍直立，断点为投球全中步骤的操作中；缺少逐瓶状态和入瓶运动数据，无法进一步确定漏击的控制根因。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未达成、断点（投球全中·操作中）、双臂协作/其他接触与脚部支撑均与视频和 log step 196/197 的裁判结果（not a strike，2/3）一致；局部漏击现象与留空根因有证据支持，未发现需改内容的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:33:32Z",
+  "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/embodied-olympics-fun-bowling-i00-standard--d3sfheq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--d3sfheq/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "29db5ff305151747eb1c798b231aa4e17ee78e01fae53c6aee201b4be9c6682e",
+   "video_sha256": "1a6ced533abc117473b03163830686b83466a9de93bc2831485a0ef52c73ae1b",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--rsflhfc.json

@@ -0,0 +1,191 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-standard--rsflhfc",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 65,
+   "sim_steps": 13370,
+   "calls": 222,
+   "rows": 2430,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 3756.7,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "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/embodied-olympics-fun-bowling-i00-standard--rsflhfc/replay.mp4",
+  "duration_s": 0.68,
+  "fps": 25.0,
+  "frames": 17,
+  "sha256": "8225e88e51cfdec2fd7d83774729baa209e9ad2dba02108e5dd7187a1f47f74c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.64 s 的全部 17 帧中，机器人站在助走区，紫色球留在髋前双掌附近；视频 0.48–0.64 s 双臂向内收拢，球没有沿球道飞出或滚出，远端球瓶未受球撞击。未出现投球全中的验收结果。"
+     },
+     {
+      "t": 0.64,
+      "log_step": 82,
+      "text_zh": "观察：log step 82 的机器人在线服务先报告 reset 到 episode 6，随后报告 done=True、outcome=\"the ball touched a part other than the hands before the release\"、stages_reached=0/3、control_steps=65，且 resets_left=0；这是最后一次受评尝试，与视频 0.64 s 球仍在身体前方的末帧相符。前五次尝试的投球或滑落不计入本次节点。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–0.64 s 左右平掌共同夹压紫色保龄球，末段双臂继续内收。候选：夹持；平掌没有能闭合包围球的手指，不标抓握。log step 82 的裁判还确认离手前球接触了非手部位，具体部位在此视角中不能辨清。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–0.64 s 机器人双脚接触助走区并支撑身体，配合髋前双掌夹球，未见跌倒。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": null,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 0.64,
+   "log_step": 82,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 0.48–0.64 s 机器人在双掌夹球准备送球时向内收臂，球仍在身体前方；log step 82 的最后一次在线尝试（episode 6）由服务裁判以“离手前球接触非手部位”立即判负，阶段为 0/3。投球已开始准备，但尚未合法送出，之后没有修复。视频 0.64 s 是末段接触事件的导航锚点，细微接触本身由服务裁判确认。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "约束",
+    "item": "违反约束",
+    "t": 0.64,
+    "log_step": 82,
+    "evidence_zh": "观察：log step 82 的 episode 6 服务裁判明确报告“the ball touched a part other than the hands before the release”，即违反出手前碰身（pre_release_body_touch）规则；视频 0.48–0.64 s 可见双臂内收且球没有送出。远景不能分辨碰身的具体部位，也未显示可独立确认的阻挡、滑脱或跌倒，故不另加这些现象。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "已验证有",
+    "log_step": 80,
+    "evidence_zh": "推断：log step 80 的已执行代码将左右肩偏航目标设为相反方向的 2.0 rad、左右腕偏航目标设为相反方向的 1.27 rad，并在摆臂前加入持续内压的夹球动作；log step 82 执行该版本，服务随即确认离手前非手部触球，视频 0.48–0.64 s 也显示内收臂而球仍贴近身体。所选夹持接触布局���球带入了不允许的身体接触，故归入接触设计，而非仅凭失败猜测感知或参数错误。agent 在 log step 83 自述球被压到前臂，具体前臂接触仍待核实。可行替代由成功参考佐证：inputs/oracle/oracle.json 的 success=1，参考视频约 0.5–1.0 s 可见从同类开局双掌持球姿态将球向球道送出；这只证明存在合法替代，不作为本次进度证据。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "已逐帧检查完整受评视频，但只有场地远景，球与掌、前臂和躯干的局部接触受本体遮挡，且本任务未提供状态日志，不能独立确认裁判所指的具体非手部位、接触力或精确触碰时刻；“压到前臂”仅为 agent 在 log step 83 的自述。原生 pins_down 数值未由评分字段或最后一次服务结果报告，故保留 null。"
+ },
+ "summary_zh": "最后一次受评尝试只完成髋前双掌夹球及内收准备，未合法送球，投球全中节点最终为 0/1。视频 0.48–0.64 s 的内收动作与 log step 82 的裁判结果相符：离手前球接触非手部位，尝试立即终止；代码中的持续内压夹持布局支持控制·抓取和接触设计归因。具体碰身部位被遮挡，前臂说法仅来自 agent 自述，原生倒瓶数未报告。",
+ "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：最后一次受评尝试（episode 6）被服务裁判以离手前非手部触球判负，节点 0/1，断点在 strike_completed，违反约束现象与抓取和接触设计归因有 log step 80/82 与成功参考支撑；具体碰身部位在远景中不可辨，文件已如实声明。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:47:46Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--rsflhfc/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--rsflhfc/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "83fc481f874b3b6a931e7bbded31fd123e5b598f5676b3277dbc5490b811d729",
+   "video_sha256": "8225e88e51cfdec2fd7d83774729baa209e9ad2dba02108e5dd7187a1f47f74c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--xwpktq7.json

@@ -0,0 +1,193 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-standard--xwpktq7",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 2929,
+   "sim_steps": 96293,
+   "calls": 320,
+   "rows": 4697,
+   "resets": 49,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 6979.5,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.585,
+   "peak_processes": 12,
+   "peak_disk_gb": 0.853,
+   "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/embodied-olympics-fun-bowling-i00-standard--xwpktq7/replay.mp4",
+  "duration_s": 29.32,
+  "fps": 25.0,
+  "frames": 733,
+  "sha256": "86aa2643a83254d74a372b06fffd9cc053ab5bd172f77d782d5971d3884b4e0c"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.12,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.08 s，机器人在髋前以双掌接触紫色球；视频 0.12 s 双掌已抬起并与球分开，球随后落到近处木质球道，视频 0.60–23.80 s 沿球道缓慢向瓶区滚动。"
+     },
+     {
+      "t": 24.4,
+      "log_step": 285,
+      "text_zh": "观察：视频 24.00–25.20 s，紫色球到达瓶阵前部，局部球瓶有移动，但后排仍见多只白色红颈球瓶直立；视频 28.32–29.28 s 末段仍有直立球瓶，机器人始终站在助走区。log step 285 中机器人服务返回 outcome=\"not a strike\"、stages_reached=2/3、control_steps=2929；这是 log step 284 启动的最后一次尝试，log step 286 确认为 episode=50、done=true，与本视频一致。"
+     },
+     {
+      "t": 29.28,
+      "log_step": 285,
+      "text_zh": "推断：最终在线尝试已通过送球及有效碰瓶阶段，但在碰瓶后的点数验收未全中，因此“投球全中”从未成立。视频 29.28 s 仍有直立球瓶，与裁判的 not a strike 相符；不从 reward=0 单独推断未投球。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–0.08 s，左右平掌在髋前从两侧夹压同一紫色保龄球，随后双臂摆动、掌球分离。候选：夹持；两只无手指掌垫不各自包围物体，故不适用抓握。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–29.28 s，机器人双脚接触木质助走区承重，送球及等待倒瓶期间保持站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": null,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 24.3,
+   "log_step": 285,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 24.20–24.60 s，缓慢滚行的紫色球进入瓶阵前部后，未引发整排球瓶连锁倾倒；视频 29.28 s 仍见直立球瓶。log step 285 的机器人服务最终返回 not a strike、stages_reached=2/3。断点锚在这次未能造成全中的碰瓶附近，时间约有 0.3 s 不确定性。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "其他",
+    "t": 24.3,
+    "log_step": 285,
+    "note_zh": "候选：出球速度不足。球已到达并有效接触球瓶，主要可见问题是缓慢入瓶、未产生足够的连锁倒瓶；不是没有接触，也没有证据证明单纯位置偏差，现有条目不能准确描述这一速度表现。",
+    "evidence_zh": "观察：视频 0.12 s 球已离掌，视频 0.60–23.80 s 球缓慢滚过球道，约视频 24.30 s 才接触瓶阵；视频 24.40–29.28 s 瓶阵只发生局部变化，仍有多只瓶直立。推断：这次送球的入瓶速度不足以产生所需的全中倒瓶链；未测量具体速度或冲量。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "视频 / 日志",
+    "evidence_zh": "推断：机器人双臂送球与离手动作未给���形成足够有效的前向运动；视频 0.00–0.52 s 可见双掌迅速抬离、球落在机器人近处，随后至视频约 24.30 s 才缓慢到瓶阵。log step 279、283 中 agent 自述正在保留出球动作并微调入瓶方向，log step 284–285 执行最后版本仍得到 not a strike，支持送球控制未调成，而非目标理解错误。日志未保留最后几次 edit 的完整补丁，无法进一步归到某个已证实错误的参数或接触几何。"
+   }
+  ],
+  "gap_zh": "无状态日志；视频只有 640×360 场地远景，远端瓶子重叠且像素很少，无法可靠逐瓶计数或测量倾角、入瓶速度和接触冲量，原生 pins_down 数值也未报告，故该指标留 null。最后几次脚本修改仅记录为 edit，缺少完整内容，控制根因不细分。成功参考的同一远景在参考视频 9.64 s 可见瓶台已无成排直立瓶，但参考不作为本次尝试的倒瓶证据。"
+ },
+ "summary_zh": "最终评分尝试中，机器人双掌夹球后送出，球沿球道滚到瓶阵，裁判记录送球和碰瓶两阶段完成，但“投球全中”节点始终未达成，最终为 0/1。视频约 24.30 s 的入瓶缓慢，未产生全中所需的连锁倒瓶，至视频 29.28 s 仍有多只瓶直立；根因归为送球控制未调成。无状态日志且仅有远景，倒瓶个数、入瓶速度及具体控制子因无法精确确定。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "标签与视频、日志一致：最终评分尝试（log step 284–285，control_steps=2929，not a strike，stages 2/3）球离手后缓慢滚到瓶阵，视频 29.28 s 仍有直立球瓶，节点未达成、断点与时间可接受；现象“出球速度不足”有视频依据（对照成功参考球约 3–4 s 即到瓶阵，本次约 24 s），根因停在“控制”未细分也符合分类规则。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:47:59Z",
+  "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/embodied-olympics-fun-bowling-i00-standard--xwpktq7/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--xwpktq7/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "2a5cf4d35e9a96943300309d39f6de5d3bf71bed99e947fecc0e2547ac51a0f8",
+   "video_sha256": "86aa2643a83254d74a372b06fffd9cc053ab5bd172f77d782d5971d3884b4e0c",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--yyaavju.json

@@ -0,0 +1,202 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-standard--yyaavju",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 1455,
+   "sim_steps": 33483,
+   "calls": 232,
+   "rows": 17812,
+   "resets": 50,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 5504.0,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.539,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.516,
+   "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/embodied-olympics-fun-bowling-i00-standard--yyaavju/replay.mp4",
+  "duration_s": 14.56,
+  "fps": 25.0,
+  "frames": 364,
+  "sha256": "b0735882ed7f498ef9b477500704cd99a7e2f8580db22bfce435bdec1bec0979"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0 s 球仍在机器人髋前双掌之间，远端球瓶组直立，投球全中条件尚未成立。"
+     },
+     {
+      "t": 1.36,
+      "log_step": 313,
+      "text_zh": "观察：视频 1.08–1.24 s 机器人双掌共同承托紫球向前上方摆动，视频 1.36–1.48 s 球与双掌明显分离并向球道飞出；机器人双脚留在犯规线后的助走区。log step 313 执行最后一次重置后的 throw24 投球，随后没有再重置。"
+     },
+     {
+      "t": 10.3,
+      "log_step": null,
+      "text_zh": "观察：视频 9.8–10.6 s 球进入瓶区画面偏右部分，部分球瓶倾倒，而画面左侧仍有直立球瓶；视频 11.6 s 及视频 13.6–14.52 s 的末尾画面仍可见直立瓶形，机器人保持站立。"
+     },
+     {
+      "t": 14.52,
+      "log_step": 314,
+      "text_zh": "观察：log step 314 的机器人仿真服务返回 done True、outcome 为 not a strike、reason 为 failed；这是 log step 313 最后一次在线投球的裁判结果，与视频末尾仍有直立球瓶一致。成功参考视频 7.0–9.5 s 则可见瓶台上已无直立瓶形；本次没有达到该结果。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.0–0.3 s 左右掌垫在髋前两侧共同夹住紫球。候选：夹持；两块平掌垫对压，单手没有合拢包住球，故不标抓握。"
+     },
+     {
+      "who": "双臂协作",
+      "how": "托/舀",
+      "note_zh": "观察：随后双掌转成朝上的承托姿态，视频 0.8–1.24 s 球位于两掌上方并随双臂前摆升高，约视频 1.3 s 离手；此承托发生在开局夹持之后。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–14.52 s 双脚在犯规线后的助走区承重，送球和等待碰瓶期间机器人始终站立。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": null,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 10.3,
+   "log_step": 314,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 10.0–10.6 s，机器人送出的球从瓶组画面偏右处进入，部分球瓶倒下，另一侧仍有球瓶直立；视频 14.52 s 仍未清空瓶台上的直立球瓶。log step 314 的仿真服务裁判对这次最终在线投球返回 not a strike，因此失败步骤为投球全中。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 10.3,
+    "log_step": 314,
+    "evidence_zh": "观察：视频 9.8–10.6 s 球沿画面偏右的入瓶线路撞入球瓶组，只带倒部分球瓶，画面左侧的直立瓶形保留至视频 14.52 s；log step 314 裁判返回 not a strike。推断：此次送球的入瓶位置和方向未形成全中所需的碰撞效果。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：log step 313 的最终控制脚本将躯干���航目标设为 -0.055，并执行固定摆臂、开掌和制动轨迹；视频 1.36 s 可见球已正常送离双掌，视频 10.0–11.6 s 却见入瓶碰撞集中于画面偏右部分，另一侧球瓶仍站立。证据支持送球轨迹的控制未达到所需入瓶效果；不足以细分为具体增益、速度、旋转或反馈问题。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "本次视频仅有场地远景，无状态日志，评分器字段也未报告 pins_down 数值；无法可靠数清具体倒瓶总数、逐瓶身份，或测定入瓶速度、旋转和精确接触力。因此原生倒瓶数留为 null，控制根因不细分；不将 agent 在 log step 317、320 自述的“约 7–8 瓶倒下”当作评分器计数。"
+ },
+ "summary_zh": "最后一次评分投球中，机器人由双掌夹持转为承托并将球送入球道，保持站立，球到达瓶区并带倒部分球瓶，但“投球全中”节点始终未完成，终态为 0/1。视频约 10.0–10.6 s 可见入瓶碰撞偏于画面一侧，末尾仍有直立球瓶，log step 314 裁判确认 not a strike，支持位姿不准及控制层面的失败。远景和缺少逐瓶状态使准确倒瓶数及更细控制原因无法确定，早先重置试验中的跌倒和入沟不计入本次评分轨迹。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与视频一致：球约 1.4 s 离手，约 8–10.6 s 到达瓶区，部分球瓶倒下，末尾仍有直立瓶，节点未完成，终态 0/1，断点与位姿不准、控制归因有据且未过度细分。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:33:27Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--yyaavju/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--yyaavju/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "a5b0ee0d11c97cbbc8105d98f15f1613e3a8b53e440f50345b6049c65aaf8c52",
+   "video_sha256": "b0735882ed7f498ef9b477500704cd99a7e2f8580db22bfce435bdec1bec0979",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/bowling/trajectories/embodied-olympics-fun-bowling-i00-standard--z8v7dg3.json

@@ -0,0 +1,193 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-bowling-i00-standard--z8v7dg3",
+ "taskKey": "embodied-olympics-fun/bowling",
+ "dag_sha256": "958dbc6f1b15aea49b7fbb6f67f6fe11ebb2ebd0b8c66ca9ec74face659152c2",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 1465,
+   "sim_steps": 55615,
+   "calls": 272,
+   "rows": 6585,
+   "resets": 50,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 2764.0,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.481,
+   "peak_processes": 12,
+   "peak_disk_gb": 1.815,
+   "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/embodied-olympics-fun-bowling-i00-standard--z8v7dg3/replay.mp4",
+  "duration_s": 14.68,
+  "fps": 25.0,
+  "frames": 367,
+  "sha256": "853ef7bf6a2fa7070079b3febd692a2484f9f4e141356178164d0db382e8b248"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "strike_completed": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：视频 0.00–0.52 s，机器人站在犯规线后，紫球保持在髋前两只平掌之间；视频 0.56–0.64 s 双臂前摆，球与双掌分离，随后落到木质球道并向瓶阵滚动。"
+     },
+     {
+      "t": 9.64,
+      "log_step": 153,
+      "text_zh": "观察：视频 9.28–10.40 s，紫球从瓶阵画面左侧进入，部分球瓶倾倒，仍有多根白色瓶身直立。log step 153 是最后一次在线尝试 fin.json，机器人服务返回 stages_reached=2、outcome=\"not a strike\"、episode=51、control_steps=1465，与评分记录 n_actions=1465 及本段视频相符；前面的重置测试不计入本次进度。"
+     },
+     {
+      "t": 14.64,
+      "log_step": 153,
+      "text_zh": "观察：视频 13.68–14.64 s 的末段密集帧仍显示多只球瓶直立，机器人双脚站在助走区，未见跌倒；log step 153 的服务终局为 not a strike、2/3 阶段。推断：本次合法送球与有效碰瓶已经完成，但碰瓶后的点数未达到十瓶全倒，投球全中节点始终未成立。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–0.56 s，左右平掌从两侧共同夹压紫球并随双臂前摆送出。候选：夹持；两只平掌之间的夹压没有单手包围或夹爪闭合，不符合抓握，故按 D12 使用其他。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.00–14.64 s，机器人双脚接触助走区承重，配合双臂送球，随后保持站立等待倒瓶结果。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "pins_down",
+   "value": null,
+   "unit": "个",
+   "threshold": 10,
+   "direction": "higher_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "strike_completed",
+   "t": 9.64,
+   "log_step": 153,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 9.6 s，机器人送出的紫球进入瓶阵的一侧，只引发局部倒瓶；视频 10.4–14.64 s 仍有多只瓶直立，未形成全中。log step 153 的最后一次在线尝试由机器人服务判为 not a strike、stages_reached=2/3；失败发生在投球全中这一步，非前面重置测试中的入沟或跌倒。"
+  },
+  "unmet": {
+   "strike_completed": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 9.64,
+    "log_step": 153,
+    "evidence_zh": "观察：视频 9.28–10.40 s 的瓶区连续帧中，紫球由画面左侧切入瓶阵，局部瓶身倒下，画面右侧仍保留多根直立白色瓶身；视频 14.64 s 这些残瓶仍未倒下。推断：面对静止瓶阵，这次入瓶路径偏侧，未产生覆盖全阵的撞击结果，记为位姿不准；不把主动送球离手记成没保持住。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "log_step": 153,
+    "evidence_zh": "推断：送球的方向和速度控制未产生全中所需的入瓶运动。log steps 151–152 显示 agent 脚本采用固定的双肩加速、前臂转动与张臂时序；log step 153 的脚本视觉追��估计滚动速度约 2.05 m/s，并估计球在 x≈18.45 米处横向 y≈0.15 米（这些是 agent 的估计，非原生状态）。视频 0.56–9.64 s 可核对球离手后缓慢滚向瓶阵并偏侧入瓶、留下残瓶；agent 在 log step 155 自述无法稳定瞄准。证据支持送球控制不足，但不能区分接触布局、轨迹跟踪或具体参数哪个是主要原因，故停留在控制大项。"
+   }
+  ],
+  "gap_zh": "无状态日志，所提供的本次视频只有场地远景；虽可见残瓶，无法可靠逐瓶计数或测量碰撞速度、旋转和接触力。run.json 的 reward 与最终服务输出未报告 pins_down 数值，故原生指标留 null；agent 的视觉估计不足以分离球速与横向偏移对残瓶的贡献，也不支持其“无法投得更快”的物理上限说法。"
+ },
+ "summary_zh": "评分对象是最后一次在线投球：机器人双掌夹球前摆，约视频 0.6 s 离手、视频 9.6 s 有效碰瓶，但至视频 14.64 s 仍有多只球瓶直立，投球全中未完成，节点终值 0/1。断点为投球全中的操作中，偏侧入瓶表现为位姿不准，结合固定送球脚本与最终日志推断为控制不足。远景及无状态日志限制了逐瓶计数和球速、旋转、落点贡献的细分，pins_down 未报告。",
+ "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（0/1）、断点、位姿不准与控制大项均与视频及 log step 153 的服务读数（not a strike、2/3 阶段）一致；未发现需要改动的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:33:58Z",
+  "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/embodied-olympics-fun-bowling-i00-standard--z8v7dg3/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-bowling-i00-standard--z8v7dg3/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b99759783a2b7606d78e6d7bdefb2bfa5a5f9d9307ec0cbd47eea6b1617a2c09",
+   "video_sha256": "853ef7bf6a2fa7070079b3febd692a2484f9f4e141356178164d0db382e8b248",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--87ojykg.json

@@ -0,0 +1,198 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-privileged--87ojykg",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.75,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1220,
+   "sim_steps": 1220,
+   "horizon": 4000,
+   "episode_over": 0,
+   "progress": 0.75,
+   "stages_reached": 3,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 13.36,
+   "bad_trajectory": 0,
+   "metric_air_s": 1.705,
+   "metric_from_knoll_m": 0.99,
+   "metric_height_m": 2.54,
+   "metric_speed_m_s": 15.569,
+   "metric_trick_flip_deg": 16.4,
+   "metric_trick_flips": 0.0,
+   "metric_trick_half_turns": 2.0,
+   "metric_trick_pelvis_spin_deg": 278.7,
+   "metric_trick_short_deg": 50.6,
+   "metric_trick_spin_deg": 410.6,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 52,
+   "agent_resets": 33,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 1,
+   "peak_gpu_memory_gb": 0.0,
+   "peak_memory_gb": 0.0,
+   "peak_processes": 0.0,
+   "peak_disk_gb": 0.0,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-freeski-360-i00-privileged--87ojykg/replay.mp4",
+  "duration_s": 12.24,
+  "fps": 25.0,
+  "frames": 306,
+  "sha256": "229f78dacc32fcd75206db661e6e538b2dcc8157e55df097a0a64bcbbffbce94"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": null,
+      "text_zh": "观察：inputs/run.json 的评分器 reward 报告 stages_reached=3、stages_total=4、progress=0.75、success=0、replay_success=0、episode_over=0、fell=0；交付轨迹含1220个动作，deterministic=1、bad_trajectory=0。评分器记录空中雪板转角410.6度、空翻数0。"
+     },
+     {
+      "t": null,
+      "log_step": null,
+      "text_zh": "推断：结合DAG规定的起滑、起跳、落地、滑出顺序与评分器仅接受三个阶段，交付轨迹未完成受控滑出终点，故完整事件节点从未达成、final=false；原生0.75不是该单节点的完成比例。"
+     },
+     {
+      "t": 12.2,
+      "log_step": null,
+      "text_zh": "观察：逐0.5秒检查视频全程，并密查视频10.0–12.2 s及视频12.20 s最后一帧，画面均是固定雪场远景，近景坡面遮住主要运动区域，无法可靠辨认机器人及雪板；成功参考视频的全程抽帧与参考视频17.48 s帧也有同样的可见性缺口。不能从这些空景判断转体、落地接触或过线的时刻，节点值依据本次评分器字段。"
+     },
+     {
+      "t": null,
+      "log_step": 128,
+      "text_zh": "观察：agent在log step 128编写的调参脚本于done或phase等于landed时检查结果，仅在landing阶段已接受且落地评分改善时调用save_trajectory，随后break；log step 114和124也按落地阶段保存候选。它们是会话保存逻辑，未提供可与本次视频逐事件核对的最终候选状态记录。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频0.0–12.2 s无法可靠辨认机器人双靴、雪板与雪面的接触，且本次没有状态日志；不能把DAG的腿/躯干经雪板支撑方案直接当作已观察接触。评分器的阶段进度说明已有动作尝试，故保留未知接触而非省略。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "value": 410.6,
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target"
+  },
+  {
+   "name": "landing_heading_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better"
+  },
+  {
+   "name": "ride_turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "推断：本次评分器reward的三个已接受阶段及410.6度转体、0次空翻记录表明连续大跳台动作已有进展，但完整转体落地滑出事件尚未完成；episode_over=0且fell=0，并无已报告的摔倒终止。断点定位到这一唯一动作步骤，视频0.0–12.2 s无法辨认机器人，不能定位物理失误瞬间，故t=null。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [],
+  "root_causes": [],
+  "gap_zh": "观察：本次inputs/内有视频和运行日志，但没有状态日志；视频全程抽帧、视频10.0–12.2 s密查及最后一帧均无法可靠辨认机器人，成功参考也不能补足本次动作证据。评分器未报告landing_heading_deg或ride_turn_deg，二者保留null；会话log step 119的落地方向错误属于另一调参试验，不能移植到本次交付。log step 134为提供方过滤中断，发生在agent继续调参且没有完整成功验收记录时；reward报告agent_replays=0、episode_over=0，不能证实过滤前已固定完整结果，因此保留这一外因中断信息，不把中断或未滑出本身归成机器人现象或根因。推断：现有证据只足以判定完整节点未达成，不足以区分姿态、平衡、滑行控制或保存选择的责任，也不能给出视频断点时间，故两层均为空。"
+ },
+ "summary_zh": "评分器记录本次交付达到起滑、起跳、落地三个原生阶段，雪板空中转角410.6度且空翻数为0，但未完成受控滑出，DAG完整节点最终为0/1。视频全程及末段均无法可靠辨认机器人，接触和物理失误时刻无法核实；会话最后在log step 134被提供方过滤中断，未见完整结果验收，故不臆造机器人失败现象或根因。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与证据一致：视频全程为固定远景、近景坡面遮挡运动区，看不到机器人；节点final=false、断点定位到唯一节点、两层为空并有gap_zh，均由评分器字段（stages 3/4、success=0）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": "2026-10-31T00:45:40Z",
+  "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/embodied-olympics-fun-freeski-360-i00-privileged--87ojykg/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--87ojykg/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "228984e4e405f48d3624abcdef9d71c3ca3ab650d7104b30fdc05d34a79c8c3f",
+   "video_sha256": "229f78dacc32fcd75206db661e6e538b2dcc8157e55df097a0a64bcbbffbce94",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--bcpa6ao.json

@@ -0,0 +1,202 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-privileged--bcpa6ao",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.75,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1557,
+   "sim_steps": 1557,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.75,
+   "stages_reached": 3,
+   "stages_total": 4,
+   "fell": 1,
+   "replay_s": 15.32,
+   "bad_trajectory": 0,
+   "metric_air_s": 1.443,
+   "metric_from_knoll_m": -3.61,
+   "metric_height_m": 1.93,
+   "metric_speed_m_s": 13.399,
+   "metric_trick_flip_deg": 70.3,
+   "metric_trick_flips": 0.0,
+   "metric_trick_half_turns": 2.0,
+   "metric_trick_pelvis_spin_deg": 249.1,
+   "metric_trick_short_deg": 1.3,
+   "metric_trick_spin_deg": 361.3,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 238,
+   "agent_resets": 0,
+   "agent_replays": 0,
+   "agent_cost_self_reported": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 3.344,
+   "peak_processes": 15,
+   "peak_disk_gb": 0.29,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-freeski-360-i00-privileged--bcpa6ao/replay.mp4",
+  "duration_s": 15.6,
+  "fps": 25.0,
+  "frames": 390,
+  "sha256": "d2e7e5856613f4e3393152db38cf4def803c55851d1e45c87539770cb197864d"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 15.56,
+      "log_step": null,
+      "text_zh": "观察：已检查本次视频全程逐秒远景，以及视频 0.00 s、3.56 s、14.12 s 原分辨率帧和视频 14.56–15.56 s 的全部逐帧画面；近景雪坡遮挡主要动作区域，只能辨认场地及极小的暗点，无法可靠辨认机器人姿态、雪板转体、落地接触或越线。成功参考视频全程逐秒画面和视频 17.52 s 帧也有相同可见性缺口，不能用本次画面中未见动作独立判定未完成。"
+     },
+     {
+      "t": null,
+      "log_step": 139,
+      "text_zh": "观察：inputs/run.json 的评分器字段为 success=0、replay_success=0、deterministic=1、stages_reached=3/4、fell=1、metric_trick_spin_deg=361.3、metric_trick_flips=0；log step 139 重新运行并保存交付生成脚本，服务结果也为 start、take-off、landing 三阶段和 failed=fell，动作总数1557与评分记录相同。"
+     },
+     {
+      "t": null,
+      "log_step": 139,
+      "text_zh": "推断：评分器明确未接受 ride-away，且整套动作是仅在受控滑出后才锁存的唯一 event 节点，因此该节点从未完整达到、终值为 false；原生3/4阶段不折算为DAG节点进度。此结论依赖评分结果，不将会话测试中的落地宣称当作视觉完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：本次视频全程远景及视频 14.56–15.56 s 逐帧画面均不能可靠辨认双靴、雪板或身体的触雪关系，且未提供状态日志；实际作用端和接触方式无法核实，不照抄任务级支撑方案。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "value": 361.3,
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target"
+  },
+  {
+   "name": "landing_heading_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better"
+  },
+  {
+   "name": "ride_turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 139,
+   "where": "操作中",
+   "evidence_zh": "推断：本任务只有“转体落地滑出”一个连续动作节点。评分器在 inputs/run.json 报告三阶段完成后 fell=1，未完成第四阶段；log step 133、139 的交付生成及保存过程同样报告落地后 failed=fell，故失败发生在该节点的落地滑出过程中。视频 14.56–15.56 s 逐帧检查不能辨认摔倒瞬间，未把日志中的仿真时间直接写成视频断点时间。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 129 保存所选控制器轨迹时，脚本诊断输出的骨盆竖轴水平分量由仿真 15.11 s 的(-0.13,0.21)增至仿真 15.51 s 的(-0.38,0.81)，随后服务先报告 landing 再报告 fell、触雪部位head；log step 133、139 运行并保存同一生成脚本，复现361.3度转体、0次空翻、1557步及相同失败结果，与评分器字段一致。这支持所交付控制器未能稳定落地姿态的控制根因；这些是会话内的数值诊断，不能据此宣称视频直接显示了头部触雪，也不足以确定反馈、接触设计或某个参数的具体责任。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 提供本次及成功参考视频，但均为主要动作区被雪坡遮挡的固定远景；本次视频 14.56–15.56 s 已逐帧检查，仍无法核实摔倒姿态、雪板及身体接触或精确失败时刻。没有 inputs/states.jsonl。评分器的 fell=1 支持失败事实，但不提供可见的失败运动细节，因此现象层留空、contact_actual记未知、视频断点时间留null；交付脚本的会话诊断只用于控制根因推断。landing_heading_deg 与 ride_turn_deg 未出现在 run.json 的 reward 字段中，原生指标值留null，未用会话结果代填。"
+ },
+ "summary_zh": "评分器记录起滑、起跳、落地三个原生阶段完成，雪板转体361.3度、空翻0次，但随后摔倒，未完成受控滑出；DAG唯一节点“转体落地滑出”未达到，最终0/1。交付生成脚本的重复运行和姿态诊断支持落地姿态控制未稳定的根因。两段视频均为遮挡严重的场地远景，无法直接核实接触、失败现象及视频断点时刻，相关字段保留未知或空缺。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未达成（评分器 success=0、fell=1、stages 3/4）、终值0、断点ref正确且t留null、contact_actual记未知均有依据；视频为被雪坡遮挡的远景，无法核实动作，现象层留空并带gap_zh可接受。未发现需改动的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:44:01Z",
+  "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/embodied-olympics-fun-freeski-360-i00-privileged--bcpa6ao/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--bcpa6ao/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "73090ce7ad6e781cc2a62600a3d90127e647b3dc1e3c5d45317625730a59ce81",
+   "video_sha256": "d2e7e5856613f4e3393152db38cf4def803c55851d1e45c87539770cb197864d",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--bgejpwd.json

@@ -0,0 +1,235 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-privileged--bgejpwd",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.75,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1602,
+   "sim_steps": 1602,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.75,
+   "stages_reached": 3,
+   "stages_total": 4,
+   "fell": 1,
+   "replay_s": 12.86,
+   "bad_trajectory": 0,
+   "metric_air_s": 1.628,
+   "metric_from_knoll_m": -0.33,
+   "metric_height_m": 2.35,
+   "metric_speed_m_s": 20.501,
+   "metric_trick_flip_deg": -32.7,
+   "metric_trick_flips": 0.0,
+   "metric_trick_half_turns": 2.0,
+   "metric_trick_pelvis_spin_deg": 177.9,
+   "metric_trick_short_deg": 13.3,
+   "metric_trick_spin_deg": 373.3,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 378,
+   "agent_resets": 3159,
+   "agent_replays": 4,
+   "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": 4.619,
+   "peak_processes": 34,
+   "peak_disk_gb": 0.246,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-freeski-360-i00-privileged--bgejpwd/replay.mp4",
+  "duration_s": 16.04,
+  "fps": 25.0,
+  "frames": 401,
+  "sha256": "1e84a90f4ff01c45cfab8d68324bd9a1b9d1371bfa8cf7458d35fafd36e32673"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 158,
+      "text_zh": "观察：inputs/run.json 的评分字段为 success=0、replay_success=0、stages_reached=3/4、fell=1、deterministic=1、n_actions=1602；log step 157 重新生成最终交付文件，log step 158 对该文件重放，输出1602行且全部有限，裁判阶段只有 start、take-off、landing，随后 failed=fell。其转体373.3度、空翻角-32.7度和腾空1.628秒与评分字段一致。"
+     },
+     {
+      "t": null,
+      "log_step": 158,
+      "text_zh": "推断：该节点要求整次转体、落地并受控滑出终点，而最终文件的裁判在落地后因摔倒终止，未到滑出阶段，故整个事件从未完成；原生三阶段通过不等于此单节点完成。log step 156 的雪上转体失败属于随后被覆盖的另一个文件，不作为本次失败现象。"
+     },
+     {
+      "t": 16.0,
+      "log_step": null,
+      "text_zh": "观察：本次视频0.0–15.5 s以0.5秒间隔检查，并逐帧检查视频15.0–16.0 s，画面主要是固定场地远景，近景雪坡遮挡滑行区域；原分辨率视频13.8 s及16.0 s也不能可靠辨认机器人和雪板。成功参考视频0.0–17.0 s及17.5 s同样缺少可辨认的完整动作，不能据这些空景判断是否起跳、落地或触雪；节点结论来自明确的裁判结果。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：本次视频全程抽帧及视频15.0–16.0 s逐帧画面无法可靠显示机器人的作用端与雪面的接触，也没有状态日志。log step 158 的最终文件重放裁判报告左手触雪，但帧中不能复核实际支承及触雪方式，故不从任务的典型雪板支撑方案复制接触标签。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "value": 373.3,
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target"
+  },
+  {
+   "name": "landing_heading_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better"
+  },
+  {
+   "name": "ride_turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 158,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 157 生成并保存最终文件，log step 158 重放该文件，裁判记录 landing 后 failed=fell、on_the_snow=[left_hand]，与本次评分 fell=1、三阶段通过相符。推断：失败发生在整套转体动作的落地后滑出过程中，故断点为转体落地滑出；本次视频13.8 s和末段视频15.0–16.0 s无法辨认失衡发生时刻，t留空。log step 158 的仿真13.802 s落地、仿真16.019 s触雪仅作为重放事件记录，未换算为视频时刻。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": null,
+    "log_step": 158,
+    "source_zh": "日志",
+    "evidence_zh": "观察：inputs/run.json 的评分字段为 fell=1、stages_reached=3/4、deterministic=1。log step 157 生成最终交付文件，log step 158 对该文件的重放裁判报告 landing 后 failed=fell、on_the_snow=[left_hand]；转体373.3度、空翻角-32.7度、腾空1.628秒及1602行动行均与本次评分一致，支持这是被评分交付的摔倒现象，而非其他会话测试或 agent 自述。视频全程逐秒抽帧、视频13.8 s原图及视频15.0–16.0 s逐帧画面无法可靠辨认摔倒姿态或时刻，t留空；不把仿真事件时间换算为视频时间。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": null,
+    "log_step": 158,
+    "source_zh": "日志",
+    "evidence_zh": "推断：落地后稳定滑出的控制未能维持机器人姿态。会话诊断 log step 146 使用后来最终交付所选的cfg6配置，输出落地后躯干倾角从37度逐渐增至78度、骨盆高度从0.52米降至0.24米，最终手触雪；这些是会话测试证据，不作为本次视频的姿态或时间。log step 137 修改航向阻尼反馈后仍有摔倒或雪上转体的测试结果，最终生成与重放（log steps 157–158）仍在落地后以fell结束，评分也报告fell=1，支持宽泛控制归因；不足以证明某一增益、逆解或具体反馈环节是本次根因。"
+   }
+  ],
+  "gap_zh": "观察：inputs/有本次与成功参考视频、run.json，但没有逐帧状态日志。本次视频全程抽帧、视频13.8 s原图及视频15.0–16.0 s逐帧检查均无法可靠分辨机器人、雪板或手触雪；参考视频也存在同样可见性限制。评分明确报告fell=1，最终文件重放裁判在log step 158报告左手触雪，但无法独立视觉复核摔倒姿态、实际接触方式及其视频时刻，故依评分字段及最终文件重放裁判记录摔倒现象“全身·平衡”，接触记未知、现象与断点的t均留空；不把会话里其他文件的摔倒或雪上转体直接当作本次可见现象。根因仅由会话诊断及最终文件裁判结果支持宽泛控制归因，具体失稳机制未确认。评分未报告landing_heading_deg和ride_turn_deg，两个原生指标保留null。"
+ },
+ "summary_zh": "评分及最终文件重放报告机器人完成起滑、起跳和落地三个原生阶段，雪板空中转体373.3度且无计入空翻，但落地后滑出时因左手触雪被裁判判为摔倒；DAG唯一节点“转体落地滑出”未完成，最终为0/1。会话姿态诊断及最终文件失败结果支持宽泛的控制问题归因。本次与参考视频的固定远景均无法可靠辨认动作，缺少状态日志，故摔倒现象“全身·平衡”依据评分及最终文件重放裁判记录，具体姿态、接触方式和视频时刻仍无法视觉复核。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 7,
+    "summary_zh": "节点未完成、0/1、断点落在转体落地滑出均与评分字段（success=0、stages 3/4、fell=1）一致，视频确为固定空景，未发现时钟或节点值错误。但现象层留空不当：评分 fell=1 与 log step 158 重放裁判 failed=fell 是服务/裁判读数，已足以支持「全身·平衡」现象，不需要视频辨认。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "节点未完成、0/1、断点在转体落地滑出、现象全身·平衡（fell=1，log 158 裁判读数）及宽泛控制根因均有评分字段和日志支持；视频为固定空景，时间均留空，未发现时钟或节点错误。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": "2026-10-31T00:48:54Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--bgejpwd/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--bgejpwd/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--bgejpwd/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--bgejpwd/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9f235a3a0d2fee834d1e267f68a3ca652bc53c36cfe7eb8ea0b1c6db6e3e3674",
+   "video_sha256": "1e84a90f4ff01c45cfab8d68324bd9a1b9d1371bfa8cf7458d35fafd36e32673",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--pabks3e.json

@@ -0,0 +1,203 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-privileged--pabks3e",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.25,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1382,
+   "sim_steps": 1382,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.25,
+   "stages_reached": 1,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 14.44,
+   "bad_trajectory": 0,
+   "metric_speed_m_s": 14.573,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 382,
+   "agent_resets": 13,
+   "agent_replays": 3,
+   "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": 2.744,
+   "peak_processes": 16,
+   "peak_disk_gb": 0.272,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-freeski-360-i00-privileged--pabks3e/replay.mp4",
+  "duration_s": 13.84,
+  "fps": 25.0,
+  "frames": 346,
+  "sha256": "b134ede4458142772526f7184c4dcff6835dd562ff4c9f840253fc523f7ca1d8"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 160,
+      "text_zh": "观察：log step 160 中，基准 CLI 对已交付 trajectory.npz 的独立重放返回 failed=\"no jump\"、stages=[\"start\"]、executed_rows=1382、bad_row=null、instance_ok=true、speed_m_s=14.573。log step 176 最终重新生成并验证的文件仍为1382行、同一状态散列和同一阶段；这些行数、阶段数及速度与本次 reward 一致。"
+     },
+     {
+      "t": null,
+      "log_step": 176,
+      "text_zh": "推断：交付轨迹只被裁判接受了起滑，未完成起跳及其后的空中转体、落地滑出；本节点要求整次动作完成，因此全程未达成，final=false。原生阶段1/4不等于语义节点1/1。"
+     },
+     {
+      "t": 13.8,
+      "log_step": null,
+      "text_zh": "观察：本次视频全程按约0.5秒间隔检查，另检查视频12.8–13.8 s每约0.1秒的末段画面及视频12.0 s、13.8 s原分辨率帧，均为近景雪坡遮挡的固定场地远景，无法可靠分辨机器人及雪板。成功参考视频全程及参考视频12.0 s、17.6 s原分辨率帧也有同样可见性缺口；不能用空景证明本次没有动作。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "裁判重放确认本次已起滑并尝试整次滑雪动作，但视频无法可靠辨认机器人、双靴、雪板及雪面之间的接触，且无状态接触日志；不把任务级典型的腿/躯干支撑直接当作实际观测。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 160,
+   "where": "操作中",
+   "evidence_zh": "推断：log step 151 保存的VY07轨迹、log step 160 独立重放及log step 176 最终重放均只达到start，裁判明确返回no jump；本次reward也只有1/4阶段且fell=0。故断在转体落地滑出这一个连续动作的起跳环节，属于操作中。视频12.8–13.8 s看不清机器人，无法定位动作出错的可见瞬间，t留空，不把日志仿真时间直接换成视频时间。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": null,
+    "log_step": 160,
+    "evidence_zh": "推断：已交付轨迹的基准重放器在log step 160 返回failed=\"no jump\"且stages仅为start，最终交付复核log step 176仍相同；本次reward的fell=0。可确定机器人已开始滑行但未实现合格起跳，归为全身·步态。此判断来自交付文件的裁判重放结果，远景视频本身不能显示起跳不足的具体姿态，不将其描述为目视摔倒。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": null,
+    "log_step": 141,
+    "evidence_zh": "推断：agent自写的VY07诊断探针在log step 141 的会话测试中读到接近唇口时竖直速度约-0.26至-0.14米每秒，探针标记的雪面法向力仍约380–458牛；log step 150 的VY07对照也给出接近唇口时vz=-0.24、fn=380.5。该配置随后在log step 151 被保存，log step 160及176的交付重放持续报no jump。综合这些会话诊断与交付验收，支持控制器未产生有效上跳运动的控制根因；探针不是逐帧状态日志，不据其断言精确接触、仿真穿透或某个参数必然错误。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "本次视频与成功参考视频都主要呈现被近景雪坡遮挡的固定远景；本次视频末秒密集检查及原分辨率帧仍无法可靠辨认机器人和雪板，因此无法视觉核验实际接触、腾空窗口、具体姿态或失败时刻。inputs/没有states.jsonl；节点未达及步态现象依据最终交付文件的基准重放结果和一致的reward，控制根因仅由同配置会话诊断辅助推断，不能细分为反馈控制、轨迹或参数问题。DAG所列三个原生转体/落地方向指标均未由本次reward报告，保留null。"
+ },
+ "summary_zh": "本次交付轨迹经独立重放只达到起滑阶段（原生阶段1/4），因no jump结束，未完成“转体落地滑出”，语义节点最终为0/1。失败位于连续动作的起跳环节，依据裁判结果标全身·步态，同配置诊断支持控制器未产生有效起跳的控制根因；本次及成功参考视频均无法可靠辨认机器人，实际接触和失败视频时刻无法核验。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未达、final=false、断点（操作中、t留空）、步态现象及控制根因均与证据一致：视频全程为被雪坡遮挡的空景，无法辨认机器人，裁判重放（failed=no jump, stages=[start], reward 0.25, fell=0）一致。log step 引用采用 step id，与run.json对应。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:46:04Z",
+  "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/embodied-olympics-fun-freeski-360-i00-privileged--pabks3e/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--pabks3e/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "b5ad1a721aa7954bc41fc379b007c1c674f27ac5c3b950be585a879a87b0ea0f",
+   "video_sha256": "b134ede4458142772526f7184c4dcff6835dd562ff4c9f840253fc523f7ca1d8",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--uismxwn.json

@@ -0,0 +1,209 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-privileged--uismxwn",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.5,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1991,
+   "sim_steps": 1991,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.5,
+   "stages_reached": 2,
+   "stages_total": 4,
+   "fell": 0,
+   "replay_s": 23.45,
+   "bad_trajectory": 0,
+   "metric_air_s": 1.393,
+   "metric_from_knoll_m": -4.89,
+   "metric_height_m": 1.92,
+   "metric_speed_m_s": 12.379,
+   "metric_trick_flip_deg": -178.7,
+   "metric_trick_flips": 0.0,
+   "metric_trick_half_turns": 2.0,
+   "metric_trick_pelvis_spin_deg": -199.8,
+   "metric_trick_short_deg": 72.3,
+   "metric_trick_spin_deg": -287.7,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 163,
+   "agent_resets": 151,
+   "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/embodied-olympics-fun-freeski-360-i00-privileged--uismxwn/replay.mp4",
+  "duration_s": 19.92,
+  "fps": 25.0,
+  "frames": 498,
+  "sha256": "50b08f00bd06a3b461cf8ec05a506adfbe51db5579289e92417984f074af91dd"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 18.6,
+      "log_step": null,
+      "text_zh": "观察：视频全程按约1秒间隔检查，并放大检查视频17.8–19.88 s；视频约18.6–19.8 s可见远处小型机器人轮廓离开跳台、向台面末端移动并下降，视频19.88 s仍在末端蓝线的上游侧；分辨率不足以辨认两块雪板的完整转角或首次触雪部位。"
+     },
+     {
+      "t": 19.88,
+      "log_step": null,
+      "text_zh": "观察：本次run.json评分器字段为stages_reached=2/4、progress=0.5、success=0、replay_success=0、deterministic=1；metric_trick_spin_deg=-287.7、metric_trick_flips=0、metric_from_knoll_m=-4.89。"
+     },
+     {
+      "t": null,
+      "log_step": 598,
+      "text_zh": "推断：原生四阶段只到起滑和起跳，落点距台面末端边界仍差4.89米，故完整的“转体落地滑出”事件从未成立，区间为空且final=false；不把原生阶段2/4换算为语义节点完成。agent在log step 598对交付文件也自述落在边界之前4.89米并明确承认失败，与评分器相符。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频约18.1–19.88 s可辨认机器人沿跳台运动及短暂腾空，但机器人和固定雪板只有少量像素，不能可靠区分双靴、雪板与其他身体部位的触雪情况；无状态日志，实际接触保留未知。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "value": -287.7,
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target"
+  },
+  {
+   "name": "landing_heading_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better"
+  },
+  {
+   "name": "ride_turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": 19.8,
+   "log_step": 598,
+   "where": "操作中",
+   "evidence_zh": "推断：视频约19.4–19.88 s的放大帧显示机器人轮廓下降至蓝线之前的台面区域；评分器metric_from_knoll_m=-4.89、stages_reached=2/4确认未通过落地阶段。失败发生于整次转体跳跃的落地过程，尚未进入合格滑出；视频19.8 s仅作下降末段的导航锚点，并非精确触雪时刻。agent在log step 598确认交付的是这一落短的轨迹。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": 19.8,
+    "log_step": 598,
+    "evidence_zh": "推断：视频约18.6–19.88 s可见机器人从跳台腾空后下降，末帧的轮廓仍处在蓝线之前；结合本次评分器metric_from_knoll_m=-4.89、metric_air_s=1.393和fell=0，失败表现为跳跃距离不足，未越过台面末端完成要求的落地，而非已证实的摔倒。agent在log step 598也自述交付轨迹落短4.89米。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 541,
+    "evidence_zh": "推断：log step 541展示交付生成脚本：agent用walk模式助滑，在骨盆横坐标达到69米后切换joints模式，持续使用固定的髋、膝、踝和腰目标；log step 524将腰俯仰目标设为+0.2，log step 593再次运行该生成器，log step 595确认交付有1991行有限动作。评分器确认这组控制实现了-287.7度雪板转体、零计数空翻，却仍落短4.89米；据此归为起跳及飞行控制未实现所需距离。缺乏逐帧速度、受力和关节跟踪数据，不能进一步认定某一参数、反馈或逆解子类。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "没有状态日志。本次视频全程粗查，视频17.8–19.88 s按约0.1秒放大检查；成功参照全程粗查并放大检查参照视频11–17.5 s，均为远景且主要过程被近景雪坡遮挡，机器人只占少量像素。能判断末段腾空和下降的大致轨迹，但无法直接复核精确板转角、首次触雪部位、落地航向或具体接触，故contact_actual为未知；landing_heading_deg及ride_turn_deg未由评分器报告，保留null。断点时间为粗略视觉锚点，距离及转角来自本次评分器；控制根因只到大类，不推断具体力学失配。"
+ },
+ "summary_zh": "本次评分器记录起滑、起跳两阶段完成，雪板空中净转角−287.7度、计数空翻为零，但落点在台面末端之前4.89米，未完成“转体落地滑出”，语义节点最终为0/1。断点在跳跃落地过程中，现象为距离不足（步态），日志中的交付控制未实现所需飞行距离，根因记控制。远景只能辨认末段的小型机器人轮廓，实际接触与精确姿态无法核实，未报告的落地航向和滑出转向指标保留未知。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点final=false、断点为spin_360_ridden_away、现象步态（跳跃距离不足，无摔倒）、根因控制（只到大类）均与评分器字段（stages 2/4、metric_from_knoll_m=-4.89、fell=0）和远景视频一致；远景下机器人仅数像素，contact_actual记未知合理，未发现需要修改的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:44:45Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--uismxwn/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--uismxwn/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0528d42a9f76cb5fb44afe73dee6e7dc1ec2f19834d41501f3c2fb62d285b31a",
+   "video_sha256": "50b08f00bd06a3b461cf8ec05a506adfbe51db5579289e92417984f074af91dd",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-privileged--wz6sfpn.json

@@ -0,0 +1,214 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-privileged--wz6sfpn",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.75,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 1628,
+   "sim_steps": 1628,
+   "horizon": 4000,
+   "episode_over": 1,
+   "progress": 0.75,
+   "stages_reached": 3,
+   "stages_total": 4,
+   "fell": 1,
+   "replay_s": 17.1,
+   "bad_trajectory": 0,
+   "metric_air_s": 1.756,
+   "metric_from_knoll_m": 2.07,
+   "metric_height_m": 2.46,
+   "metric_speed_m_s": 19.969,
+   "metric_trick_flip_deg": -20.4,
+   "metric_trick_flips": 0.0,
+   "metric_trick_half_turns": 2.0,
+   "metric_trick_pelvis_spin_deg": -283.8,
+   "metric_trick_short_deg": 33.5,
+   "metric_trick_spin_deg": -326.5,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 479,
+   "agent_resets": 0,
+   "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": 2.957,
+   "peak_processes": 20,
+   "peak_disk_gb": 0.984,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-freeski-360-i00-privileged--wz6sfpn/replay.mp4",
+  "duration_s": 16.32,
+  "fps": 25.0,
+  "frames": 408,
+  "sha256": "dcc8c236ee1aea3b187ebdead5931566b41f71fc7b17c98c0ee7af1f31b505ac"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 190,
+      "text_zh": "观察：log step 190 的轨迹生成脚本保存交付文件后，基准重放工具对该文件执行1628行，报告 success=false、failed=fell、stages=[start,take-off,landing]、progress=0.75；这是基准裁判的重放结果，不是 agent 对动作的口头判断。log step 198 的两次新进程重放终态散列一致，log step 215 最终交付验证仍为同一散列和摔倒结果；inputs/run.json 的 reward 同样报告 deterministic=1、fell=1、stages_reached=3/4。"
+     },
+     {
+      "t": 16.24,
+      "log_step": null,
+      "text_zh": "观察：已检查视频全程逐秒远景、视频 12.4–14.2 s 及视频 13.8–16.0 s 每约0.2秒放大帧、视频 16.24 s 末段帧；机器人仅呈细小模糊轮廓，近侧雪坡遮挡动作细节，无法可靠判断雪板转角、首次触雪或摔倒姿态。成功参考同样使用远景，参考视频 17.52 s 也不能提供可可靠辨认的完成动作对照。"
+     },
+     {
+      "t": null,
+      "log_step": 215,
+      "text_zh": "推断：交付轨迹的基准裁判确认已起滑、起跳并接受落地，但在滑出阶段摔倒，完整的“转体落地滑出”从未成立；因此 event 节点没有达成区间，final=false。原生阶段进度0.75不换算为这个单节点的完成。视频不具备判定该完整动作的可用细节，故以交付文件重放结果判定，method=log；未把会话测试的仿真时刻换成视频时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频 12.4–16.24 s 放大后仍无法可靠分辨双靴、雪板与雪面的实际接触，也没有逐帧状态接触日志；不能仅依据任务的固定雪板配置或裁判阶段记录，把典型的腿部支撑接触当作本轨迹的可见接触。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "value": -326.5,
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target"
+  },
+  {
+   "name": "landing_heading_deg",
+   "value": [
+    6.4,
+    6.4
+   ],
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better"
+  },
+  {
+   "name": "ride_turn_deg",
+   "value": null,
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better"
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 215,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 215 对最终交付文件的重放报告 outcome=fell、stages=[start,take-off,landing]，且与 log step 190 的重放终态散列一致；评分器 reward 报告 fell=1、success=0。失败发生在机器人落地后继续滑出的连续动作中，该节点未完成；视频 15.4–16.24 s 远景无法辨认失衡开始的瞬间，故不指定视频断点时间。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": null,
+    "log_step": 215,
+    "evidence_zh": "观察：评分器 reward 的 fell=1，log step 190 和 log step 215 对交付文件的基准重放均报告 failed/outcome=fell，已接受 landing 而没有 ride-away；这支持落地滑出时摔倒的现象。视频 15.4–16.24 s 中机器人太小且受遮挡，无法独立辨认具体倒向或身体触雪部位，未从其他测试移用摔倒画面。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 153,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 153 用与 log step 184 最终生成配置相同的参数进行会话诊断，记录落地后航向来回偏转，仿真 15.91 s 至仿真 16.21 s 横滚从12.1度增至47.4度、骨盆法向净高从0.67米降至0.44米，随后裁判判 fell；最终交付重放在 log step 190、198、215 一致复现相同阶段和摔倒结果。结合这些诊断，支持落地后姿态控制未能稳定滑出的控制问题；会话诊断只用于根因，不作为本次视频节点的时间或接触证据，现有信息不足以细分反馈控制、轨迹或参数问题。"
+   }
+  ],
+  "gap_zh": "视频为640×360固定远景，已检查全程粗采样、起跳至落地附近及片尾最后一秒密集帧并放大；机器人细小且被坡面遮挡，不能可靠辨认实际支撑接触、雪板姿态或失衡起点，故 contact_actual=未知、breakpoint.t=null。inputs/没有逐帧状态日志；已提供的成功参考视频也存在相同可见性限制。现象依据最终交付文件的裁判重放与评分器 fell 字段，控制根因仅由相同配置的会话诊断作有限推断，具体失稳机制未定。landing_heading_deg=[6.4,6.4]来自 log step 190 交付文件重放的裁判 metrics；reward直接报告 trick_spin_deg=-326.5，ride_turn_deg在评分与交付重放结果中均未报告，填null。",
+  "unmet": {
+   "spin_360_ridden_away": false
+  }
+ },
+ "summary_zh": "基准裁判确认交付轨迹完成起滑、起跳和落地，雪板空中转角为−326.5度、空翻计数为零，但机器人在继续滑出时摔倒，唯一的“转体落地滑出”节点最终为0/1。相同配置的会话诊断支持落地后姿态控制失稳，交付文件多次确定性重放均得到摔倒结果。视频与成功参考均为细节不足的远景，实际接触和视频失衡时刻无法可靠辨认，已保留未知及空时间。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与证据一致：视频为远景，机器人不可见（已核对），节点 final=false、断点在唯一节点、现象为平衡（fell=1，基准重放 log step 190/198/215 一致）、根因为有限推断的控制，均有支持；未发现需要修改的错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:44:33Z",
+  "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/embodied-olympics-fun-freeski-360-i00-privileged--wz6sfpn/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-privileged--wz6sfpn/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8695c11ef987626d4b8b34898970e7a6939f3194db20219b4d96e514e2545c58",
+   "video_sha256": "dcc8c236ee1aea3b187ebdead5931566b41f71fc7b17c98c0ee7af1f31b505ac",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--cgsbh2j.json

@@ -0,0 +1,199 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-standard--cgsbh2j",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 1109,
+   "sim_steps": 8552,
+   "calls": 665,
+   "rows": 8552,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 2359.0,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 3,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 4.209,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.51,
+   "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/embodied-olympics-fun-freeski-360-i00-standard--cgsbh2j/replay.mp4",
+  "duration_s": 11.12,
+  "fps": 25.0,
+  "frames": 278,
+  "sha256": "3b589783a580a919341df43988ec49426b02621c677248d08c8327ff4229cd47"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 87,
+      "text_zh": "观察：log step 87 中 agent 用最后一次 reset 启动 episode 6，机器人服务返回的最终 reason、outcome、stages_reached、stages_total 被脚本原样打印为“END failed fell 1 4”；打印代码见 log step 80。log step 88 的服务状态确认 episode=6、done=true、control_steps=1109、rows=8552、calls=665、resets_left=0，动作数、总步数和调用数与本次评分器字段一致，后续未再执行机器人动作。这是评分末局的服务结果，完整转体落地滑出事件未成立，不能把原生阶段1/4算作语义节点完成。"
+     },
+     {
+      "t": 11.08,
+      "log_step": null,
+      "text_zh": "观察：本次视频从视频 0 s 起每0.5秒检查至末段，并密查视频 9.2–11.08 s；画面为近处雪坡遮挡的固定远景，末帧不能可靠辨认机器人或雪板。成功参考视频全程每秒抽帧及参考视频 12 s 原分辨率帧也有同样可见性缺口，因此空景本身不证明未完成；节点否定依据是本次末局服务的明确终止结果。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：评分末局确有动作执行且服务报告阶段1/4后摔倒（log step 87–88），但本次视频全程及视频 9.2–11.08 s 密集帧均无法可靠辨认机器人、双靴或雪板；没有状态接触日志，不能核定实际支承链路或摔倒时触雪部位。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 87,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 87 的最后一次 reset 明确进入 episode 6，机器人运行起滑脚本后，服务终态为 failed / fell、stages_reached=1、stages_total=4；log step 88 确认本局已终止、无剩余重置，且动作计数与评分器一致。推断：失败发生在整次转体落地滑出动作的起滑阶段，尚未完成跳台动作；本次视频末段仍为遮挡远景，无法从帧定位失衡时刻，故视频时间留空。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": null,
+    "log_step": 87,
+    "evidence_zh": "观察：评分末局的机器人服务裁判 outcome 为 fell（log step 87），随后 log step 88 查询同一 episode 6 显示 done=true；并非取自早先练习局的摔倒或仅引用 agent 的总结。log step 88 对该局已存传感读数的诊断显示，仿真 9.46 s 骨盆俯仰约14.5度，仿真 10.59 s 约-72.7度；这些读数支持明显后仰失衡，视频本身因遮挡未直接显示摔倒。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 87,
+    "evidence_zh": "推断：评分末局的姿态控制未维持起滑平衡。log step 80 的脚本在起滑序列后立即用30个控制步切换到 ride 姿态；log step 87 将双髋、双膝、双踝目标改为-0.8、1.2、-0.25弧度并运行该脚本。log step 88 对同一末局保存读数的诊断显示，切换时髋膝读数跟随目标改变，骨盆随后连续后仰，最终服务报告 fell（log step 87）。这支持姿态切换导致控制失稳的解释，但没有真实质心、雪面接触力或力矩记录，故不细分为参数调试或接触设计，也不把 log step 86 本地模型估计的质心偏移当作评分场景真值。",
+    "source_zh": "日志"
+   }
+  ],
+  "gap_zh": "本次视频和成功参考视频都是被近处雪坡遮挡的场地远景，无法可靠辨认机器人和雪板；已检查本次全程抽帧、视频 9.2–11.08 s 密集帧和视频 11.08 s 原分辨率末帧。inputs/ 有运行日志和两段视频，但没有 states.jsonl 或其他接触状态日志；实际接触记未知。末局的 episode、裁判结果和计数可与评分器对应，但没有可见物理事件可核对仿真到视频的精确映射，不将仿真时间转换成视频断点时间。转角、落地航向和滑出转向三个原生指标均未由本次评分器报告，值留空；控制根因属于脚本与同局诊断共同支持的推断。",
+  "unmet": {
+   "spin_360_ridden_away": false
+  }
+ },
+ "summary_zh": "评分末局为 episode 6：机器人服务仅报告完成1/4个原生阶段，随后以 fell 终止，完整“转体落地滑出”节点未达到，语义进度为0/1（log step 87–88）。脚本在起滑后立即切换为深屈膝后坐姿态，同局诊断随后显示骨盆持续后仰，支持控制失稳导致起滑阶段摔倒的解释。两段视频均无法可靠辨认机器人，实际接触和视频断点时间无法核定；未把 早先重置局的动作或本地自由飞行测试计入本次进度。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与证据一致：评分末局（episode 6）服务终态 failed/fell、阶段 1/4，节点未达为 false；视频为被雪坡遮挡的空景，看不到机器人，所以接触记未知、断点时间留空合理；平衡现象与控制根因（停在“控制”不细分）有 log step 87–88 的同局读数支持，且已标为推断。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:47:26Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--cgsbh2j/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--cgsbh2j/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "eb27f7b3b59cdb17fe7c90e1967eab680038dae52d094a78d895682899da0ef5",
+   "video_sha256": "3b589783a580a919341df43988ec49426b02621c677248d08c8327ff4229cd47",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--ftisjrc.json

@@ -0,0 +1,193 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-standard--ftisjrc",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 738,
+   "sim_steps": 3164,
+   "calls": 58,
+   "rows": 3158,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 488.8,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.313,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.834,
+   "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/embodied-olympics-fun-freeski-360-i00-standard--ftisjrc/replay.mp4",
+  "duration_s": 7.4,
+  "fps": 25.0,
+  "frames": 185,
+  "sha256": "313ae9e700169ce0fc7a15c5142f76dc12d07bf669d5f058718cf9b6c880e43a"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 7,
+      "text_zh": "观察：log step 7 的机器人服务规格明确规定只评分最后一个 episode；log step 59 的 agent 脚本发起最后一次 reset，log step 64 继续运行同一助滑脚本，随后未再重置。inputs/run.json 的评分器字段为 graded_by_live=1、finalized=1、live_success=0、infra_error=0，说明最后一次实时尝试已完成评分且未通过。"
+     },
+     {
+      "t": null,
+      "log_step": null,
+      "text_zh": "推断：本任务只有“转体落地滑出”这一个与整次成功绑定的 event 节点；评分器 live_success=0 且已 finalized，因此该完整动作未被验收，final=false、达成区间为空。这不证明某个助滑或转体子阶段完全没有发生。"
+     },
+     {
+      "t": 7.36,
+      "log_step": null,
+      "text_zh": "观察：按视频 0.00–7.20 s 每隔约0.40 s 检查全程，并逐帧检查视频 6.36–7.36 s；固定远景显示雪坡、跳台和落地区，近景雪坡遮挡部分赛道，无法可靠辨认本次机器人、雪板或摔倒动作。成功参考视频全程及参考视频 17.60 s 的原分辨率末帧也不足以提供本次动作的可核查接触证据；不能仅凭本次画面无人判定未尝试。"
+     },
+     {
+      "t": null,
+      "log_step": 66,
+      "text_zh": "观察：agent 在 log step 66 自述最后一局“起滑后、起跳前摔倒，阶段1/4”，并在 log step 68 重复该结果；log step 67 输出的 agent 自写记录也作同样描述。这些是 agent 自述，相关服务调用的实际裁判输出未保留在可读日志中，故不据此填入摔倒时刻、接触或独立的阶段达成时间。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：log step 59、64 记录了 agent 启动并延续助滑控制脚本，评分器记录最后尝试 n_actions=738；但视频 0.00–7.36 s 无法可靠辨认机器人的靴、雪板或躯干，且没有状态日志，无法确认实际支承或身体触雪，故接触标未知。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 64,
+   "where": "操作中",
+   "evidence_zh": "推断：log step 59 的 agent 脚本重置后执行助滑控制，log step 64 延续该尝试；评分器记录最后尝试 n_actions=738、live_success=0、finalized=1。助滑是唯一节点“转体落地滑出”的连续动作组成部分，故断点归该节点的操作中。视频 0.00–7.36 s 不能可靠辨认机器人，无法定位动作何时变坏，t留空；不把 agent 在 log step 66 自述的仿真 7.38 s 当作已核实的视频锚点。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [],
+  "root_causes": [],
+  "gap_zh": "本次视频已检查全程与末秒逐帧，但固定远景和近景雪坡遮挡使机器人及其雪板无法可靠辨认；成功参考视频也没有可供本次接触核验的清晰动作视角。inputs/ 中没有状态日志；log step 59–65 的相关服务工具输出只有会话编号或空文本，log step 67 没有实际裁判状态输出，保存脚本的中段也被日志截断。agent 在 log step 66、68 自述起跳前摔倒、阶段1/4，不能替代本次动作观测；评分器只报告未通过而未给出可解释的原生失败状态，first_reason_code=2在现有输入中无可核对的码义。因此能判定唯一目标未通过和尝试已启动，实际接触、摔倒现象、断点视频时刻及具体根因均缺乏独立证据，现象和根因留空。trick_spin_deg、landing_heading_deg、ride_turn_deg均未由本次评分器字段报告。"
+ },
+ "summary_zh": "最后一次实时尝试已终结且评分未通过，唯一的“转体落地滑出”节点未获验收，终态为0/1。agent 在 log step 66、68 自述起滑后、起跳前摔倒，阶段1/4，但本次视频全程及末秒逐帧检查均无法可靠辨认机器人，无法独立核实该描述。断点归连续滑雪动作的操作中，接触标未知，视频时刻、失败现象和根因因缺少清晰动作画面与原生状态记录而留空。",
+ "conf": 5,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "标注与证据一致：视频固定远景中看不到机器人（本次与参考视频均如此），评分器 live_success=0、finalized=1，故唯一节点 final=false；接触未知、现象与根因留空且有 gap_zh，处理保守但合理。日志中没有保留服务裁判输出，摔倒/阶段1/4 只来自 agent 自述，未被当作事实。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:45:10Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--ftisjrc/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--ftisjrc/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "e61b4dc3e68127fc66467c82c33569cc2110df556d1f5026799bd45a5ee21f5e",
+   "video_sha256": "313ae9e700169ce0fc7a15c5142f76dc12d07bf669d5f058718cf9b6c880e43a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--moc8bqh.json

@@ -0,0 +1,198 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-standard--moc8bqh",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 1410,
+   "sim_steps": 7309,
+   "calls": 636,
+   "rows": 2006,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 3167.9,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.306,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.853,
+   "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/embodied-olympics-fun-freeski-360-i00-standard--moc8bqh/replay.mp4",
+  "duration_s": 14.12,
+  "fps": 25.0,
+  "frames": 353,
+  "sha256": "827424f7e3fcca06c14559a6792da962c941b9b53511f909277296f8f6514e87"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 134,
+      "text_zh": "观察：log step 134 中机器人实时服务报告最后评分局 episode=6、control_steps=1410、sim_time_s=14.1、done=true、outcome=\"fell\"、stages_reached=1、stages_total=4；log step 132 发起最后一次 reset，log step 133 的该次执行返回同一摔倒结果。此为服务裁判结果，不是前一次试验或 agent 自述；完整转体落地滑出未被接受，节点为 false。"
+     },
+     {
+      "t": 14.08,
+      "log_step": null,
+      "text_zh": "观察：本次视频 0–13 s 的逐秒帧、视频 12.4–14.0 s 的约0.2秒间隔帧及视频 14.08 s 的末帧均为固定场地远景，近景雪坡遮挡助滑区域，无法可靠辨认机器人、雪板或触雪部位；成功参考视频 0–17 s 的逐秒帧和参考视频 17.56 s 的末帧也有同样可见性限制，不能根据空景判断动作完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：本次视频全程远景及视频 12.4–14.08 s 的末段检查无法辨认机器人与雪板接触；无状态日志。最后评分局确有执行并被服务判为摔倒（log step 133–134），但实际承重及身体触雪的作用部位不能从画面核实，故接触记未知。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 134,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 131–132 的 agent 脚本为最后评分局执行此前记录的助滑关节序列；log step 133 返回 replay 14.1 True / fell，log step 134 的服务最终结果为 episode=6、1/4阶段、outcome=\"fell\"。该连续动作在助滑执行中失败，未完成转体落地滑出。视频 12.4–14.08 s 看不清机器人，无法给出视觉核实的失衡瞬间；仿真 14.1 s 不直接当作视频断点时间。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": null,
+    "log_step": 134,
+    "evidence_zh": "观察：log step 134 的实时服务裁判对最后评分局 episode=6 明确返回 outcome=\"fell\"、done=true、stages_reached=1/4；log step 133 的执行输出也为 fell，说明机器人在该局发生摔倒。视频 12.4–14.08 s 的远景无法核实具体倒向或触雪部位；不把前一局的空中倾斜与落地方向错误当成本局现象。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志",
+    "evidence_zh": "推断：最终助滑执行的控制未维持平衡。log step 64、80 的原助滑脚本每次读取新IMU及关节反馈并更新踝、髋目标；log step 131–132 却把其已���录的目标插值成固定序列，分两批提交，不根据本局传感反馈重新计算助滑目标。log step 134 显示在仿真 9.2 s 的对比检查点关节和IMU已存在差异，最终服务在仿真 14.1 s 判 fell；log step 135 的 agent 记录也承认最终改成了开环回放。证据支持控制执行失稳，但没有末段姿态或接触记录，无法单独确定反馈缺失、插值改变或某个参数各自的贡献，故停在控制大项。"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 提供本次及成功参考视频和完整运行日志，但无逐帧状态日志。本次视频全程逐秒检查及视频 12.4–14.08 s 加密检查均无法可靠辨认机器人；成功参考同一视角也无法解除遮挡，因此实际接触、摔倒方向和视频发生时刻均不能独立核实。最后评分局由 log step 132–134 的 reset、执行结果和服务 episode=6 明确定位；节点未完成和摔倒取自该局服务裁判，而非 agent 的前五次试验。未找到可见事件核对控制步到视频的映射，故断点 t 留空；三个原生指标未在本局评分字段报告，均为 null。控制细分原因缺少末段传感与接触证据。"
+ },
+ "summary_zh": "最后评分局为第6次实时尝试，服务在仿真 14.1 s 报告摔倒、仅达到1/4个原生阶段，完整“转体落地滑出”节点未完成，语义进度为0/1。断点在该连续动作的助滑执行中；日志支持控制执行失稳，最终脚本将此前依赖传感反馈生成的目标改为固定开环序列。两段视频均受远景及雪坡遮挡，不能核实机器人实际接触和摔倒的视频时刻，接触记未知、断点时间留空。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "标注与证据一致：视频（时长14.12 s，对应最后评分局）全程远景且雪坡遮挡，看不到机器人，接触记未知合理；服务在 log step 134 报 episode=6、outcome=fell、1/4 阶段，节点 false、断点在助滑、平衡现象、控制根因（标为推断）均有日志支撑，t 留空保守可接受。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:46:37Z",
+  "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/embodied-olympics-fun-freeski-360-i00-standard--moc8bqh/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--moc8bqh/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "929a8341cfc0b40a8349baccfb9099882c4063d6a69886b91f8d938150cfdf4e",
+   "video_sha256": "827424f7e3fcca06c14559a6792da962c941b9b53511f909277296f8f6514e87",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--peodoez.json

@@ -0,0 +1,203 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-standard--peodoez",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 1116,
+   "sim_steps": 3578,
+   "calls": 189,
+   "rows": 3578,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 5264.8,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.934,
+   "peak_processes": 13,
+   "peak_disk_gb": 2.548,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-freeski-360-i00-standard--peodoez/replay.mp4",
+  "duration_s": 11.2,
+  "fps": 25.0,
+  "frames": 280,
+  "sha256": "c6a26a926b538666e5b4e67e932ab048af0c26d3187434488237fb02b0f6c449"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 128,
+      "text_zh": "观察：log step 125 重置并启动最后一次实时尝试；log step 128 的服务返回标明 episode=6、control_steps=1116、sim_time_s=11.16、done=true、reason=failed，其累计 control_steps_total=3578 与本次 reward.sim_steps 一致。此前五局为会话试验，不合并计入本视频。"
+     },
+     {
+      "t": null,
+      "log_step": 129,
+      "text_zh": "观察：log step 129 读取最后一局 att_6 的服务裁判 meta，得到 outcome=fell、stages_reached=1、stages_total=4；log step 112 说明这些 meta 字段是服务裁判对已结束尝试的结果。推断：该局只完成起滑阶段，尚未完成起跳、合格转体落地与滑出这一完整事件，节点未达到。"
+     },
+     {
+      "t": 11.12,
+      "log_step": null,
+      "text_zh": "观察：已按视频每约0.5秒检查全程，并加密检查视频10.0–11.1 s及视频11.12 s原分辨率帧；画面为固定场地远景，近景雪坡遮住路线，不能可靠辨认机器人或雪板。成功参考视频每约1秒的全程画面及参考视频12.0 s帧也有相同可见性缺口；因此不以空景作为未起滑或未完成的证据。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：本次视频0.0–11.12 s所检帧不能可靠辨认机器人、雪板和触雪点，且无逐帧状态日志，无法确认实际作用端与接触方式；未复制任务DAG中的典型支撑方案。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 129,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 128 显示最后一次实时尝试由控制脚本持续发出滑行关节指令，随后服务以 failed 结束；log step 129 从该局服务裁判 meta 读取 fell、1/4 阶段。推断：机器人在同一次转体滑出动作的助滑至起跳准备过程中失衡，完整动作未完成；本视频不能看清失败瞬间，故不填视频断点时刻。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": null,
+    "log_step": 129,
+    "evidence_zh": "观察：log step 129 读取评分最后一局的服务裁判结果 outcome=fell、stages_reached=1/4；这是真实服务返回的摔倒判定，而非仅引用agent最终总结。视频10.0–11.12 s所检帧被雪坡遮挡，无法视觉定位身体失去支撑或触雪的瞬间。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "反馈控制",
+    "source_zh": "日志",
+    "evidence_zh": "观察：log step 125 的最终控制脚本保留姿态修正环，Kpb=1.2、trim_lim=0.25；log step 129 对最后一局保存的服务本体传感读数计算出仿真8.84–11.00 s骨盆俯仰约1.06–1.23 rad、左踝约−0.88至−0.91 rad，期间俯仰未恢复，随后该局裁判报告fell。推断：滑行过程中持续前倾未被姿态反馈纠正，支持姿态稳定控制失败；不能据此独立证明某个增益或起步参数必然不当。"
+   }
+  ],
+  "gap_zh": "本次视频全程所检帧（视频0.0–11.12 s）及成功参考视频均无法可靠辨认机器人，输入中无状态日志，实际支撑链、起跳地点和摔倒瞬间不可核验。log step 128 的服务控制步数1116与0.01秒控制周期对应仿真11.16 s，接近视频11.2 s长度，但没有可见动作事件可核对映射，故未把日志仿真时刻转成视频断点时间。反馈控制归因为末段本体传感数据与裁判结果支持的推断；agent在log step 134自述起步踝部塌陷、修正权限不足，其精确因果链未获独立验证。评分器未报告DAG的三个原生指标，均留null。"
+ },
+ "summary_zh": "最后一次实时评分尝试完成起滑，但服务裁判报告fell、仅达到1/4原生阶段，因此完整“转体落地滑出”节点未完成，最终进度0/1。末段传感记录显示持续大幅前倾未恢复，支持姿态反馈控制失败；本次及成功参考视频均被近景雪坡遮挡，实际接触、摔倒视频时刻及更具体的控制参数因果关系无法独立核验。",
+ "conf": 6,
+ "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、contact 未知、断点 spin_360_ridden_away（操作中）、平衡现象与反馈控制根因均有服务读数（outcome=fell、stages 1/4、本体传感）支持；视频被雪坡遮挡，标注如实，未把仿真时间转成视频时间。",
+    "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": "2026-10-31T00:49:53Z",
+  "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/embodied-olympics-fun-freeski-360-i00-standard--peodoez/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--peodoez/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "c1ecf32cbbbcc94070840a8aee4f565672c67ea204c59c8cc1793e70fa3c0be0",
+   "video_sha256": "c6a26a926b538666e5b4e67e932ab048af0c26d3187434488237fb02b0f6c449",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--pqakdp5.json

@@ -0,0 +1,196 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-standard--pqakdp5",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 894,
+   "sim_steps": 5822,
+   "calls": 249,
+   "rows": 5822,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 1819.2,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 0.207,
+   "peak_processes": 11,
+   "peak_disk_gb": 2.531,
+   "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/embodied-olympics-fun-freeski-360-i00-standard--pqakdp5/replay.mp4",
+  "duration_s": 8.96,
+  "fps": 25.0,
+  "frames": 224,
+  "sha256": "a82e74cc88ea5e0679b836d4f88d7c2e00652a4c18ba0b64de330b0fc542c6b9"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 90,
+      "text_zh": "观察：log step 90 中 agent 的脚本执行最后一次 reset（resets_left=0）并启动 att6；该次机器人服务返回 DONE: failed | outcome: fell | stages: 1 / 4。log step 53 的驱动代码表明这些字段直接读取服务响应 res.meta，而非 agent 对画面的描述；因此完整转体落地滑出事件未达成。"
+     },
+     {
+      "t": null,
+      "log_step": 117,
+      "text_zh": "观察：log step 117 的机器人实时服务 /status 返回 episode=6、done=true、reason=failed、control_steps=894；log step 114 的后续 att7 提交已被拒绝，未产生新的评分尝试。"
+     },
+     {
+      "t": 8.92,
+      "log_step": null,
+      "text_zh": "观察：本次视频 0.00–8.50 s 的每隔约0.5秒采样、视频 7.96–8.92 s 的逐帧检查以及视频 8.92 s 原尺寸末帧均为雪坡遮挡的远景，无法可靠辨认机器人和雪板；成功参考视频 0.00–17.00 s 的逐秒采样及参考视频 12.00 s 原尺寸帧也如此，不能用空景判断动作发生与否。本节点最终值依据本次服务的阶段与终止结果。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "本次视频无法辨认机器人、雪板及触雪部位，且没有状态接触日志。log step 89–90 表明脚本已尝试以 walk 转 ride 执行起滑，但无法由此确认实际接触方式，不照抄任务的典型支撑方案。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 90,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 89–90 的脚本以 walk 起滑并切换 ride，随后本次实时服务返回 outcome=fell、stages_reached=1、stages_total=4；log step 117 确认第6局已经失败终止。推断：失败发生在完整转体落地滑出动作的起滑/助滑尝试中；视频无法定位失衡开始，故不把日志中的估计时间换成视频断点。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": null,
+    "log_step": 90,
+    "evidence_zh": "观察：log step 90 的 att6 执行输出中，机器人服务明确返回 outcome=fell，且仅达到1/4原生裁判阶段；log step 53 显示驱动直接打印 res.meta.outcome。该摔倒结论来自本次评分局的服务反馈，视频远景不能独立复核倒地姿态或触雪部位。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "本次视频全程及最后一秒逐帧均无法可靠辨认机器人；成功参考同样被近景雪坡遮挡。没有状态日志，实际接触、失衡开始的视频时刻及姿态变化无法核实；保留服���明确报告的摔倒现象。log step 90 的 pitch、slope、速度和雪板方向为 agent 脚本的估计量，log step 119–120 的“原地蹭步、楔形自锁”等为 agent 自述，缺少独立的姿态、接触力及控制跟踪证据，无法可靠区分具体控制问题或其他根因，根因留空。DAG三项原生指标均未在本次评分字段中报告。"
+ },
+ "summary_zh": "第6次实时评分尝试仅达到1/4原生裁判阶段，完整“转体落地滑出”节点未完成，最终为0/1。机器人服务明确返回摔倒，断点属于该连续动作的操作中；后续 att7 被拒绝，没有替代此结果。本次与成功参考视频均无法可靠辨认机器人，实际接触、视频失衡时刻和根因无法独立核实。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标注与证据一致：视频8.96 s对应control_steps=894（约100 Hz），末段仅见雪坡远景无法辨认机器人；log id 90的服务返回 fell、stages 1/4，节点未达成、平衡现象、根因留空并写明gap_zh，均有依据，未发现需要修改的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:46:28Z",
+  "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/embodied-olympics-fun-freeski-360-i00-standard--pqakdp5/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--pqakdp5/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "76d4a3b247b107085f331e95473923f97b46de7b04595fa455e3bd423b67311b",
+   "video_sha256": "a82e74cc88ea5e0679b836d4f88d7c2e00652a4c18ba0b64de330b0fc542c6b9",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--srekgg4.json

@@ -0,0 +1,205 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-standard--srekgg4",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 574,
+   "sim_steps": 5552,
+   "calls": 224,
+   "rows": 2608,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 1225.0,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "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/embodied-olympics-fun-freeski-360-i00-standard--srekgg4/replay.mp4",
+  "duration_s": 5.76,
+  "fps": 25.0,
+  "frames": 144,
+  "sha256": "9820720124ec304d89d87ba7c0dade76799f1382aa27fab710e4f037429492f8"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 87,
+      "text_zh": "观察：log step 85 的控制脚本重置后执行本次人字形蹬伸起滑；log step 87 打印实时机器人服务的裁判返回值（服务内置裁判，而非 agent 的完成自评）：episode=6、control_steps=574、done=true、reason=failed、outcome=stopped、stages_reached=1。log step 89 再次确认第6局已结束且无法重置；因此评分对象是该局，不是此前五局试验或之后改写的 gen.py。"
+     },
+     {
+      "t": null,
+      "log_step": 87,
+      "text_zh": "推断：本节点要求起滑、起跳、转体落地及滑出完整完成；第6局裁判仅接受起滑阶段并以 stopped 终止，故该整段事件未达成，final=false、无达成区间；这依据具体阶段与结束原因，不是仅由 success=0 反推。"
+     },
+     {
+      "t": 5.72,
+      "log_step": null,
+      "text_zh": "观察：已查看本次视频全程约每0.25秒的帧、视频1.4—3.3 s的密集帧及视频4.72—5.72 s的逐帧画面，均为场地远景，近景雪坡遮挡部分赛道，无法可靠辨认机器人和雪板；视频5.72 s亦不能检查转体、落地或过线。成功参考视频全程抽帧也存在相同可见性缺口，不能用空景断定没有动作。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频0—5.72 s无法可靠辨认机器人、雪板及触雪点，且无状态接触日志；虽第6局执行了起滑控制并被裁判接受起滑，仍无法核实实际作用端及接触方式，故不照抄DAG的典型支撑方案。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 87,
+   "where": "操作中",
+   "evidence_zh": "观察：log step 85 的脚本已实际执行起滑蹬伸及后续固定姿态控制；log step 87 的第6局服务裁判报告 stages_reached=1、outcome=stopped、done=true，未进入完成整次转体滑出的结果。视频全程不能辨认机器人，不能定位动作出错的视频时刻，故 t=null；断点为正在执行且未修复的“转体落地滑出”节点。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "步态",
+    "t": null,
+    "log_step": 87,
+    "evidence_zh": "推断：log step 87 的实时服务裁判对第6局返回 outcome=stopped、stages_reached=1；按本任务滑行停滞判据，这是起滑后速度连续低于0.3米/秒达到3秒，机器人未能保持向跳台的滑行。因此记全身·步态（滑行停滞）；依据是该局裁判的物理规则结果，而非早先局的摔倒或 agent 最终总结。画面不能核实支撑丢失，故不另标��衡。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": null,
+    "log_step": 85,
+    "source_zh": "日志",
+    "evidence_zh": "推断：log step 85 的已执行脚本在蹬伸时设置双踝 lean=0.18，检测前俯后切换到固定 Q0，后续循环持续保持 Q0而未按姿态误差调整。log step 87 中 agent 对该局服务本体感受数据的计算输出显示，仿真2.40 s俯仰约−57度、仿真2.85 s约−73度，后续仍大幅前俯，最后服务裁判返回 stopped；这些代码及反馈支持起滑到滑行的控制未能维持有效运动。仅归总体控制：无法从远景核实接触布局，也没有验证删除 lean 参数就能修复，故不细分参数调试或抓取和接触设计。"
+   }
+  ],
+  "gap_zh": "本次及成功参考视频均缺乏可辨认机器人和雪板的视角；无 inputs/states.jsonl 或逐帧状态路径，无法核实支撑、身体触雪、精确停滞起点及具体接触，contact_actual记未知。第6局具体裁判结果可判定整段事件未完成及滑行停滞，但姿态变化仅来自agent对本体感受数据的计算，控制根因为推断。未找到能与控制步计数核对的可见事件，故未把仿真5.74 s或任何agent时钟映射为视频断点，breakpoint.t为空。评分器未报告三个原生指标的数值，均记null。"
+ },
+ "summary_zh": "评分对应最后一次第6局：服务裁判仅接受起滑（原生阶段1/4），随后以 stopped 结束，完整“转体落地滑出”节点未达成，最终0/1。起滑后的滑行停滞记全身·步态；执行脚本切换为固定姿态后仍持续大幅前俯的日志支持总体控制问题。视频及成功参考均无法可靠辨认机器人，实际接触和出错的视频时刻无法核实，控制细分原因亦未获验证。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "节点未达成、断点、t=null、contact_actual 未知与根因停在「控制」均有 log step 87 的服务裁判读数（stopped、stages_reached=1）支持；视频全程为无机器人的远景，与标签所述一致，未发现需改动的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:46:44Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--srekgg4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--srekgg4/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "107bd788ad69a592baf35f4eb4f8b9831061cdd123d7a8974f2d64f92238e59d",
+   "video_sha256": "9820720124ec304d89d87ba7c0dade76799f1382aa27fab710e4f037429492f8",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/freeski_360/trajectories/embodied-olympics-fun-freeski-360-i00-standard--tachtmw.json

@@ -0,0 +1,193 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-freeski-360-i00-standard--tachtmw",
+ "taskKey": "embodied-olympics-fun/freeski_360",
+ "dag_sha256": "861efa8ce64a6c121798b9927345b65cbb761cb3b91579bd06a292aa528db572",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 1041,
+   "sim_steps": 4961,
+   "calls": 107,
+   "rows": 4951,
+   "resets": 5,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 5834.9,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 2,
+   "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/embodied-olympics-fun-freeski-360-i00-standard--tachtmw/replay.mp4",
+  "duration_s": 10.44,
+  "fps": 25.0,
+  "frames": 261,
+  "sha256": "c6790c7681a10a4e43f4ac8a87aa1491dd2f87ee6562f2638eccf8249716dc49"
+ },
+ "state": null,
+ "progress": {
+  "method": "log",
+  "denominator": 1,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "spin_360_ridden_away": {
+    "method": "log",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": null,
+      "log_step": 178,
+      "text_zh": "观察：log step 178 的机器人实时服务返回重置至 episode 6、resets_left=0，随后返回六轮双腿关节指令后的骨盆 IMU 读数；log steps 179–182 继续提交起滑和助滑指令。该回合为最后一次实时尝试，先前回合和本地 MuJoCo 模型测试不计入本节点。"
+     },
+     {
+      "t": 5.88,
+      "log_step": null,
+      "text_zh": "观察：本次视频按约0.5秒间隔检查全程，并查看视频 5.88 s 原分辨率帧及视频 9.20–10.40 s 每帧；画面为固定场地远景，近景雪坡遮挡运动区域，无法可靠辨认机器人、雪板及其接触。成功参考视频全程粗查及参考视频 12.00 s 原分辨率帧同样无法辨认动作，故不能凭本次空景判定起跳、转体或摔倒。"
+     },
+     {
+      "t": null,
+      "log_step": 186,
+      "text_zh": "观察：log step 186 的机器人实时服务状态为 episode=6、control_steps=1041、done=True、reason=failed；run.json 的评分字段为 graded_by_live=1、live_success=0、finalized=1、infra_error=0。"
+     },
+     {
+      "t": null,
+      "log_step": 186,
+      "text_zh": "推断：本 DAG 唯一事件节点绑定整次有效转体、落地并滑出终点的成功验收；最后实时回合已由服务终结为 failed，且评分确认 live_success=0，故完整节点从未锁存达成，intervals=[]、final=false。这不能判定各中间动作是否发生，也不表示没有尝试。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "未知",
+      "how": "未知",
+      "note_zh": "观察：视频全程远景无法可靠辨认双靴、雪板或身体与雪面的接触，且未提供状态接触日志；log steps 178–183 表明已实施本节点的关节控制尝试，但不能据指令将典型的腿部经雪板支撑接触当作已观察事实。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "trick_spin_deg",
+   "unit": "deg",
+   "threshold": 360,
+   "direction": "target",
+   "value": null
+  },
+  {
+   "name": "landing_heading_deg",
+   "unit": "deg",
+   "threshold": 45,
+   "direction": "lower_better",
+   "value": null
+  },
+  {
+   "name": "ride_turn_deg",
+   "unit": "deg",
+   "threshold": 90,
+   "direction": "lower_better",
+   "value": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "spin_360_ridden_away",
+   "t": null,
+   "log_step": 186,
+   "where": "操作中",
+   "evidence_zh": "推断：log steps 178–182 显示最后的 episode 6 已执行双腿起滑与助滑控制，log step 183 启动继续助滑的闭环脚本，log step 186 的实时服务随后报告该回合 done=True、reason=failed，因此失败归于正在尝试的完整转体落地滑出节点。视频全程及视频 9.20–10.40 s 的逐帧检查均不能可靠辨认机器人，无法定位失败发生的画面时刻，t 保留 null；不把视频结束点当作失败时刻。"
+  },
+  "unmet": {
+   "spin_360_ridden_away": false
+  },
+  "phenomena": [],
+  "root_causes": [],
+  "gap_zh": "观察：inputs/ 提供本次视频、run.json、video.json 和成功参考视频，但没有逐帧状态日志。视频全程及末尾密集检查均无法可靠辨认机器人和雪板，成功参考视频同样存在遮挡；最终 episode 6 的 log step 186 只返回 failed，未记录具体裁判 outcome、stages_reached、接触对或三项原生指标，log step 183 的后台脚本后续逐段观测输出也未见于提供的日志。推断：这些证据足以确认完整事件节点未达成及其已被尝试，却不能区分失衡、滑行偏航、起跳或转体不合格等具体失败现象，更不能确定根因；两层留空，接触标未知。log step 70 的 outcome=fell、stages_reached=1 属于重置前的 episode 1，不能套用到评分回合；log step 188 在评分回合结束后写出的含转体阶段的 gen.py 未执行，也不能当作视频动作。"
+ },
+ "summary_zh": "最后一次实时评分尝试为 episode 6，已执行起滑及助滑关节控制，但服务在 log step 186 报告 failed，完整“转体落地滑出”节点未达成，最终为0/1。断点归于该节点操作中；本次及成功参考视频均为无法可靠辨认机器人的遮挡远景，且最终回合缺少具体裁判判负原因和后段观测，故失败时刻、实际接触、失败现象及根因无法确定。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "标签保守且与证据一致：视频（约10.44 s，对应最后回合 control_steps=1041×0.01 s）全程为遮挡远景，无法辨认机器人；log step 186 服务报告 episode 6 done=True、reason=failed，run.json 评分 live_success=0，故节点 final=false、断点在该节点、t 为 null、现象与根因留空并有 gap_zh 均成立。log 中 episode 6 最后一段闭环脚本（step 183）输出缺失，无法区分具体失败现象。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:43:44Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--tachtmw/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-freeski-360-i00-standard--tachtmw/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "06da8c7d1d49009bc42784e858b2a601ee6e54de7ff60cc0a63329379d7091e4",
+   "video_sha256": "c6790c7681a10a4e43f4ac8a87aa1491dd2f87ee6562f2638eccf8249716dc49",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--6gqj8cq.json

@@ -0,0 +1,338 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-privileged--6gqj8cq",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 331,
+   "sim_steps": 331,
+   "horizon": 3000,
+   "episode_over": 1,
+   "progress": 0.0,
+   "stages_reached": 0,
+   "stages_total": 10,
+   "fell": 0,
+   "replay_s": 2.79,
+   "bad_trajectory": 0,
+   "metric_ball_height_m": 0.095,
+   "metric_touches": 0.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 71,
+   "agent_resets": 2,
+   "agent_replays": 3,
+   "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": 0.512,
+   "peak_processes": 11,
+   "peak_disk_gb": 0.239,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-juggling-i00-privileged--6gqj8cq/replay.mp4",
+  "duration_s": 3.32,
+  "fps": 25.0,
+  "frames": 83,
+  "sha256": "cdc4d65381318ba46c1aa2c6a907cd8680f5b1b3510be8c3d2bd2c68efc232ff"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.4,
+      "log_step": null,
+      "text_zh": "观察：视频 0.0–2.92 s，机器人先在双掌之间持球，随后把重心移到支撑腿并抬起右腿，球仍在手前，尚无离体击起后的回升。"
+     },
+     {
+      "t": 3.12,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 2.92–3.28 s，机器人张开双掌，球持续下降，从抬起的腿旁经过并落到草地；未见合法部位把球击起后回升一个球直径。inputs/run.json 的评分器字段 metric_touches=0、stages_reached=0，与画面一致，第1次有效颠球未成立。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.0–2.92 s，两臂末端平掌共同夹持并移动足球，之后分开释放。候选：双掌挤压夹持；单掌没有合围或夹爪闭合，故不标抓握。此为首次颠球尝试的持球、释放准备接触。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.0–1.8 s，双脚接触草地支撑身体；视频约 2.0–3.28 s，左脚承重，右腿抬起尝试迎球。未观察到右腿实际击起足球，不把尝试击打写成已发生的击打接触。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第2次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第3次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第4次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第5次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第6次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第7次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第8次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第9次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 3.28,
+      "log_step": null,
+      "text_zh": "推断：视频 0.0–3.28 s 仅出现持球、抬腿、放球和持续下落，球未出现第1次有效颠球所需的离体回升；评分器 metric_touches=0、stages_reached=0 进一步确认计数为零。因此依次锁存的第10次颠球未成立，其前置计数也未达到，评分尝试没有进入这一轮击球。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "value": 0.0,
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better"
+  },
+  {
+   "name": "ball_height_m",
+   "value": 0.095,
+   "unit": "m",
+   "threshold": null,
+   "direction": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "valid_touch_1",
+   "t": 3.12,
+   "log_step": null,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 2.0 s 起机器人抬右腿准备首次颠球；视频 3.04–3.20 s，已释放的球从右腿旁持续下降，右腿的迎击没有使球回升，视频 3.28 s 球到达草地。评分器记录 0 次有效颠球且 episode_over=1；log step 86 对当时331行交付文件的官方回放也返回 the ball touched the ground。首次颠球一直未修复，故断点为 valid_touch_1。"
+  },
+  "unmet": {
+   "valid_touch_1": false,
+   "valid_touch_2": false,
+   "valid_touch_3": false,
+   "valid_touch_4": false,
+   "valid_touch_5": false,
+   "valid_touch_6": false,
+   "valid_touch_7": false,
+   "valid_touch_8": false,
+   "valid_touch_9": false,
+   "valid_touch_10": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "位姿不准",
+    "t": 3.12,
+    "log_step": null,
+    "evidence_zh": "观察：视频 3.00–3.28 s，球离开双掌后从抬起的右腿旁持续下降并落到草地，未见被腿击起后回升；机器人保持左脚支撑。推断：放球路线横向偏离右大腿迎击面；inputs/run.json 的 log step 84 会话测试在控制步295记录球 y=+0.095 m、右大腿参考点 y=-0.03 m，支持空间错位，画面没有提供接触过早或过晚的独立证据。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "已验证有",
+    "t": 3.12,
+    "log_step": 84,
+    "evidence_zh": "推断：视频 3.04–3.28 s 的落球路线没有与右大腿迎击面接合，符合持球到腿部击球之间的接触几何未对齐。inputs/run.json 的 log step 84 测试采用随后写入 gen.py 并在 log step 86 回放验证的参数：控制步295的球 y=+0.095 m、横向速度 vy=+0.27 m/s，右大腿参考点 y=-0.03 m；控制步310的球 y=+0.135 m，而右大腿参考点 y=-0.04 m、右膝参考点 y=-0.06 m。log step 88 的 NOTES 是 agent 对该版本横向漏球的自述，与这些测试读数一致，仅作补充。log step 89 改大腰部调平参数后的另一场会话测试，控制步310的球 y=+0.133 m、右大腿参考点 y=-0.05 m、右膝参考点 y=-0.06 m，也呈横向分离；这组读数与 log step 84 分开引用，两者均不当作本次视频的逐帧状态。成功参考 inputs/oracle/replay.mp4 的参考视频 8.24–8.64 s 可见头球后离体上升，oracle.json 标明该参考成功，验证了另一种合法接触方案存在。",
+    "source_zh": "日志",
+    "note_zh": "已验证的替代方案是成功参考中的头球路线；本次右大腿方案的可行修正没有经过成功验证。"
+   }
+  ],
+  "gap_zh": "本次没有状态日志或逐物理步接触记录；单一远景可确认持续下落和无有效回升，但不能排除未计数的极短擦碰，也不能精确测量持球滑移及接触力。横向错位根因结合视频与会话诊断推断。inputs/run.json 的 log step 86 对当时331行文件回放并记录了末状态哈希；log step 89 改参数后运行 fftest.py，没有再调用官方回放。该脚本已在 log step 79 加入 s.save_trajectory()，log step 81 的后续编辑未删除该调用，且 log step 89 的输出筛选会隐藏 saved 提示，因此不能称其为不保存的纯测试。log step 90 仅有分钟级时间戳和相同的331行形状，不能据此证明文件未被改写；最后文件与 log step 86 哈希的对应关系仍未直接核实。log step 84 和89的读数分别作为不同参数的会话证据，不与视频逐帧绑定。"
+ },
+ "summary_zh": "机器人完成了持球、向左支撑腿转移重心、抬右腿和放球准备，但视频 3.04–3.28 s 中球从腿旁持续下落并触地，十次颠球节点均未达到，原生有效计数为 0。断点是首次颠球，现象为位姿不准，视频与会话诊断支持放球路线和右大腿接触几何横向错位的控制根因。缺少本次状态及接触日志，横向滑移的细节属于推断，最后文件与 log step 86 回放哈希的对应关系仍未直接核实；会话早期的摔倒不属于本次评分视频。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 7,
+    "summary_zh": "进度全假、断点 valid_touch_1 与视频一致（球约 2.96 s 离手，3.28 s 落地，无回升，metric_touches=0）。但现象“动态时机”与自己的根因证据（球横向漂移、释放时落点偏离大腿）自相矛盾：日志指出的是横向落点错位，不是时机早晚，应为位姿不准。gap_zh 对 log step 89 的描述也不准确。",
+    "n_issues": 2,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "十个节点全为假，断点 valid_touch_1 与视频及评分器（metric_touches=0）一致；位姿不准与根因证据自洽，gap_zh 如实标明版本对应缺口。未发现需改动之处。",
+    "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": "2026-10-31T00:50:50Z",
+  "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/embodied-olympics-fun-juggling-i00-privileged--6gqj8cq/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--6gqj8cq/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--6gqj8cq/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--6gqj8cq/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "1d179fbbced8e76d70149f1228ec6aec5d8e4baf3c48c31261a245d7802118fe",
+   "video_sha256": "cdc4d65381318ba46c1aa2c6a907cd8680f5b1b3510be8c3d2bd2c68efc232ff",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--ae4cve4.json

@@ -0,0 +1,477 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-privileged--ae4cve4",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 905,
+   "sim_steps": 905,
+   "horizon": 3000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 10,
+   "stages_total": 10,
+   "fell": 0,
+   "replay_s": 5.41,
+   "bad_trajectory": 0,
+   "metric_ball_height_m": 1.628,
+   "metric_touches": 10.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 104,
+   "agent_resets": 106,
+   "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": 14.721,
+   "peak_processes": 25,
+   "peak_disk_gb": 0.241,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-juggling-i00-privileged--ae4cve4/replay.mp4",
+  "duration_s": 9.08,
+  "fps": 25.0,
+  "frames": 227,
+  "sha256": "3e198791e13f775d11036ae95b55e3d98577732ad011385685dcece1b539046a"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 10,
+  "final": 10,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [
+     [
+      3.84,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 3.68,
+      "log_step": null,
+      "text_zh": "观察：视频约 3.68 s，足球从已分开的双掌之间下落至机器人头顶，随后被顶起；已逐帧检查视频 3.56 s 至视频 3.96 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 3.84,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 3.84 s 的离头回升画面定位第1次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第1次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.32,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 4.2,
+      "log_step": null,
+      "text_zh": "观察：视频约 4.20 s，足球再次落到机器人头顶，随后离头回升；已逐帧检查视频 4.08 s 至视频 4.44 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 4.32,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 4.32 s 的离头回升画面定位第2次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第2次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.84,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 4.68,
+      "log_step": null,
+      "text_zh": "观察：视频约 4.68 s，足球下落至机器人头顶，机器人伸膝抬头后球重新上升；已逐帧检查视频 4.56 s 至视频 4.96 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 4.84,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 4.84 s 的离头回升画面定��第3次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第3次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [
+     [
+      5.36,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 5.24,
+      "log_step": null,
+      "text_zh": "观察：视频约 5.24 s，足球落近机器人头顶后再次向上弹离；已逐帧检查视频 5.12 s 至视频 5.48 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 5.36,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 5.36 s 的离头回升画面定位第4次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第4次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [
+     [
+      5.92,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 5.76,
+      "log_step": null,
+      "text_zh": "观察：视频约 5.76 s，机器人在球下落时抬高头部，球离头向上飞起；已逐帧检查视频 5.64 s 至视频 6.00 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 5.92,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 5.92 s 的离头回升画面定位第5次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第5次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [
+     [
+      6.44,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 6.28,
+      "log_step": null,
+      "text_zh": "观察：视频约 6.28 s，足球落到机器人偏左倾的头顶后再度升高；已逐帧检查视频 6.16 s 至视频 6.60 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 6.44,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 6.44 s 的离头回升画面定位第6次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第6次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [
+     [
+      7.08,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 6.96,
+      "log_step": null,
+      "text_zh": "观察：视频约 6.96 s，足球从较高位置落回机器人头顶，随后离头向上飞起；已逐帧检查视频 6.84 s 至视频 7.24 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 7.08,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 7.08 s 的离头回升画面定位第7次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第7次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [
+     [
+      7.72,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 7.6,
+      "log_step": null,
+      "text_zh": "观察：视频约 7.60 s，足球下落接近机器人头顶后再次回升；已逐帧检查视频 7.48 s 至视频 7.92 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 7.72,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 7.72 s 的离头回升画面定位第8次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第8次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [
+     [
+      8.44,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 8.32,
+      "log_step": null,
+      "text_zh": "观察：视频约 8.32 s，足球落至机器人头顶附近后再次上升，与头部分离；已逐帧检查视频 8.20 s 至视频 8.60 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 8.44,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 8.44 s 的离头回升画面定位第9次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第9次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [
+     [
+      9.04,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 8.96,
+      "log_step": null,
+      "text_zh": "观察：视频约 8.96 s，足球最后一次落近机器人头顶后被顶起，末帧球已在头顶上方且机器人仍站立；已逐帧检查视频 8.84 s 至视频 9.04 s 的下落、头顶邻接和离体回升画面，双掌处于球两侧，未见再次夹球。"
+     },
+     {
+      "t": 9.04,
+      "log_step": 112,
+      "text_zh": "推断：以视频约 9.04 s 的离头回升画面定位第10次有效颠球达成，视频时间为近似值（约有 0.1 s 误差）；log step 112 对实际交付轨迹的官方回放明确列出 touch 1 至 touch 10，metrics.touches=10、success=True，支持本次通过高度、时窗及速度判据，非仅凭一次可见碰球计数。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "候选：击打；第10次由机器人头顶短促撞击足球，使球离头上升。现有怎么碰取值不能准确表示击打，故用其他；不是脚或腿击球。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人双脚在草地上承重，双腿屈伸配合头部迎球，未见倒地。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": 10.0
+  },
+  {
+   "name": "ball_height_m",
+   "unit": "m",
+   "threshold": null,
+   "direction": null,
+   "value": 1.628
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人先把双掌夹持的足球举过头顶并释放，在视频约 3.68 s 至视频 9.04 s 连续用头顶击球，十个事件节点全部完成并锁存，末帧机器人仍站立、球仍在空中。评分器记录 touches=10、stages_reached=10/10、success=1；log step 112、113 显示实际交付轨迹两次官方回放均成功且末状态哈希一致。无状态日志，也无独立参考成功视频；各次达成时间按逐帧回升画面近似定位，毫秒级接触时窗和离球/入球速度比例不能从此视角独立测定，其有效性由原生裁判结果确认。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "成功运行，十个触球节点均到达，视频 3.84–9.04 s 的时间与 log step 112 的官方回放（touch 1–10、time_s 9.05）一致；接触标注（头击打=腿/躯干+其他、双脚支撑）符合 S16，无 failure 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": "2026-10-31T00:46:05Z",
+  "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/embodied-olympics-fun-juggling-i00-privileged--ae4cve4/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--ae4cve4/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "98ba1d8f809a7fe686fcc43fbc5cf44d5ed867151fa6e0f48d5dfb95cf2891a2",
+   "video_sha256": "3e198791e13f775d11036ae95b55e3d98577732ad011385685dcece1b539046a",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--d9mbrka.json

@@ -0,0 +1,425 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-privileged--d9mbrka",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 1,
+  "reward": 1.0,
+  "source": "catalog",
+  "raw": {
+   "success": 1,
+   "replay_success": 1,
+   "deterministic": 1,
+   "n_actions": 1015,
+   "sim_steps": 1015,
+   "horizon": 3000,
+   "episode_over": 1,
+   "progress": 1.0,
+   "stages_reached": 10,
+   "stages_total": 10,
+   "fell": 0,
+   "replay_s": 13.15,
+   "bad_trajectory": 0,
+   "metric_ball_height_m": 1.609,
+   "metric_touches": 10.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 92,
+   "agent_resets": 698,
+   "agent_replays": 6,
+   "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/embodied-olympics-fun-juggling-i00-privileged--d9mbrka/replay.mp4",
+  "duration_s": 10.16,
+  "fps": 25.0,
+  "frames": 254,
+  "sha256": "0432f93aaf531c3bac1a02c47e815046202f34743eb1d3a1855ab177f28b7f01"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 10,
+  "final": 10,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.32,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 4.32,
+      "log_step": null,
+      "text_zh": "观察：视频 4.12–4.20 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 4.24–4.40 s，球反向上升并与头分离，构成按顺序观察到的第1次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 4.32,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第1次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第1次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。此前视频 0–4.04 s，机器人用左右平掌共同挤持足球并举至头顶，随后双掌分开释放；该先行接触为双臂挤持，无手指包围或单手夹紧，不能称为抓握。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [
+     [
+      4.76,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 4.76,
+      "log_step": null,
+      "text_zh": "观察：视频 4.56–4.64 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 4.68–4.88 s，球反向上升并与头分离，构成按顺序观察到的第2次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 4.76,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第2次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第2次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [
+     [
+      5.28,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 5.28,
+      "log_step": null,
+      "text_zh": "观察：视频 5.08–5.20 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 5.24–5.48 s，球反向上升并与头分离，构成按顺序观察到的第3次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 5.28,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第3次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第3次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [
+     [
+      5.92,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 5.92,
+      "log_step": null,
+      "text_zh": "观察：视频 5.68–5.80 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 5.84–6.08 s，球反向上升并与头分离，构成按顺序观察到的第4次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 5.92,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第4次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第4次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [
+     [
+      6.56,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 6.56,
+      "log_step": null,
+      "text_zh": "观察：视频 6.36–6.48 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 6.52–6.76 s，球反向上升并与头分离，构成按顺序观察到的第5次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 6.56,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第5次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第5次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [
+     [
+      7.2,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 7.2,
+      "log_step": null,
+      "text_zh": "观察：视频 7.00–7.12 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 7.16–7.40 s，球反向上升并与头分离，构成按顺序观察到的第6次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 7.2,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第6次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第6次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [
+     [
+      7.92,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 7.92,
+      "log_step": null,
+      "text_zh": "观察：视频 7.68–7.84 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 7.88–8.12 s，球反向上升并与头分离，构成按顺序观察到的第7次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 7.92,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第7次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第7次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [
+     [
+      8.64,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 8.64,
+      "log_step": null,
+      "text_zh": "观察：视频 8.44–8.56 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 8.60–8.84 s，球反向上升并与头分离，构成按顺序观察到的第8次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 8.64,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第8次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第8次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [
+     [
+      9.4,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 9.4,
+      "log_step": null,
+      "text_zh": "观察：视频 9.20–9.32 s，足球下降至机器人头顶，机器人屈伸身体迎球；视频 9.36–9.60 s，球反向上升并与头分离，构成按顺序观察到的第9次头顶回弹。节点时间取回升阶段的近似帧锚点。"
+     },
+     {
+      "t": 9.4,
+      "log_step": null,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第9次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第9次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [
+     [
+      10.12,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 10.12,
+      "log_step": null,
+      "text_zh": "观察：视频 9.92–10.08 s，足球下降至机器人头顶；视频 10.12 s 的末帧中球已重新上升，机器人仍双脚站立，双掌分在球两侧。末次回升未完整录出。"
+     },
+     {
+      "t": 10.12,
+      "log_step": 220,
+      "text_zh": "推断：本次评分器字段 replay_success=1、deterministic=1、stages_reached=10、metric_touches=10，结合视频中顺序可辨的十次头顶回弹，判第10次有效计数成立并锁存；接触时长和速度比例采用评分器验收结果，未从像素独立测量。末次起点取视频 10.12 s 的最后可见回升帧，时间约有 0.1 s 不确定性；log step 220 核对最终候选两次成功回放，log step 222–223 确认保存文件及生成器复现一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "机器人以头顶短促击起足球，双腿支撑并通过身体屈伸配合第10次顶球；候选“击打”，现有推/扫、压/擦、托/舀均不能表达短促碰撞后离体上升。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better",
+   "value": 10.0
+  },
+  {
+   "name": "ball_height_m",
+   "unit": "m",
+   "threshold": null,
+   "direction": null,
+   "value": 1.609
+  }
+ ],
+ "failure": null,
+ "summary_zh": "机器人先用双掌将球举至头顶并释放，视频约 4.32–10.12 s 可顺序辨认十次头顶回弹，十个颠球节点最终全部完成。评分器记录 touches=10、stages_reached=10、replay_success=1、deterministic=1，与最终保存轨迹的验证日志及视频表现一致，无失败断点。没有状态日志，毫秒级触球时长、精确回升高度与速度比例不能由视频独立核验，采用评分器验收结果；第十次回升仅录至视频 10.12 s，完成时间为末帧附近的近似锚点。",
+ "conf": 9,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "十个颠球节点与评分器 touches=10、stages_reached=10、replay_success=1 及视频一致，约 4.3–10.1 s 可顺序辨认头顶回弹；节点时间、contact_actual（腿/躯干·其他+击打）和无失败断点均合理，未发现需修改之处。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:42:32Z",
+  "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/embodied-olympics-fun-juggling-i00-privileged--d9mbrka/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--d9mbrka/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "9ca639476ce35ccf71cb3d97cfa27211aaac39101daa529b8a17c7b874f4733d",
+   "video_sha256": "0432f93aaf531c3bac1a02c47e815046202f34743eb1d3a1855ab177f28b7f01",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--de6wseu.json

@@ -0,0 +1,326 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-privileged--de6wseu",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 0,
+  "reward": 0.2,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 120,
+   "sim_steps": 120,
+   "horizon": 3000,
+   "episode_over": 0,
+   "progress": 0.2,
+   "stages_reached": 2,
+   "stages_total": 10,
+   "fell": 0,
+   "replay_s": 4.56,
+   "bad_trajectory": 0,
+   "metric_ball_height_m": 0.651,
+   "metric_touches": 2.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 868,
+   "agent_resets": 1496,
+   "agent_replays": 4,
+   "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": 5.967,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.271,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-juggling-i00-privileged--de6wseu/replay.mp4",
+  "duration_s": 1.24,
+  "fps": 25.0,
+  "frames": 31,
+  "sha256": "5fcabb1fa198db85d4637b0dcc33ec03101c227ff6557dbfb5ddc2abecb3e970"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 2,
+  "final": 2,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [
+     [
+      0.6,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.6,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.12–0.40 s，机器人张开双掌让球下落；视频 0.44–0.60 s，右腿迎球后黑白球由低处回升并与腿分离。log step 590 对最终交付文件的重放记录只含顺序阶段 touch 1、touch 2，第一阶段记录于仿真 0.565 s，确认此次回升通过原生判据；视频达到时刻取约 0.60 s，受帧间隔影响。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "观察：视频 0.44–0.52 s，机器人右脚/小腿迎击足球，另一脚支撑草地；候选：击打，现有 how 项不能准确表达短促踢球。初始双掌夹球及释放发生在击球之前。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [
+     [
+      0.84,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.84,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.64–0.84 s，机器人抬起左腿再次击球，球离腿后继续向前上方飞行；log step 590 对最终交付文件的重放记录 touch 2 于仿真 0.785 s 成立，metric touches=2，评分器字段 metric_touches=2.0。视频达到时刻取约 0.84 s；此前的第一次计数已成立，此后到视频 1.20 s 未见第三次触球。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "观察：视频 0.68–0.80 s，机器人左脚/小腿击足球，右脚在草地支撑；候选：击打，现有 how 项不能准确表达短促踢球。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.96,
+      "log_step": 405,
+      "text_zh": "观察：视频 0.84–1.20 s，机器人收回左腿，球持续向前飞离，躯干从近直立逐渐向后倾倒，始终没有第三次击球。log step 405 展示的交付生成脚本只安排右、左各一次踢球及随后收腿；log step 589–590 重新生成并重放该文件，仍只有两次计数，未达到第三次。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 0.92–1.20 s，机器人在第二次击球后的收腿转场中以脚接触草地尝试恢复站姿，身体持续后倾；没有第三次腿球接触。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.00–1.20 s 全部帧只见两次腿部击球，末段球已飞离机器人，未见第4次击球；log step 590 的最终交付重放阶段仅 touch 1、touch 2，评分器 stages_reached=2、metric_touches=2.0，故第4次顺序计数未达到。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.00–1.20 s 全部帧只见���次腿部击球，末段球已飞离机器人，未见第5次击球；log step 590 的最终交付重放阶段仅 touch 1、touch 2，评分器 stages_reached=2、metric_touches=2.0，故第5次顺序计数未达到。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.00–1.20 s 全部帧只见两次腿部击球，末段球已飞离机器人，未见第6次击球；log step 590 的最终交付重放阶段仅 touch 1、touch 2，评分器 stages_reached=2、metric_touches=2.0，故第6次顺序计数未达到。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.00–1.20 s 全部帧只见两次腿部击球，末段球已飞离机器人，未见第7次击球；log step 590 的最终交付重放阶段仅 touch 1、touch 2，评分器 stages_reached=2、metric_touches=2.0，故第7次顺序计数未达到。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.00–1.20 s 全部帧只见两次腿部击球，末段球已飞离机器人，未见第8次击球；log step 590 的最终交付重放阶段仅 touch 1、touch 2，评分器 stages_reached=2、metric_touches=2.0，故第8次顺序计数未达到。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.00–1.20 s 全部帧只见两次腿部击球，末段球已飞离机器人，未见第9次击球；log step 590 的最终交付重放阶段仅 touch 1、touch 2，评分器 stages_reached=2、metric_touches=2.0，故第9次顺序计数未达到。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.2,
+      "log_step": 590,
+      "text_zh": "观察：视频 0.00–1.20 s 全部帧只见两次腿部击球，末段球已飞离机器人，未见第10次击球；log step 590 的最终交付重放阶段仅 touch 1、touch 2，评分器 stages_reached=2、metric_touches=2.0，故第10次顺序计数未达到。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "value": 2.0,
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better"
+  },
+  {
+   "name": "ball_height_m",
+   "value": 0.651,
+   "unit": "m",
+   "threshold": null,
+   "direction": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "valid_touch_3",
+   "t": 0.96,
+   "log_step": 405,
+   "where": "转场",
+   "evidence_zh": "观察：视频 0.84–0.92 s，第二次击球后机器人收腿，球向前远离；视频 0.96–1.20 s，机器人躯干持续后倾、身体下降，未恢复稳定支撑或接续第三次击球。log step 405 的最终生成脚本在两次踢球后只收腿，log step 590 的交付重放只有两阶段；评分器 episode_over=0、fell=0，交付在正式判倒或球落地之前结束。"
+  },
+  "unmet": {
+   "valid_touch_1": true,
+   "valid_touch_2": true,
+   "valid_touch_3": false,
+   "valid_touch_4": false,
+   "valid_touch_5": false,
+   "valid_touch_6": false,
+   "valid_touch_7": false,
+   "valid_touch_8": false,
+   "valid_touch_9": false,
+   "valid_touch_10": false
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 0.96,
+    "log_step": null,
+    "evidence_zh": "观察：视频 0.96–1.20 s，机器人收腿过程中躯干明显向后翻倾，骨盆及上身逐帧降低，到视频 1.20 s 已大幅偏离直立姿态，期间未恢复；球仍在空中飞离。这里标记可见的失衡，不将视频结束之后的倒地或触地当作已发生。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "source_zh": "日志 / 视频",
+    "evidence_zh": "推断：交付控制未能在第二次踢球后稳定支撑并接续后续颠球。log step 385、405 的生成代码使用预定关节目标，在控制步 80–90 将左腿插值收回，未依据身体姿态或球位修正；log step 589–590 确认该脚本生成的最终文件仍只有两次计数，视频 0.96–1.20 s 显示该收腿阶段持续后倾失衡。log step 552–553 的会话第三次踢球测试另显示脚未跟上飞行中的球，但该测试没有写入最终交付，不用作本次轨迹的触球事实；现有证据不足以细分控制子类。"
+   }
+  ],
+  "gap_zh": "无状态日志，无法逐物理步复核触球部位、接触时窗、回升及速度比例，前两次通过以最终文件重放的裁判阶段和评分器计数佐证，视频时刻为近似。交付在回合尚未终止时结束，不能宣称机器人已触发跌倒判负或足球已经触地；会话中更长测试及第三次踢球测试不属于交付轨迹，不能用来延长节点或补记接触。"
+ },
+ "summary_zh": "机器人释放足球后完成右、左两次有效颠球，最终达到 2/10；两次计数由视频回升动作及最终交付重放的裁判阶段共同支持。第二次击球后的收腿转场中机器人持续后倾失衡，球向前飞离，断点为第3次颠球，根因归为控制；最终脚本没有第三次踢球动作。缺少状态日志，精确接触阈值无法独立复核，且交付结束时裁判尚未判倒或判球触地。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 6,
+    "summary_zh": "两次击球节点、第3次起未达到、断点valid_touch_3与后倾失衡现象均与视频帧及评分器字段(stages_reached=2、metric_touches=2、fell=0)一致，未发现需改动的内容。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:44:30Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--de6wseu/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--de6wseu/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "1410a6d920f8cfb1dc7682a39ea6c6eec9b419a1ec270f16b651c8c3ce04c87e",
+   "video_sha256": "5fcabb1fa198db85d4637b0dcc33ec03101c227ff6557dbfb5ddc2abecb3e970",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--mwme26l.json

@@ -0,0 +1,299 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-privileged--mwme26l",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 180,
+   "sim_steps": 180,
+   "horizon": 3000,
+   "episode_over": 0,
+   "progress": 0.0,
+   "stages_reached": 0,
+   "stages_total": 10,
+   "fell": 0,
+   "replay_s": 4.47,
+   "bad_trajectory": 0,
+   "metric_ball_height_m": 0.837,
+   "metric_touches": 0.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 179,
+   "agent_resets": 13,
+   "agent_replays": 8,
+   "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/embodied-olympics-fun-juggling-i00-privileged--mwme26l/replay.mp4",
+  "duration_s": 1.84,
+  "fps": 25.0,
+  "frames": 46,
+  "sha256": "687f9e510a9cfcb66f557a579103787b658de59759e0bbe76060793703a59ba3"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0.0,
+      "log_step": null,
+      "text_zh": "观察：逐帧检查视频 0.00–1.80 s（全片46帧），机器人保持双掌夹球，仅轻微调整站姿；球始终在腹前双掌之间，没有离手下落，也没有脚、腿或头击球后的离体回升，第1次颠球未启动。初始持球接触未转入颠球操作，故不列该节点的 contact_actual。"
+     },
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：log step 249 的最终生成脚本只运行 settle 与 narrow 两阶段；log step 250 对该180行交付文件两次回放均报告 touches=0、ball_height_m=0.837，末状态哈希相同；inputs/run.json 的评分器字段 metric_touches=0、stages_reached=0、fell=0、episode_over=0，与视频末帧仍持球一致。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第2次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第3次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第4次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第5次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第6次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第7次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第8次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第9次颠球未发生。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.8,
+      "log_step": 250,
+      "text_zh": "观察：视频 0.00–1.80 s 全片中球一直留在机器人双掌间，未出现离手后连续弹起的球路；log step 250 的最终交付回放及评分器 metric_touches 均为0，第10次颠球未发生。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "value": 0.0,
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better"
+  },
+  {
+   "name": "ball_height_m",
+   "value": 0.837,
+   "unit": "m",
+   "threshold": null,
+   "direction": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "valid_touch_1",
+   "t": 0.0,
+   "log_step": 249,
+   "where": "首步前",
+   "evidence_zh": "观察：视频 0.00–1.80 s 机器人始终双掌持球，虽收窄站姿，但未释放球或发起首次颠球。log step 249 明确将最终脚本限制为 settle 0.30 s 与 narrow 1.50 s 两个仿真准备阶段并保存；log step 250 证实此交付仅180行、0次有效颠球。评分器未报告跌倒或回合终止，失败发生在首次颠球尚未启动的交付前缀。"
+  },
+  "unmet": {
+   "valid_touch_1": false,
+   "valid_touch_2": false,
+   "valid_touch_3": false,
+   "valid_touch_4": false,
+   "valid_touch_5": false,
+   "valid_touch_6": false,
+   "valid_touch_7": false,
+   "valid_touch_8": false,
+   "valid_touch_9": false,
+   "valid_touch_10": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0.0,
+    "log_step": 249,
+    "evidence_zh": "观察：视频 0.00–1.80 s 只有机器人持球站立和双脚收窄，球始终没有离开双掌，未出现向脚、腿或头落球的颠球尝试。log step 249 保存的脚本主动省略 lift、release、juggle 等后续阶段；这是未启动第1次颠球的基线前缀。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "log_step": 248,
+    "source_zh": "日志",
+    "evidence_zh": "推断：未能形成稳定的抬脚控制使 agent 最终只交付持球准备前缀。会话测试 log step 243、245、247、248 的工具输出均显示 narrow 阶段后，在 lift 阶段侧向重心偏移并以 failed=fell 结束、touches=0；log step 248 的低踢变体仍未解决，随后 log step 249 重新生成只含 settle+narrow 的交付，log step 250 验证0次颠球。上述测试解释控制能力缺口，不是本次视频内发生的跌倒；证据不足以细分具体反馈或逆解机制。"
+   }
+  ],
+  "gap_zh": "本目录没有状态日志，无法测量毫秒级接触与力矩；会话中的较长测试不属于评分视频，不能把其中跌倒作为本次失败现象。日志支持抬脚阶段控制未稳定，但无法独立确认 agent 所猜测的髋偏航反力等单一机制，根因仅标控制。参考成功视频仅用于确认可见的离手及头顶弹球形态，不用于本次计数。"
+ },
+ "summary_zh": "观察：本次评分视频 0.00–1.80 s 仅显示机器人双掌持球并收窄站姿，球没有离手或被击起，十个颠球节点均未达到，评分器有效次数为0。推断：断点为第1次颠球的首步前；会话抬脚测试反复控制失稳，agent 最终交付仅含持球准备的180行前缀，本次视频未跌倒。缺少状态日志及会话测试视频，无法确认侧向失稳的具体机制。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "视频全程持球、无颠球，10个节点均未达成，断点首步前+没尝试、根因控制（会话证据）均符合规则；日志步号与视频时钟无误。",
+    "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": "2026-10-31T00:42:44Z",
+  "review_prompt": "prompts/04_trajectory_review.md",
+  "review_prompt_sha256": "9afbc4ac19b1e1cf03732c8961ec46a618d682ee229343f66e3316794b14f9ba",
+  "reconcile_prompt": "prompts/05_reconcile.md",
+  "reconcile_prompt_sha256": "8c827af3ae3f77d09229399fb41283c50a8f725302cbe3b263b8051213dde2b2",
+  "calls": [
+   {
+    "stage": "traj",
+    "round": 0,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--mwme26l/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--mwme26l/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "0b838041f476b9ffc54e2b27a6b0a31d6022febc5d8591753631d99417b977b5",
+   "video_sha256": "687f9e510a9cfcb66f557a579103787b658de59759e0bbe76060793703a59ba3",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--zg4fqfs.json

@@ -0,0 +1,309 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-privileged--zg4fqfs",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 221,
+   "sim_steps": 221,
+   "horizon": 3000,
+   "episode_over": 1,
+   "progress": 0.0,
+   "stages_reached": 0,
+   "stages_total": 10,
+   "fell": 0,
+   "replay_s": 4.39,
+   "bad_trajectory": 0,
+   "metric_ball_height_m": 0.097,
+   "metric_touches": 0.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 69,
+   "agent_resets": 1,
+   "agent_replays": 1,
+   "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": 0.977,
+   "peak_processes": 9,
+   "peak_disk_gb": 0.98,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-juggling-i00-privileged--zg4fqfs/replay.mp4",
+  "duration_s": 2.24,
+  "fps": 25.0,
+  "frames": 56,
+  "sha256": "dab2feb1257d87cc05c114aae36f4cbbc86889bcac7f14b96dedd36283c62387"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–1.84 s 足球保持在机器人双掌之间；视频 1.88–2.20 s 松掌后球下落，机器人抬左腿，球经过左膝/小腿前侧后继续下降至草地，没有出现离体回升一个球直径的过程。log step 126 对最终交付文件的重放返回 touches=0、stages=[]、failed=\"the ball touched the ground\"，与视频末段一致。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "观察：视频 0.00–1.84 s 机器人两只平掌从两侧共同夹住足球，随后双掌分开使球下落。这是首次颠球前的持球及释放准备，依次先于腿部迎球；候选：双掌挤压。没有单手包合，现有方式无准确对应项，故不标抓握。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "观察：视频 1.96–2.08 s 机器人抬左腿迎向下落足球，球与左膝/小腿前侧相接并偏移，随后继续下落；未见脚背从球底将球击起。候选：击打；现有接触方式无击打项，具体碰撞连杆仅凭该视角不能精确区分。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第2次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第3次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第4次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第5次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第6次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第7次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第8次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第9次有效颠球也未达到。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 2.2,
+      "log_step": 126,
+      "text_zh": "观察：视频 0.00–2.20 s 仅有初始持球、松掌和一次左腿迎球，随后球落地，没有连续颠球。log step 126 的最终文件重放返回 touches=0、stages=[]；评分器 metric_touches=0。推断：第1次有效颠球尚未成立，按计数顺序，第10次有效颠球也未达到。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "value": 0.0,
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better"
+  },
+  {
+   "name": "ball_height_m",
+   "value": 0.097,
+   "unit": "m",
+   "threshold": null,
+   "direction": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "valid_touch_1",
+   "t": 2.04,
+   "log_step": 126,
+   "where": "操作中",
+   "evidence_zh": "观察：视频 1.96–2.08 s 机器人已抬左腿尝试迎球，但下落足球经过膝/小腿前侧后未转为上升，视频 2.20 s 球到达草地，机器人仍站立。log step 123 保存221行最终轨迹，log step 126 重放同一交付文件返回0次触球及球触地终止；首次颠球始终未补成。"
+  },
+  "unmet": {
+   "valid_touch_1": false,
+   "valid_touch_2": false,
+   "valid_touch_3": false,
+   "valid_touch_4": false,
+   "valid_touch_5": false,
+   "valid_touch_6": false,
+   "valid_touch_7": false,
+   "valid_touch_8": false,
+   "valid_touch_9": false,
+   "valid_touch_10": false
+  },
+  "phenomena": [
+   {
+    "cat": "运动",
+    "item": "动态时机",
+    "t": 2.04,
+    "log_step": 126,
+    "evidence_zh": "观察：视频 1.96–2.08 s 足球下落时，机器人左腿正在抬起，球先经过膝/小腿前侧，脚仍在球下方偏后，未形成由脚背向上的迎击；视频 2.12–2.20 s 球继续落地。推断：左腿的迎球动作没有与球的下落配合到有效上击的时空位置，首次颠球失配；不是已经有效颠起后再次漏接。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "抓取和接触设计",
+    "alt_feasible": "未知",
+    "log_step": 117,
+    "evidence_zh": "推断：机器人所用左脚迎球布局没有避开同侧膝/小腿对足球下落路径的先行接触，导致脚背尚未到球底时球已被腿部偏转。依据是视频 2.00–2.08 s 的膝/小腿邻接及随后侧移下降；会话诊断 log step 113、115、117 也记录抬腿时球在膝附近出现水平速度突变，而脚仍较低（这些是调试测试，只用于解释根因）。log step 123 最终只修改目标触球高度和脚端高度，重放 log step 126 仍为0次并触地。未验证该左脚布局的可行替代。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "观察：inputs/ 无状态日志；视频仅有一个远景视角，放大后可定位腿部迎球及球落地，但无法精确区分首次接触的膝或小腿碰��壳，也不能量出毫秒级接触时长和击球速度。推断：接触布局根因结合会话诊断与末段画面判断，未把会话中的其他测试时刻或计数移入本次评分轨迹。"
+ },
+ "summary_zh": "本次评分轨迹完成了双掌释放和左腿迎球尝试，但没有完成第1次有效颠球，最终为0/10。视频约 2.04 s 左腿未在球下落时形成有效向上迎击，球随后触地；推断主要问题是迎球布局让膝/小腿先碰到球，最终文件重放同样报告0次。缺少状态日志，具体碰撞连杆及接触参数不能精确复核。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "视频 0–1.84 s 双掌持球，随后松掌，约 2.0 s 球经过左膝/小腿前侧后继续下落，2.2 s 触地；没有离体回升一个球直径的触球，valid_touch_1..10 全为 false、final=0 与评分器 touches=0、stages 0/10 一致。breakpoint 在 valid_touch_1、动态时机与根因判断有视频支持，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": "2026-10-31T00:42:21Z",
+  "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/embodied-olympics-fun-juggling-i00-privileged--zg4fqfs/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--zg4fqfs/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "5f25abd1521831220ffd6f36cac383bf0a1e1eb649ce20f7371af67f47b90a4a",
+   "video_sha256": "dab2feb1257d87cc05c114aae36f4cbbc86889bcac7f14b96dedd36283c62387",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-privileged--zmql697.json

@@ -0,0 +1,338 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-privileged--zmql697",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 0,
+  "reward": 0.3,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "replay_success": 0,
+   "deterministic": 1,
+   "n_actions": 174,
+   "sim_steps": 174,
+   "horizon": 3000,
+   "episode_over": 0,
+   "progress": 0.3,
+   "stages_reached": 3,
+   "stages_total": 10,
+   "fell": 0,
+   "replay_s": 3.39,
+   "bad_trajectory": 0,
+   "metric_ball_height_m": 0.909,
+   "metric_touches": 3.0,
+   "grader_error": 0,
+   "video_rendered": 1,
+   "agent_sessions": 611,
+   "agent_resets": 0,
+   "agent_replays": 16,
+   "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": 1.119,
+   "peak_processes": 10,
+   "peak_disk_gb": 0.849,
+   "resource_emergency_kills": 0,
+   "resource_oom_kills": 0,
+   "resource_host_oom_kills": 0,
+   "resource_service_oom_kills": 0,
+   "session_log_tampered": 0,
+   "session_log_from_monitor": 0
+  }
+ },
+ "video": {
+  "kind": "hf",
+  "url": "https://huggingface.co/datasets/MLL-Lab/embodied-exam-media/resolve/main/runs/embodied-olympics-fun-juggling-i00-privileged--zmql697/replay.mp4",
+  "duration_s": 1.76,
+  "fps": 25.0,
+  "frames": 44,
+  "sha256": "089ed7fb778a15bff32a589945c514ac1e5efacb8b380d82d38c9c623841dac6"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 3,
+  "final": 3,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [
+     [
+      0.48,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 0.48,
+      "log_step": 703,
+      "text_zh": "观察：视频 0.20–0.36 s，机器人张开双掌，球下落到抬起的腿脚处；视频 0.40–0.68 s 球离开腿部并明显回升。log step 701 保存最终文件后，log step 703 的两次独立回放均报告 touch 1、touch 2、touch 3，末状态哈希相同，且球高 0.909 m，与本次评分器字段一致；据画面回升将本节点达到时刻标在视频约 0.48 s，精确接触时窗及速度比例由裁判计数支持。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "观察：机器人用腿脚短促击起足球，球随后离体回升；候选“击打”，现有方式中的推/扫、压/擦和支撑均不能准确表示这种短促碰撞。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [
+     [
+      1.04,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 1.04,
+      "log_step": 703,
+      "text_zh": "观察：视频 0.80–0.96 s，机器人另一条腿迎向下落的球；视频 1.00–1.20 s 球再次离腿上升。log step 701 保存最终文件后，log step 703 的两次独立回放均报告 touch 1、touch 2、touch 3，末状态哈希相同，且球高 0.909 m，与本次评分器字段一致；据画面回升将本节点达到时刻标在视频约 1.04 s，精确接触时窗及速度比例由裁判计数支持。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "观察：机器人用腿脚短促击起足球，球随后离体回升；候选“击打”，现有方式中的推/扫、压/擦和支撑均不能准确表示这种短促碰撞。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [
+     [
+      1.6,
+      null
+     ]
+    ],
+    "final": true,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 703,
+      "text_zh": "观察：视频 1.40–1.48 s，球下降到机器人抬起的腿部附近；视频 1.52–1.72 s 球再次离体上升。log step 701 保存最终文件后，log step 703 的两次独立回放均报告 touch 1、touch 2、touch 3，末状态哈希相同，且球高 0.909 m，与本次评分器字段一致；据画面回升将本节点达到时刻标在视频约 1.60 s，精确接触时窗及速度比例由裁判计数支持。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "其他",
+      "note_zh": "观察：机器人在侧倾调整身体时，抬起的腿部附近触球并使球弹起；候选“击打”，现有方式无短促击球项。腿部与球有局部遮挡，无法可靠细分为脚、小腿或大腿。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.64,
+      "log_step": 703,
+      "text_zh": "观察：视频 1.60–1.72 s，第三次回升后的足球仍在空中，与机器人逐渐分离；机器人侧后倾并继续降低身体，没有出现第四次击球及回升。log step 703 的最终文件回放均只列出三个 touch 阶段；评分器 metric_touches=3、stages_reached=3，故第4次未完成。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "观察：视频 1.60–1.72 s，机器人脚部接地并调整双腿支撑姿态，但身体继续侧后倾；该段未再次触到足球。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.72,
+      "log_step": 703,
+      "text_zh": "观察：检查视频 0.00–1.72 s 全部帧，仅见三段击球回升，末段球仍悬空；log step 703 的最终文件两次回放及评分器均确认累计仅3次。推断：顺序计数未达到第5次，本节点未完成，视频中也无该节点的独立尝试。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.72,
+      "log_step": 703,
+      "text_zh": "观察：检查视频 0.00–1.72 s 全部帧，仅见三段击球回升，末段球仍悬空；log step 703 的最终文件两次回放及评分器均确认累计仅3次。推断：顺序计数未达到第6次，本节点未完成，视频中也无该节点的独立尝试。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.72,
+      "log_step": 703,
+      "text_zh": "观察：检查视频 0.00–1.72 s 全部帧，仅见三段击球回升，末段球仍悬空；log step 703 的最终文件两次回放及评分器均确认累计仅3次。推断：顺序计数未达到第7次，本节点未完成，视频中也无该节点的独立尝试。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.72,
+      "log_step": 703,
+      "text_zh": "观察：检查视频 0.00–1.72 s 全部帧，仅见三段击球回升，末段球仍悬空；log step 703 的最终文件两次回放及评分器均确认累计仅3次。推断：顺序计数未达到第8次，本节点未完成，视频中也无该节点的独立尝试。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.72,
+      "log_step": 703,
+      "text_zh": "观察：检查视频 0.00–1.72 s 全部帧，仅见三段击球回升，末段球仍悬空；log step 703 的最终文件两次回放及评分器均确认累计仅3次。推断：顺序计数未达到第9次，本节点未完成，视频中也无该节点的独立尝试。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.72,
+      "log_step": 703,
+      "text_zh": "观察：检查视频 0.00–1.72 s 全部帧，仅见三段击球回升，末段球仍悬空；log step 703 的最终文件两次回放及评分器均确认累计仅3次。推断：顺序计数未达到第10次，本节点未完成，视频中也无该节点的独立尝试。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "value": 3.0,
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better"
+  },
+  {
+   "name": "ball_height_m",
+   "value": 0.909,
+   "unit": "m",
+   "threshold": null,
+   "direction": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "valid_touch_4",
+   "t": 1.6,
+   "log_step": 703,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 1.60 s 第三次回升已发生，机器人仍在移动双腿调整身体，随后至视频 1.72 s 持续侧后倾、骨盆下沉，球与身体分开而未再被击起。log step 701 保存的174行文件在 log step 703 两次回放均停在3次计数、done=false、failed=null；评分器 episode_over=0、fell=0。推断：连续颠球在准备后续触球时中断，下一未完成步骤为第4次；没有证据把本段称为已经倒地或球已触地。"
+  },
+  "phenomena": [
+   {
+    "cat": "全身",
+    "item": "平衡",
+    "t": 1.6,
+    "log_step": null,
+    "evidence_zh": "观察：视频 1.60、1.64、1.68、1.72 s 连续帧中，机器人上身持续侧后倾，骨盆逐渐下沉，末帧身体已明显偏离直立支撑姿态，未恢复。推断：机器人未能稳住连续踢球后的全身平衡；这里记录可见的失衡过程，评分器 fell=0，尚无倒地判罚。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "evidence_zh": "推断：视频 1.60–1.72 s 中，机器人调整双腿仍未抑制持续侧后倾和身体下沉，无法维持后续击球所需的支撑姿态，支持全身控制不足。log step 701、704 显示最终脚本仍使用双腿支撑力矩（左0.54、右0.535倍 support_tau），最终回放 log step 703 只达到3次；现有证��不足以进一步区分轨迹跟踪、反馈策略或具体参数问题。",
+    "source_zh": "视频 / 日志"
+   }
+  ],
+  "gap_zh": "没有状态日志；25 fps 的远景无法独立测出毫秒级接触时窗、0.95速度比及第三次击球的精确腿段，前三次有效性由与画面一致的最终回放裁判阶段支持。交付轨迹在球仍悬空、裁判尚未判负时结束，不能推断继续执行后一定跌倒或漏球；全身控制的具体失效机制也无法确定。",
+  "unmet": {
+   "valid_touch_1": true,
+   "valid_touch_2": true,
+   "valid_touch_3": true,
+   "valid_touch_4": false,
+   "valid_touch_5": false,
+   "valid_touch_6": false,
+   "valid_touch_7": false,
+   "valid_touch_8": false,
+   "valid_touch_9": false,
+   "valid_touch_10": false
+  }
+ },
+ "summary_zh": "观察：交付轨迹释放足球后，以腿部完成三次有效颠球，视频约 0.48、1.04、1.60 s 的回升与最终文件两次一致回放的3次计数相符，最终完成3/10节点。推断：第4次颠球的准备过程中，机器人未稳住身体并持续侧后倾，支撑控制不足；轨迹在球仍悬空且未触发裁判终局时结束。没有状态日志，精细接触阈值依靠裁判记录，无法确认截断后的动作及具体控制故障。",
+ "conf": 7,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "三次有效颠球（约0.48/1.04/1.60 s）与评分器 touches=3 一致，第4次为断点，机器人后段侧后倾失衡的画面可见；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": "2026-10-31T00:44:47Z",
+  "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/embodied-olympics-fun-juggling-i00-privileged--zmql697/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-privileged--zmql697/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "bb1cc7e6fce8c5a2b2591e16f89c50c1c4e092b34215089e191761a4151bb470",
+   "video_sha256": "089ed7fb778a15bff32a589945c514ac1e5efacb8b380d82d38c9c623841dac6",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-standard--5dytmqi.json

@@ -0,0 +1,311 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-standard--5dytmqi",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 0,
+   "sim_steps": 0,
+   "calls": 1,
+   "rows": 0,
+   "resets": 0,
+   "reset_budget": 5,
+   "sim_startups": 1,
+   "elapsed_s": 7089.4,
+   "video_rendered": 1,
+   "first_success": 0,
+   "first_reason_code": 4,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.296,
+   "peak_processes": 10,
+   "peak_disk_gb": 2.547,
+   "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/embodied-olympics-fun-juggling-i00-standard--5dytmqi/replay.mp4",
+  "duration_s": 0.04,
+  "fps": 25.0,
+  "frames": 1,
+  "sha256": "3a2b3d2acae150890a4ff672e79a8dfb55d1ab15b43f610833cc238c6b1d2aef"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "观察：视频唯一一帧（视频 0.00 s，片长 0.04 s）中机器人直立，足球仍在身前双掌之间，没有离手或合法部位击球；log step 8 的正式机器人服务观测给出 control_steps=0、sim_time_s=0、rows=0。评分器最终记录 n_actions=0、sim_steps=0、rows=0，正式 episode 未推进，第1次颠球未开始。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第2次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第3次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第4次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第5次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第6次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍��双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第7次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第8次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第9次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 0,
+      "log_step": 8,
+      "text_zh": "推断：视频 0.00 s 的唯一画面仍是双掌夹球初态；log step 8 正式服务记录 control_steps=0，评分器最终 n_actions=0、sim_steps=0、rows=0。整个受评 episode 未执行动作，第1次有效颠球尚未发生，故顺序累计的第10次颠球亦未达到；本地测试的计数不计入本次视频。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "value": null,
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better"
+  },
+  {
+   "name": "ball_height_m",
+   "value": null,
+   "unit": "m",
+   "threshold": null,
+   "direction": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "valid_touch_1",
+   "t": 0,
+   "log_step": 8,
+   "where": "首步前",
+   "evidence_zh": "观察：视频 0.00 s（全部视频仅此一帧）显示机器人保持初始双掌夹球姿态，未释放或击球。log step 8 仅调用正式服务的 get_obs（读取相机与传感器，不推进仿真），返回 control_steps=0；评分器最终 n_actions=0、sim_steps=0、rows=0。第1次颠球的工作没有启动，视频 0.00 s 仅作为初态导航锚点。"
+  },
+  "unmet": {
+   "valid_touch_1": false,
+   "valid_touch_2": false,
+   "valid_touch_3": false,
+   "valid_touch_4": false,
+   "valid_touch_5": false,
+   "valid_touch_6": false,
+   "valid_touch_7": false,
+   "valid_touch_8": false,
+   "valid_touch_9": false,
+   "valid_touch_10": false
+  },
+  "phenomena": [
+   {
+    "cat": "没尝试",
+    "item": "没尝试",
+    "t": 0,
+    "log_step": 8,
+    "evidence_zh": "观察：视频 0.00 s 的唯一帧中球仍夹在双掌间，机器人站在初位；log step 8 正式服务返回 control_steps=0、rows=0，评分器最终也记录 n_actions=0、sim_steps=0。日志后续运行的是 agent 自建本地模型，正式受评机器人从未开始第1次颠球。"
+   }
+  ],
+  "root_causes": [],
+  "gap_zh": "没有状态日志，正式受评视频仅包含视频 0.00 s 的初始一帧，不能展示任何运动过程；但评分器记录 n_actions=0、sim_steps=0，足以确认正式 episode 未推进。log step 12、17-18 显示 agent 自建 MuJoCo 模型并开发控制器，log steps 74、153、154 的 fall_pelvis 均为本地测试结果；log step 155 预算服务报告剩余墙钟 69.7 秒，run.wall_s=7207、time_limit_s=7200，整个会话结束前仍未向正式服务提交动作。自建模型与正式服务的物理一致性未验证，本地失衡不能迁移为受评机器人的控制根因；按 DECISIONS.md D23 Now (a) 保留根因空层，不以预算耗尽作根因。正式 episode 有零动作初态视频，按 03_trajectory.md Step 3.1-3.2 保留首步前及没尝试。评分器未报告 touches、ball_height_m，原生指标值保留 null；本地测试输出不替代正式指标。"
+ },
+ "summary_zh": "受评视频只有视频 0.00 s 的初始双掌夹球画面，评分器记录动作数与仿真步数均为 0，十个颠球节点均未达到（0/10）。断点为第1次颠球／首步前，现象是没尝试；agent 在自建模型中开发控制器直至墙钟预算结束，未向正式服务提交动作。正式 episode 没有运动证据，自建模型与正式服务的一致性未核实，本地失衡不足以确定受评机器人的失败根因，故根因留空。",
+ "conf": 6,
+ "review": {
+  "status": "agreed",
+  "rounds": 2,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "fix",
+    "conf": 6,
+    "summary_zh": "节点全 false、断点首步前/没尝试均有评分器 n_actions=0、sim_steps=0 与唯一帧支持，无异议。问题在根因：“控制”完全依据 agent 在自建 MuJoCo 模型里的失衡测试，那不是受评机器人，日志中正式服务从未提交动作（所有步 sim.submit=0），不能据此归因正式机器人的控制能力；按 DECISIONS D23 Now 应留空并写 gap_zh。",
+    "n_issues": 1,
+    "n_major": 1
+   },
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 2,
+    "verdict": "agree",
+    "conf": 8,
+    "summary_zh": "第1轮问题已修复：根因层为空并有 gap_zh；节点全 false、断点首步前/没尝试均有评分器 n_actions=0、sim_steps=0 与唯一帧支持，未发现新错误。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:50:57Z",
+  "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/embodied-olympics-fun-juggling-i00-standard--5dytmqi/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-standard--5dytmqi/04_review_r1"
+   },
+   {
+    "stage": "reconcile",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-standard--5dytmqi/05_reconcile_r1"
+   },
+   {
+    "stage": "traj_review",
+    "round": 2,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-standard--5dytmqi/04_review_r2"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "8b39c3f261b8f62c166e0bdefd835095a98204285103def09f947a24635927f8",
+   "video_sha256": "3a2b3d2acae150890a4ff672e79a8dfb55d1ab15b43f610833cc238c6b1d2aef",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}

phase2/embodied-olympics-fun/juggling/trajectories/embodied-olympics-fun-juggling-i00-standard--eeqc8ou.json

@@ -0,0 +1,318 @@
+{
+ "schema": "efe-trajectory-labels/1",
+ "taxonomy": "v2",
+ "runId": "embodied-olympics-fun-juggling-i00-standard--eeqc8ou",
+ "taskKey": "embodied-olympics-fun/juggling",
+ "dag_sha256": "25e819f9f1251e0a7ce9add4f150f9227df00d011426943d8c29867bd2eff282",
+ "outcome": {
+  "success": 0,
+  "reward": 0.0,
+  "source": "catalog",
+  "raw": {
+   "success": 0,
+   "live_success": 0,
+   "graded_by_live": 1,
+   "replays": 0,
+   "live_rule": 1,
+   "live_verdict_error": 0,
+   "finalized": 1,
+   "missing_log": 0,
+   "infra_error": 0,
+   "grader_error": 0,
+   "sim_error": 0,
+   "sim_crashed_on_call": 0,
+   "instance_mismatch": 0,
+   "n_actions": 163,
+   "sim_steps": 7189,
+   "calls": 681,
+   "rows": 7189,
+   "resets": 47,
+   "reset_budget": 50,
+   "sim_startups": 1,
+   "elapsed_s": 4864.3,
+   "video_rendered": 1,
+   "resource_enforced": 0,
+   "limits_violation_observed": 0,
+   "resource_monitor_missing": 0,
+   "peak_gpu_memory_gb": 0,
+   "peak_memory_gb": 1.144,
+   "peak_processes": 15,
+   "peak_disk_gb": 2.641,
+   "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/embodied-olympics-fun-juggling-i00-standard--eeqc8ou/replay.mp4",
+  "duration_s": 1.64,
+  "fps": 25.0,
+  "frames": 41,
+  "sha256": "4e1040735a041317bda072ea545225ecbe9e3a950fb2e2e30b7db4292143d5ce"
+ },
+ "state": null,
+ "progress": {
+  "method": "visual",
+  "denominator": 10,
+  "base": 0,
+  "peak": 0,
+  "final": 0,
+  "nodes": {
+   "valid_touch_1": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.24,
+      "log_step": null,
+      "text_zh": "观察：已逐帧检查视频 0.00–1.60 s。机器人先用双掌夹球并屈膝、调整脚位；视频 1.16 s 球仍在双掌附近，视频 1.24–1.40 s 球降到掌下和腿前，视频 1.44–1.60 s 继续落向草地，未见合法部位将球击起并回升一个球直径。"
+     },
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：log step 242 明确运行交付的 gen.py 作为最后评分局；机器人服务返回 episode=48、control_steps=163、sim_time_s=1.63、outcome=\"the ball touched the ground\"、stages_reached=0、stages_total=10。视频 1.60 s 末帧球已在脚前接近草地，和该局结束状态一致，第1次有效颠球未成立。"
+     },
+     {
+      "t": 1.24,
+      "log_step": 197,
+      "text_zh": "推断：log step 197 保存的交付脚本以每控制步 0.01 s 运行，log step 242 最后评分局的 163 步对应仿真 1.63 s，与视频总长 1.64 s 及视频 1.56–1.60 s 球落近地面的事件相符，可在一帧量级误差内对照。脚本设仿真 1.55 s 才开始张掌，视频约 1.24 s 球已在掌下，因此是持球摆臂过程中提前脱落，而非预定释放后完成颠球。"
+     }
+    ],
+    "contact_actual": [
+     {
+      "who": "双臂协作",
+      "how": "其他",
+      "note_zh": "候选：双掌挤夹。机器人两只平掌从足球两侧同时挤压持球并随双臂移球，随后球从掌间脱出；没有手指包围球，抓握不适用。该初始持球及释放过程并入首次颠球的尝试。"
+     },
+     {
+      "who": "腿/躯干",
+      "how": "支撑",
+      "note_zh": "机器人用双脚与草地接触承重，屈膝、移重并调整脚位，为首次脚部接球作准备；未观察到把球击起的有效脚、腿或头部接触，不据计划添加击打接触。"
+     }
+    ]
+   },
+   "valid_touch_2": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第2次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_3": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第3次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_4": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：��4次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_5": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第5次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_6": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第6次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_7": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第7次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_8": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第8次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_9": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第9次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   },
+   "valid_touch_10": {
+    "method": "visual",
+    "intervals": [],
+    "final": false,
+    "evidence": [
+     {
+      "t": 1.6,
+      "log_step": 242,
+      "text_zh": "观察：视频 0.00–1.60 s 仅有首次持球准备、球离掌并落向脚前草地的过程，没有连续颠起；最后评分局 episode=48 的机器人服务在 log step 242 返回 stages_reached=0/10。推断：第10次有效颠球及其顺序前置均未成立，该节点未被实际执行。"
+     }
+    ]
+   }
+  },
+  "constraints": {}
+ },
+ "native_metrics": [
+  {
+   "name": "touches",
+   "value": 0,
+   "unit": "count",
+   "threshold": 10,
+   "direction": "higher_better"
+  },
+  {
+   "name": "ball_height_m",
+   "value": null,
+   "unit": "m",
+   "threshold": null,
+   "direction": null
+  }
+ ],
+ "failure": {
+  "breakpoint": {
+   "ref": "valid_touch_1",
+   "t": 1.24,
+   "log_step": 242,
+   "where": "操作中",
+   "evidence_zh": "观察：视频约 1.24 s，机器人为首次脚部颠球而持球调整姿势时，足球已降到双掌下方，随后在视频 1.28–1.60 s 持续落向脚前草地，没有被有效颠起。log step 242 最后评分局的机器人服务报告球触地、有效阶段 0/10；首次颠球的失败没有得到修复。"
+  },
+  "unmet": {
+   "valid_touch_1": false,
+   "valid_touch_2": false,
+   "valid_touch_3": false,
+   "valid_touch_4": false,
+   "valid_touch_5": false,
+   "valid_touch_6": false,
+   "valid_touch_7": false,
+   "valid_touch_8": false,
+   "valid_touch_9": false,
+   "valid_touch_10": false
+  },
+  "phenomena": [
+   {
+    "cat": "接触",
+    "item": "没保持住",
+    "t": 1.24,
+    "log_step": 197,
+    "evidence_zh": "观察：视频 1.16 s 足球仍在机器人双掌附近，视频 1.24–1.32 s 已滑到双掌下方、腿部前方并继续下落。log step 197 保存的交付动作将张掌安排在仿真 1.55 s，log step 242 确认最后评分局运行此交付脚本。推断：球在预定张掌前从双掌挤夹中提前脱落，机器人未保持住持球接触，直至球触地也未补救。"
+   }
+  ],
+  "root_causes": [
+   {
+    "cat": "控制",
+    "item": "控制",
+    "t": 1.24,
+    "log_step": 197,
+    "evidence_zh": "推断：交付控制未能在移重、屈膝与持球摆臂时保持球和双掌的稳定相对位置，导致预定释放前滑落。依据是视频 0.48–1.24 s 机器人明显摇摆、球随后降到掌下；log step 197 的交付脚本在��真 1.20–1.55 s 执行持球摆臂，仿真 1.55 s 才张掌，脚踝颠球动作更晚才启动；log step 242 的同一评分局还报告仿真 1.30 s 起左踝俯仰目标被限位裁剪。视频和指令偏离支持控制层归因，但不足以确定掌间受力、轨迹跟踪或反馈增益中哪项主导。",
+    "source_zh": "日志 / 视频"
+   }
+  ],
+  "gap_zh": "没有状态日志或双掌接触力记录，无法确定滑脱的毫秒级时刻和具体受力原因，也无法排除下降中极短的腿部擦碰；视频及最后评分局服务的 0/10 阶段记录足以确认所有有效颠球节点均未达成。视频最后采样帧为视频 1.60 s，球触地的终止判定来自 log step 242 的仿真 1.63 s 服务结果；ball_height_m 未报告，保留 null。会话中较早的跌倒和抬膝试验均未作为本评分局现象。"
+ },
+ "summary_zh": "最后评分局中，机器人完成了持球移重和脚位调整，但在视频约 1.24 s 持球摆臂时足球提前滑出双掌，未形成第1次有效颠球，十个节点最终均未达成。log step 242 的机器人服务以球触地结束该局并报告 0/10；现象为没保持住，根因归于控制。缺少掌间接触力和状态日志，无法进一步判定具体控制子类或短暂擦碰，球心绝对高度指标也未报告。",
+ "conf": 8,
+ "review": {
+  "status": "agreed",
+  "rounds": 1,
+  "reviewers": [
+   {
+    "engine": "claude",
+    "model": "claude-sonnet-5-5",
+    "round": 1,
+    "verdict": "agree",
+    "conf": 7,
+    "summary_zh": "视频 0–1.64 s 与标注一致：球在双掌间被带动，约 1.24 s 起滑落并在 1.60 s 前接近草地；log step 242 服务返回 episode=48、control_steps=163、stages_reached=0/10、球触地，十个节点均未达成，断点 valid_touch_1、现象没保持住、根因控制均有证据支持，时钟与 log 步号有效。",
+    "n_issues": 0,
+    "n_major": 0
+   }
+  ],
+  "open_zh": []
+ },
+ "provenance": {
+  "prompt": "prompts/03_trajectory.md",
+  "prompt_sha256": "2385b35930667330fc4ea181a15ea0818b49cfeae0df1700b8b08c988e72ad8e",
+  "engine": "codex",
+  "model": "gpt-6.1-sol",
+  "effort": "high",
+  "created": "2026-10-31T00:46:59Z",
+  "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/embodied-olympics-fun-juggling-i00-standard--eeqc8ou/03_label"
+   },
+   {
+    "stage": "traj_review",
+    "round": 1,
+    "ok": true,
+    "space": "work/runs/embodied-olympics-fun-juggling-i00-standard--eeqc8ou/04_review_r1"
+   }
+  ],
+  "review_engine": "claude",
+  "review_model": "claude-sonnet-5-5",
+  "inputs": {
+   "run_json_sha256": "d89b8b8757c038defa674313e7d367550e684082909a3f99a1b08c8f3b33766f",
+   "video_sha256": "4e1040735a041317bda072ea545225ecbe9e3a950fb2e2e30b7db4292143d5ce",
+   "states_sha256": null,
+   "video_choice": "HF replay.mp4"
+  }
+ }
+}