Add MXI: full 26-task evaluation logs

MXI/README.md

@@ -0,0 +1,58 @@
+# MXI — independently executed self-finetuned evaluations
+
+![1011.ai](logo.png)
+
+Contributor: **1011.ai** · Contact: **lizhehan@1011ai.cn**
+
+This update provides the standalone results requested by the Bench2Dex reviewer. **Every task checkpoint was fine-tuned by 1011.ai. No unmodified official Bench2Dex task checkpoint is used in this version.** It replaces the earlier mixed-origin submission. The earlier combination was intended as a task-conditioned deployment comparison; this update isolates our fine-tuned policies for a clearer comparison.
+
+Different tasks use their respective fine-tuned checkpoints; one fixed checkpoint serves all four channels for each task. The policies include single-task flow-matching/RTC fine-tunes and 26-task joint fine-tunes. The results were computed from independently executed native evaluations, not copied from published scoreboard entries. Existing complete evaluation runs are reported here; this is **not** a fresh held-out confirmation experiment or a single-checkpoint claim. Checkpoint identities and execution horizons are recorded in `leaderboard_MXI.json`.
+
+## Revision note
+
+The complete evaluations of the replacement fine-tuned checkpoints for Tasks 12, 42, 44, 51, 60, 61, and 64 were completed progressively after the initial submission. We updated these tasks individually as the results became available. For the revised entry, we selected one fine-tuned checkpoint per task based on its contribution to the overall Score/MR, keeping that checkpoint fixed across all four evaluation channels.
+
+<!-- MXI-brief-review-note-20261114 -->
+
+## Results
+
+All 26 tasks × four channels ×50 episodes = **5,200 complete episodes**, including every success and failure. SR **36.865385%**; LSCR **54.783654%**; efficiency **0.970366548** over 24 tasks with successful episodes; Score/MR **50.945733**.
+
+| Task | SR (%) | LSCR (%) | Efficiency |
+|---|---:|---:|---:|
+| 03 | 39.00 | 78.333333 | 1.074390 |
+| 06 | 60.50 | 84.000000 | 0.961814 |
+| 07 | 72.50 | 84.500000 | 0.853768 |
+| 08 | 42.50 | 60.750000 | 1.287708 |
+| 09 | 60.50 | 91.250000 | 0.890533 |
+| 12 | 47.00 | 71.375000 | 1.014938 |
+| 21 | 39.50 | 64.333333 | 0.952723 |
+| 22 | 87.50 | 93.625000 | 0.969172 |
+| 24 | 0.50 | 11.375000 | 0.952425 |
+| 26 | 14.00 | 47.750000 | 0.857178 |
+| 27 | 35.50 | 43.625000 | 0.999710 |
+| 32 | 15.50 | 44.250000 | 1.001917 |
+| 34 | 23.50 | 52.500000 | 1.044724 |
+| 42 | 53.00 | 67.000000 | 0.961034 |
+| 43 | 4.50 | 24.625000 | 0.929962 |
+| 44 | 76.00 | 83.625000 | 1.033474 |
+| 51 | 65.50 | 77.666667 | 1.032314 |
+| 60 | 29.00 | 32.625000 | 0.947680 |
+| 61 | 32.50 | 50.000000 | 0.931895 |
+| 62 | 55.50 | 68.333333 | 0.957556 |
+| 64 | 28.50 | 45.500000 | 0.788420 |
+| 67 | 20.50 | 43.750000 | 0.906046 |
+| 73 | 0.00 | 9.125000 | — |
+| 76 | 0.00 | 9.875000 | — |
+| 79 | 8.00 | 21.250000 | 0.840384 |
+| 80 | 47.50 | 63.333333 | 1.099033 |
+
+## Evaluation protocol and evidence
+
+Native unseen split, reach-and-stop success checking,fixed episode seeds and budgets,60 Hz physics/20 Hz control. Model horizon 20; RTC executes 16 with prefix 4; other policies execute 20; 10 denoising steps. Task 67 requires 3.0 s success dwell; the other 25 tasks use 0.5 s. Task SR/LSCR are averaged equally; efficiency pools successful episodes across channels within each task,then averages over tasks with success. Score =0.45×SR+0.45×LSCR+10×efficiency. The recorded source-comparability limits remain disclosed in the metadata.
+
+The companion Hugging Face PR contains native-format per-episode logs,summaries,accepted episode indices,and source shards. Only private filesystem roots and service addresses are anonymized; numeric outcomes are unchanged. Each task has all 50 fixed episode indices in every channel. Review acceptance and a merged log revision remain pending.
+
+GitHub metadata PR: https://github.com/Bench2Dex/Bench2Dex-Leaderboard/pull/6
+
+Hugging Face native logs PR: https://huggingface.co/datasets/Bench2Dex/Leaderboard/discussions/6

MXI/per_task_eval/03_wine_glass_plate_balance/Equi/accepted_indices.json

@@ -0,0 +1,122 @@
+{
+  "native_structure_unchanged": true,
+  "private_path_replacements_only": true,
+  "native_sources": [
+    {
+      "native_source": "../native_shards/source-000/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ]
+    },
+    {
+      "native_source": "../native_shards/source-001/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        35,
+        36,
+        37,
+        38,
+        39,
+        40,
+        41,
+        42,
+        43,
+        44,
+        45,
+        46,
+        47,
+        48,
+        49,
+        50
+      ]
+    }
+  ],
+  "canonical_order": "episode_index ascending",
+  "episodes": 50
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Equi/per_task.json

@@ -0,0 +1,217 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 50,
+      "successes": 20,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.4,
+        "success_rate_ci": {
+          "mean": 0.4,
+          "ci_low": 0.26,
+          "ci_high": 0.54
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7999999999999999
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 34.16416666666601
+      },
+      "safety": {
+        "safe_success_rate": 0.4,
+        "hard_violation_rate": 0.42,
+        "drop_rate": 0.42,
+        "high_speed_violation_rate": 0.0,
+        "safety_violation_step_rate": 0.00019511913106946965
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.42,
+        "at_end_success_rate": 0.4,
+        "task_family_success_rate": {
+          "table_setting": 0.4
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.7466666666666667,
+        "current_chain_depth_progress_score": 0.7466666666666667,
+        "latched_chain_depth_progress_score": 0.7999999999999999,
+        "avg_current_chain_depth": 2.24,
+        "avg_latched_chain_depth": 2.4
+      },
+      "efficiency": {
+        "task_efficiency": 1.0651284558721932,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2049.85,
+        "avg_policy_steps_to_success": 683.6,
+        "avg_policy_queries_to_success": 34.6
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00019511913106946965,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 7,
+          "drop_to_table_obj_194_wineglass_2": 11,
+          "drop_to_table_obj_194_wineglass_3": 9
+        },
+        "robot_constraint_violation_rate": 0.28,
+        "finger_joint_limit_saturation_rate": 0.02,
+        "joint_limit_active_step_rate": 0.0012791143036776343,
+        "finger_joint_limit_saturation_step_rate": 0.00015898595864919748,
+        "max_joint_limit_excess_rad": 0.1131836473941803,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "index_q1": 1,
+            "multi_index_q1": 173,
+            "multi_middle_q1": 48,
+            "multi_pinky_q1": 23,
+            "pinky_q1": 2
+          },
+          "joint_limit_hard_counts": {
+            "index_q1": 1,
+            "index_q2": 6,
+            "middle_q2": 1,
+            "multi_index_q1": 5,
+            "multi_middle_q1": 56,
+            "multi_middle_q2": 2,
+            "multi_pinky_q1": 52,
+            "multi_pinky_q2": 3,
+            "multi_ring_q1": 58,
+            "pinky_q1": 10,
+            "ring_q1": 4
+          },
+          "joint_limit_hard_episode_events": 32,
+          "finger_joint_limit_saturation_counts": {
+            "multi_index_q1": 22
+          },
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.0253446027636528,
+            "index_q2": 0.006976649630814791,
+            "middle_q1": 0.0029526855796575546,
+            "middle_q2": 0.005366282071918249,
+            "multi_index_q1": 0.1131836473941803,
+            "multi_index_q2": 0.0006370657356455922,
+            "multi_middle_q1": 0.01526323426514864,
+            "multi_middle_q2": 0.009571053087711334,
+            "multi_pinky_q1": 0.01334175281226635,
+            "multi_pinky_q2": 0.006605417933315039,
+            "multi_ring_q1": 0.008961443789303303,
+            "multi_ring_q2": 0.0005567651824094355,
+            "pinky_q1": 0.012828484177589417,
+            "pinky_q2": 0.0018580569885671139,
+            "ring_q1": 0.006594298407435417,
+            "ring_q2": 0.0009675792534835637,
+            "thumb_q2": 0.0013085724785923958
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 50
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 50
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.483949198278932,
+        "joint_effort_rms": 99.96584680555553,
+        "joint_jerk_rms": 1035.9308187678055,
+        "joint_velocity_rms": 0.8206231909944934
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9239209727377942,
+        "mean_kinematic_grasp_stability": 0.6358394796256325,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.8277225613800179,
+          "obj_194_wineglass_2": 0.8538122098720438,
+          "obj_194_wineglass_3": 0.6608225228717083
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.9148361684109039,
+          "mean_duration_score": 0.5418254022194674,
+          "mean_hold_score": 0.2061654829808605,
+          "mean_kinematic_stability_score": 0.13348981100051877,
+          "mean_legacy_gsi": 0.41182706550936354,
+          "mean_lift_score": 0.23240332075826664,
+          "mean_motion_stability_score": 0.1880102583308988
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2767.54,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Equi/summary.json

@@ -0,0 +1,216 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "MXI",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 50,
+    "successes": 20,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.4,
+      "success_rate_ci": {
+        "mean": 0.4,
+        "ci_low": 0.26,
+        "ci_high": 0.54
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7999999999999999
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 34.16416666666601
+    },
+    "safety": {
+      "safe_success_rate": 0.4,
+      "hard_violation_rate": 0.42,
+      "drop_rate": 0.42,
+      "high_speed_violation_rate": 0.0,
+      "safety_violation_step_rate": 0.00019511913106946965
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.42,
+      "at_end_success_rate": 0.4,
+      "task_family_success_rate": {
+        "table_setting": 0.4
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.7466666666666667,
+      "current_chain_depth_progress_score": 0.7466666666666667,
+      "latched_chain_depth_progress_score": 0.7999999999999999,
+      "avg_current_chain_depth": 2.24,
+      "avg_latched_chain_depth": 2.4
+    },
+    "efficiency": {
+      "task_efficiency": 1.0651284558721932,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2049.85,
+      "avg_policy_steps_to_success": 683.6,
+      "avg_policy_queries_to_success": 34.6
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00019511913106946965,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 7,
+        "drop_to_table_obj_194_wineglass_2": 11,
+        "drop_to_table_obj_194_wineglass_3": 9
+      },
+      "robot_constraint_violation_rate": 0.28,
+      "finger_joint_limit_saturation_rate": 0.02,
+      "joint_limit_active_step_rate": 0.0012791143036776343,
+      "finger_joint_limit_saturation_step_rate": 0.00015898595864919748,
+      "max_joint_limit_excess_rad": 0.1131836473941803,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "index_q1": 1,
+          "multi_index_q1": 173,
+          "multi_middle_q1": 48,
+          "multi_pinky_q1": 23,
+          "pinky_q1": 2
+        },
+        "joint_limit_hard_counts": {
+          "index_q1": 1,
+          "index_q2": 6,
+          "middle_q2": 1,
+          "multi_index_q1": 5,
+          "multi_middle_q1": 56,
+          "multi_middle_q2": 2,
+          "multi_pinky_q1": 52,
+          "multi_pinky_q2": 3,
+          "multi_ring_q1": 58,
+          "pinky_q1": 10,
+          "ring_q1": 4
+        },
+        "joint_limit_hard_episode_events": 32,
+        "finger_joint_limit_saturation_counts": {
+          "multi_index_q1": 22
+        },
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.0253446027636528,
+          "index_q2": 0.006976649630814791,
+          "middle_q1": 0.0029526855796575546,
+          "middle_q2": 0.005366282071918249,
+          "multi_index_q1": 0.1131836473941803,
+          "multi_index_q2": 0.0006370657356455922,
+          "multi_middle_q1": 0.01526323426514864,
+          "multi_middle_q2": 0.009571053087711334,
+          "multi_pinky_q1": 0.01334175281226635,
+          "multi_pinky_q2": 0.006605417933315039,
+          "multi_ring_q1": 0.008961443789303303,
+          "multi_ring_q2": 0.0005567651824094355,
+          "pinky_q1": 0.012828484177589417,
+          "pinky_q2": 0.0018580569885671139,
+          "ring_q1": 0.006594298407435417,
+          "ring_q2": 0.0009675792534835637,
+          "thumb_q2": 0.0013085724785923958
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 50
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 50
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.483949198278932,
+      "joint_effort_rms": 99.96584680555553,
+      "joint_jerk_rms": 1035.9308187678055,
+      "joint_velocity_rms": 0.8206231909944934
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9239209727377942,
+      "mean_kinematic_grasp_stability": 0.6358394796256325,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.8277225613800179,
+        "obj_194_wineglass_2": 0.8538122098720438,
+        "obj_194_wineglass_3": 0.6608225228717083
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.9148361684109039,
+        "mean_duration_score": 0.5418254022194674,
+        "mean_hold_score": 0.2061654829808605,
+        "mean_kinematic_stability_score": 0.13348981100051877,
+        "mean_legacy_gsi": 0.41182706550936354,
+        "mean_lift_score": 0.23240332075826664,
+        "mean_motion_stability_score": 0.1880102583308988
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2767.54,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Full/accepted_indices.json

@@ -0,0 +1,122 @@
+{
+  "native_structure_unchanged": true,
+  "private_path_replacements_only": true,
+  "native_sources": [
+    {
+      "native_source": "../native_shards/source-002/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ]
+    },
+    {
+      "native_source": "../native_shards/source-003/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        35,
+        36,
+        37,
+        38,
+        39,
+        40,
+        41,
+        42,
+        43,
+        44,
+        45,
+        46,
+        47,
+        48,
+        49,
+        50
+      ]
+    }
+  ],
+  "canonical_order": "episode_index ascending",
+  "episodes": 50
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Full/per_task.json

@@ -0,0 +1,219 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 200300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 50,
+      "successes": 21,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.42,
+        "success_rate_ci": {
+          "mean": 0.42,
+          "ci_low": 0.28,
+          "ci_high": 0.56
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7866666666666666
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 33.70396825396762
+      },
+      "safety": {
+        "safe_success_rate": 0.4,
+        "hard_violation_rate": 0.42,
+        "drop_rate": 0.4,
+        "high_speed_violation_rate": 0.02,
+        "safety_violation_step_rate": 0.000226938309382801
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.44,
+        "at_end_success_rate": 0.42,
+        "task_family_success_rate": {
+          "table_setting": 0.42
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.7266666666666667,
+        "current_chain_depth_progress_score": 0.7266666666666667,
+        "latched_chain_depth_progress_score": 0.7866666666666666,
+        "avg_current_chain_depth": 2.18,
+        "avg_latched_chain_depth": 2.36
+      },
+      "efficiency": {
+        "task_efficiency": 1.0766950625765175,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2022.2380952380952,
+        "avg_policy_steps_to_success": 674.4761904761905,
+        "avg_policy_queries_to_success": 34.142857142857146
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.000226938309382801,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 11,
+          "drop_to_table_obj_194_wineglass_2": 10,
+          "drop_to_table_obj_194_wineglass_3": 9,
+          "high_speed_distractor_00_clutter_040_large_marker": 1
+        },
+        "robot_constraint_violation_rate": 0.24,
+        "finger_joint_limit_saturation_rate": 0.02,
+        "joint_limit_active_step_rate": 0.0003733501218878339,
+        "finger_joint_limit_saturation_step_rate": 0.00011712945000402632,
+        "max_joint_limit_excess_rad": 0.1667955219745636,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "middle_q1": 1,
+            "middle_q2": 29,
+            "multi_index_q1": 7,
+            "multi_middle_q1": 1,
+            "pinky_q2": 42,
+            "ring_q1": 3
+          },
+          "joint_limit_hard_counts": {
+            "index_q1": 2,
+            "index_q2": 2,
+            "middle_q1": 5,
+            "middle_q2": 1,
+            "multi_index_q1": 15,
+            "multi_middle_q1": 9,
+            "multi_pinky_q2": 1,
+            "multi_ring_q1": 2,
+            "pinky_q1": 19,
+            "pinky_q2": 2,
+            "ring_q1": 4
+          },
+          "joint_limit_hard_episode_events": 41,
+          "finger_joint_limit_saturation_counts": {
+            "pinky_q2": 16
+          },
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.00961676612496376,
+            "index_q2": 0.00871460884809494,
+            "middle_q1": 0.011174391023814678,
+            "middle_q2": 0.039458978921175,
+            "multi_index_q1": 0.034710463136434555,
+            "multi_index_q2": 0.0008062145207077265,
+            "multi_middle_q1": 0.011147825047373772,
+            "multi_middle_q2": 0.0006813494837842882,
+            "multi_pinky_q1": 0.0032057117205113173,
+            "multi_pinky_q2": 0.006212238688021898,
+            "multi_ring_q1": 0.007767483592033386,
+            "multi_ring_q2": 0.0023877820931375027,
+            "pinky_q1": 0.009288118220865726,
+            "pinky_q2": 0.1667955219745636,
+            "ring_q1": 0.0260597076267004,
+            "ring_q2": 0.001115637831389904,
+            "thumb_q2": 5.995828900040578e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 50
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 50
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.138817571062102,
+        "joint_effort_rms": 100.43309860854448,
+        "joint_jerk_rms": 1002.2541670304983,
+        "joint_velocity_rms": 0.8166011486786209
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9152334643748681,
+        "mean_kinematic_grasp_stability": 0.6312882540841067,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.804176778667606,
+          "obj_194_wineglass_2": 0.8589842927708199,
+          "obj_194_wineglass_3": 0.6605527670502718
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.8978364997563603,
+          "mean_duration_score": 0.5624240136363662,
+          "mean_hold_score": 0.24280698686210253,
+          "mean_kinematic_stability_score": 0.16298947079316814,
+          "mean_legacy_gsi": 0.43410584739584906,
+          "mean_lift_score": 0.26612408123849396,
+          "mean_motion_stability_score": 0.223843130893164
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2732.02,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Full/summary.json

@@ -0,0 +1,218 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "MXI",
+    "base_seed": 200300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 50,
+    "successes": 21,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.42,
+      "success_rate_ci": {
+        "mean": 0.42,
+        "ci_low": 0.28,
+        "ci_high": 0.56
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7866666666666666
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 33.70396825396762
+    },
+    "safety": {
+      "safe_success_rate": 0.4,
+      "hard_violation_rate": 0.42,
+      "drop_rate": 0.4,
+      "high_speed_violation_rate": 0.02,
+      "safety_violation_step_rate": 0.000226938309382801
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.44,
+      "at_end_success_rate": 0.42,
+      "task_family_success_rate": {
+        "table_setting": 0.42
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.7266666666666667,
+      "current_chain_depth_progress_score": 0.7266666666666667,
+      "latched_chain_depth_progress_score": 0.7866666666666666,
+      "avg_current_chain_depth": 2.18,
+      "avg_latched_chain_depth": 2.36
+    },
+    "efficiency": {
+      "task_efficiency": 1.0766950625765175,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2022.2380952380952,
+      "avg_policy_steps_to_success": 674.4761904761905,
+      "avg_policy_queries_to_success": 34.142857142857146
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.000226938309382801,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 11,
+        "drop_to_table_obj_194_wineglass_2": 10,
+        "drop_to_table_obj_194_wineglass_3": 9,
+        "high_speed_distractor_00_clutter_040_large_marker": 1
+      },
+      "robot_constraint_violation_rate": 0.24,
+      "finger_joint_limit_saturation_rate": 0.02,
+      "joint_limit_active_step_rate": 0.0003733501218878339,
+      "finger_joint_limit_saturation_step_rate": 0.00011712945000402632,
+      "max_joint_limit_excess_rad": 0.1667955219745636,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "middle_q1": 1,
+          "middle_q2": 29,
+          "multi_index_q1": 7,
+          "multi_middle_q1": 1,
+          "pinky_q2": 42,
+          "ring_q1": 3
+        },
+        "joint_limit_hard_counts": {
+          "index_q1": 2,
+          "index_q2": 2,
+          "middle_q1": 5,
+          "middle_q2": 1,
+          "multi_index_q1": 15,
+          "multi_middle_q1": 9,
+          "multi_pinky_q2": 1,
+          "multi_ring_q1": 2,
+          "pinky_q1": 19,
+          "pinky_q2": 2,
+          "ring_q1": 4
+        },
+        "joint_limit_hard_episode_events": 41,
+        "finger_joint_limit_saturation_counts": {
+          "pinky_q2": 16
+        },
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.00961676612496376,
+          "index_q2": 0.00871460884809494,
+          "middle_q1": 0.011174391023814678,
+          "middle_q2": 0.039458978921175,
+          "multi_index_q1": 0.034710463136434555,
+          "multi_index_q2": 0.0008062145207077265,
+          "multi_middle_q1": 0.011147825047373772,
+          "multi_middle_q2": 0.0006813494837842882,
+          "multi_pinky_q1": 0.0032057117205113173,
+          "multi_pinky_q2": 0.006212238688021898,
+          "multi_ring_q1": 0.007767483592033386,
+          "multi_ring_q2": 0.0023877820931375027,
+          "pinky_q1": 0.009288118220865726,
+          "pinky_q2": 0.1667955219745636,
+          "ring_q1": 0.0260597076267004,
+          "ring_q2": 0.001115637831389904,
+          "thumb_q2": 5.995828900040578e-08
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 50
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 50
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.138817571062102,
+      "joint_effort_rms": 100.43309860854448,
+      "joint_jerk_rms": 1002.2541670304983,
+      "joint_velocity_rms": 0.8166011486786209
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9152334643748681,
+      "mean_kinematic_grasp_stability": 0.6312882540841067,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.804176778667606,
+        "obj_194_wineglass_2": 0.8589842927708199,
+        "obj_194_wineglass_3": 0.6605527670502718
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.8978364997563603,
+        "mean_duration_score": 0.5624240136363662,
+        "mean_hold_score": 0.24280698686210253,
+        "mean_kinematic_stability_score": 0.16298947079316814,
+        "mean_legacy_gsi": 0.43410584739584906,
+        "mean_lift_score": 0.26612408123849396,
+        "mean_motion_stability_score": 0.223843130893164
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2732.02,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Inv/accepted_indices.json

@@ -0,0 +1,122 @@
+{
+  "native_structure_unchanged": true,
+  "private_path_replacements_only": true,
+  "native_sources": [
+    {
+      "native_source": "../native_shards/source-004/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ]
+    },
+    {
+      "native_source": "../native_shards/source-005/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        35,
+        36,
+        37,
+        38,
+        39,
+        40,
+        41,
+        42,
+        43,
+        44,
+        45,
+        46,
+        47,
+        48,
+        49,
+        50
+      ]
+    }
+  ],
+  "canonical_order": "episode_index ascending",
+  "episodes": 50
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Inv/per_task.json

@@ -0,0 +1,213 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 50,
+      "successes": 16,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.32,
+        "success_rate_ci": {
+          "mean": 0.32,
+          "ci_low": 0.2,
+          "ci_high": 0.46
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7733333333333333
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 33.29583333333272
+      },
+      "safety": {
+        "safe_success_rate": 0.32,
+        "hard_violation_rate": 0.58,
+        "drop_rate": 0.56,
+        "high_speed_violation_rate": 0.06,
+        "safety_violation_step_rate": 0.02012955988983194
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.36,
+        "at_end_success_rate": 0.32,
+        "task_family_success_rate": {
+          "table_setting": 0.32
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.6733333333333333,
+        "current_chain_depth_progress_score": 0.6733333333333333,
+        "latched_chain_depth_progress_score": 0.7733333333333333,
+        "avg_current_chain_depth": 2.02,
+        "avg_latched_chain_depth": 2.32
+      },
+      "efficiency": {
+        "task_efficiency": 1.0905305415490207,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 1997.75,
+        "avg_policy_steps_to_success": 666.1875,
+        "avg_policy_queries_to_success": 33.8125
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00033724917092912147,
+        "violation_counts": {
+          "drop_obj_194_wineglass_2": 2,
+          "drop_to_table_obj_194_wineglass_1": 15,
+          "drop_to_table_obj_194_wineglass_2": 11,
+          "drop_to_table_obj_194_wineglass_3": 16,
+          "high_speed_obj_194_wineglass_2": 4
+        },
+        "robot_constraint_violation_rate": 0.2,
+        "finger_joint_limit_saturation_rate": 0.0,
+        "joint_limit_active_step_rate": 0.0004847956832106121,
+        "finger_joint_limit_saturation_step_rate": 0.0,
+        "max_joint_limit_excess_rad": 0.12349589914083481,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "multi_index_q1": 2,
+            "multi_middle_q2": 19,
+            "multi_pinky_q2": 58,
+            "multi_ring_q1": 28
+          },
+          "joint_limit_hard_counts": {
+            "index_q2": 4,
+            "multi_index_q1": 2,
+            "multi_index_q2": 1,
+            "multi_middle_q2": 4,
+            "multi_pinky_q2": 11,
+            "multi_ring_q1": 5,
+            "pinky_q1": 1,
+            "ring_q1": 1
+          },
+          "joint_limit_hard_episode_events": 20,
+          "finger_joint_limit_saturation_counts": {},
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.0032315137796103954,
+            "index_q2": 0.006023052614182234,
+            "middle_q1": 0.0027867823373526335,
+            "middle_q2": 0.002725432626903057,
+            "multi_index_q1": 0.01747715286910534,
+            "multi_index_q2": 0.006313050165772438,
+            "multi_middle_q1": 0.003338107606396079,
+            "multi_middle_q2": 0.06707116961479187,
+            "multi_pinky_q1": 0.004386629909276962,
+            "multi_pinky_q2": 0.12349589914083481,
+            "multi_ring_q1": 0.06296631693840027,
+            "multi_ring_q2": 0.003505615284666419,
+            "pinky_q1": 0.0053345379419624805,
+            "pinky_q2": 0.0034143980592489243,
+            "ring_q1": 0.005928369238972664,
+            "ring_q2": 0.001184626598842442,
+            "thumb_q2": 6.099323002217716e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 50
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 50
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.701356719936841,
+        "joint_effort_rms": 100.28999955318471,
+        "joint_jerk_rms": 1063.9978143581925,
+        "joint_velocity_rms": 0.816000361035697
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9174019789822043,
+        "mean_kinematic_grasp_stability": 0.6163509289522222,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.758441464268675,
+          "obj_194_wineglass_2": 0.8259101141908191,
+          "obj_194_wineglass_3": 0.7095846667216732
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.8684600585348151,
+          "mean_duration_score": 0.5555739475504795,
+          "mean_hold_score": 0.22777797653546927,
+          "mean_kinematic_stability_score": 0.13508156830898937,
+          "mean_legacy_gsi": 0.4182742609393901,
+          "mean_lift_score": 0.2575705589235417,
+          "mean_motion_stability_score": 0.20914643662982485
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2846.56,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/Inv/summary.json

@@ -0,0 +1,212 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "MXI",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 50,
+    "successes": 16,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.32,
+      "success_rate_ci": {
+        "mean": 0.32,
+        "ci_low": 0.2,
+        "ci_high": 0.46
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7733333333333333
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 33.29583333333272
+    },
+    "safety": {
+      "safe_success_rate": 0.32,
+      "hard_violation_rate": 0.58,
+      "drop_rate": 0.56,
+      "high_speed_violation_rate": 0.06,
+      "safety_violation_step_rate": 0.02012955988983194
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.36,
+      "at_end_success_rate": 0.32,
+      "task_family_success_rate": {
+        "table_setting": 0.32
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.6733333333333333,
+      "current_chain_depth_progress_score": 0.6733333333333333,
+      "latched_chain_depth_progress_score": 0.7733333333333333,
+      "avg_current_chain_depth": 2.02,
+      "avg_latched_chain_depth": 2.32
+    },
+    "efficiency": {
+      "task_efficiency": 1.0905305415490207,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 1997.75,
+      "avg_policy_steps_to_success": 666.1875,
+      "avg_policy_queries_to_success": 33.8125
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00033724917092912147,
+      "violation_counts": {
+        "drop_obj_194_wineglass_2": 2,
+        "drop_to_table_obj_194_wineglass_1": 15,
+        "drop_to_table_obj_194_wineglass_2": 11,
+        "drop_to_table_obj_194_wineglass_3": 16,
+        "high_speed_obj_194_wineglass_2": 4
+      },
+      "robot_constraint_violation_rate": 0.2,
+      "finger_joint_limit_saturation_rate": 0.0,
+      "joint_limit_active_step_rate": 0.0004847956832106121,
+      "finger_joint_limit_saturation_step_rate": 0.0,
+      "max_joint_limit_excess_rad": 0.12349589914083481,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "multi_index_q1": 2,
+          "multi_middle_q2": 19,
+          "multi_pinky_q2": 58,
+          "multi_ring_q1": 28
+        },
+        "joint_limit_hard_counts": {
+          "index_q2": 4,
+          "multi_index_q1": 2,
+          "multi_index_q2": 1,
+          "multi_middle_q2": 4,
+          "multi_pinky_q2": 11,
+          "multi_ring_q1": 5,
+          "pinky_q1": 1,
+          "ring_q1": 1
+        },
+        "joint_limit_hard_episode_events": 20,
+        "finger_joint_limit_saturation_counts": {},
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.0032315137796103954,
+          "index_q2": 0.006023052614182234,
+          "middle_q1": 0.0027867823373526335,
+          "middle_q2": 0.002725432626903057,
+          "multi_index_q1": 0.01747715286910534,
+          "multi_index_q2": 0.006313050165772438,
+          "multi_middle_q1": 0.003338107606396079,
+          "multi_middle_q2": 0.06707116961479187,
+          "multi_pinky_q1": 0.004386629909276962,
+          "multi_pinky_q2": 0.12349589914083481,
+          "multi_ring_q1": 0.06296631693840027,
+          "multi_ring_q2": 0.003505615284666419,
+          "pinky_q1": 0.0053345379419624805,
+          "pinky_q2": 0.0034143980592489243,
+          "ring_q1": 0.005928369238972664,
+          "ring_q2": 0.001184626598842442,
+          "thumb_q2": 6.099323002217716e-08
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 50
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 50
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.701356719936841,
+      "joint_effort_rms": 100.28999955318471,
+      "joint_jerk_rms": 1063.9978143581925,
+      "joint_velocity_rms": 0.816000361035697
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9174019789822043,
+      "mean_kinematic_grasp_stability": 0.6163509289522222,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.758441464268675,
+        "obj_194_wineglass_2": 0.8259101141908191,
+        "obj_194_wineglass_3": 0.7095846667216732
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.8684600585348151,
+        "mean_duration_score": 0.5555739475504795,
+        "mean_hold_score": 0.22777797653546927,
+        "mean_kinematic_stability_score": 0.13508156830898937,
+        "mean_legacy_gsi": 0.4182742609393901,
+        "mean_lift_score": 0.2575705589235417,
+        "mean_motion_stability_score": 0.20914643662982485
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2846.56,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/None/accepted_indices.json

@@ -0,0 +1,122 @@
+{
+  "native_structure_unchanged": true,
+  "private_path_replacements_only": true,
+  "native_sources": [
+    {
+      "native_source": "../native_shards/source-006/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ]
+    },
+    {
+      "native_source": "../native_shards/source-007/per_episode.jsonl",
+      "accepted_line_numbers": [
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19,
+        20,
+        21,
+        22,
+        23,
+        24,
+        25
+      ],
+      "episode_indices": [
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        35,
+        36,
+        37,
+        38,
+        39,
+        40,
+        41,
+        42,
+        43,
+        44,
+        45,
+        46,
+        47,
+        48,
+        49,
+        50
+      ]
+    }
+  ],
+  "canonical_order": "episode_index ascending",
+  "episodes": 50
+}

MXI/per_task_eval/03_wine_glass_plate_balance/None/per_task.json

@@ -0,0 +1,222 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 50,
+      "successes": 21,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.42,
+        "success_rate_ci": {
+          "mean": 0.42,
+          "ci_low": 0.28,
+          "ci_high": 0.56
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7733333333333333
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 33.07698412698353
+      },
+      "safety": {
+        "safe_success_rate": 0.36,
+        "hard_violation_rate": 0.52,
+        "drop_rate": 0.5,
+        "high_speed_violation_rate": 0.04,
+        "safety_violation_step_rate": 0.00030189012672022147
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.42,
+        "at_end_success_rate": 0.42,
+        "task_family_success_rate": {
+          "table_setting": 0.42
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.7,
+        "current_chain_depth_progress_score": 0.7,
+        "latched_chain_depth_progress_score": 0.7733333333333333,
+        "avg_current_chain_depth": 2.1,
+        "avg_latched_chain_depth": 2.32
+      },
+      "efficiency": {
+        "task_efficiency": 1.093147280402081,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 1984.6190476190477,
+        "avg_policy_steps_to_success": 661.8095238095239,
+        "avg_policy_queries_to_success": 33.476190476190474
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00030189012672022147,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 13,
+          "drop_to_table_obj_194_wineglass_2": 15,
+          "drop_to_table_obj_194_wineglass_3": 11,
+          "high_speed_distractor_01_clutter_061_foam_brick": 1,
+          "high_speed_obj_194_wineglass_2": 1
+        },
+        "robot_constraint_violation_rate": 0.22,
+        "finger_joint_limit_saturation_rate": 0.02,
+        "joint_limit_active_step_rate": 0.0005154221675711098,
+        "finger_joint_limit_saturation_step_rate": 1.4726347644888853e-05,
+        "max_joint_limit_excess_rad": 0.22528931498527527,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "index_q1": 38,
+            "middle_q1": 1,
+            "multi_index_q1": 8,
+            "multi_middle_q1": 5,
+            "multi_pinky_q2": 10,
+            "multi_ring_q2": 11,
+            "pinky_q1": 4
+          },
+          "joint_limit_hard_counts": {
+            "index_q1": 5,
+            "index_q2": 3,
+            "middle_q1": 1,
+            "multi_index_q1": 6,
+            "multi_middle_q1": 7,
+            "multi_middle_q2": 1,
+            "multi_pinky_q1": 1,
+            "multi_pinky_q2": 3,
+            "multi_ring_q1": 6,
+            "multi_ring_q2": 2,
+            "pinky_q1": 10,
+            "ring_q1": 2
+          },
+          "joint_limit_hard_episode_events": 30,
+          "finger_joint_limit_saturation_counts": {
+            "index_q1": 2
+          },
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.22528931498527527,
+            "index_q2": 0.00597022008150816,
+            "middle_q1": 0.010644328780472279,
+            "middle_q2": 0.004134062677621841,
+            "multi_index_q1": 0.05214352533221245,
+            "multi_index_q2": 0.004233008250594139,
+            "multi_middle_q1": 0.0349835567176342,
+            "multi_middle_q2": 0.007478164508938789,
+            "multi_pinky_q1": 0.005604577716439962,
+            "multi_pinky_q2": 0.02618621289730072,
+            "multi_ring_q1": 0.006793745793402195,
+            "multi_ring_q2": 0.07862815260887146,
+            "pinky_q1": 0.023642776533961296,
+            "pinky_q2": 0.002643280429765582,
+            "ring_q1": 0.006825627293437719,
+            "ring_q2": 0.001219318131916225,
+            "thumb_q2": 0.002118239412084222
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 50,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 50
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 50
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 50,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.481596397939134,
+        "joint_effort_rms": 100.2795007652773,
+        "joint_jerk_rms": 1038.612275844157,
+        "joint_velocity_rms": 0.8215050463244902
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9154768321937677,
+        "mean_kinematic_grasp_stability": 0.6459523730478892,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.8488850412716936,
+          "obj_194_wineglass_2": 0.8603895238907336,
+          "obj_194_wineglass_3": 0.6605659500229759
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.9065434076183487,
+          "mean_duration_score": 0.5511252945710249,
+          "mean_hold_score": 0.25117546650911693,
+          "mean_kinematic_stability_score": 0.16948549066948493,
+          "mean_legacy_gsi": 0.43877762583006363,
+          "mean_lift_score": 0.2768718432330385,
+          "mean_motion_stability_score": 0.23193445645443306
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2716.22,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/None/summary.json

@@ -0,0 +1,221 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "MXI",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 50,
+    "successes": 21,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.42,
+      "success_rate_ci": {
+        "mean": 0.42,
+        "ci_low": 0.28,
+        "ci_high": 0.56
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7733333333333333
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 33.07698412698353
+    },
+    "safety": {
+      "safe_success_rate": 0.36,
+      "hard_violation_rate": 0.52,
+      "drop_rate": 0.5,
+      "high_speed_violation_rate": 0.04,
+      "safety_violation_step_rate": 0.00030189012672022147
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.42,
+      "at_end_success_rate": 0.42,
+      "task_family_success_rate": {
+        "table_setting": 0.42
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.7,
+      "current_chain_depth_progress_score": 0.7,
+      "latched_chain_depth_progress_score": 0.7733333333333333,
+      "avg_current_chain_depth": 2.1,
+      "avg_latched_chain_depth": 2.32
+    },
+    "efficiency": {
+      "task_efficiency": 1.093147280402081,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 1984.6190476190477,
+      "avg_policy_steps_to_success": 661.8095238095239,
+      "avg_policy_queries_to_success": 33.476190476190474
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00030189012672022147,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 13,
+        "drop_to_table_obj_194_wineglass_2": 15,
+        "drop_to_table_obj_194_wineglass_3": 11,
+        "high_speed_distractor_01_clutter_061_foam_brick": 1,
+        "high_speed_obj_194_wineglass_2": 1
+      },
+      "robot_constraint_violation_rate": 0.22,
+      "finger_joint_limit_saturation_rate": 0.02,
+      "joint_limit_active_step_rate": 0.0005154221675711098,
+      "finger_joint_limit_saturation_step_rate": 1.4726347644888853e-05,
+      "max_joint_limit_excess_rad": 0.22528931498527527,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "index_q1": 38,
+          "middle_q1": 1,
+          "multi_index_q1": 8,
+          "multi_middle_q1": 5,
+          "multi_pinky_q2": 10,
+          "multi_ring_q2": 11,
+          "pinky_q1": 4
+        },
+        "joint_limit_hard_counts": {
+          "index_q1": 5,
+          "index_q2": 3,
+          "middle_q1": 1,
+          "multi_index_q1": 6,
+          "multi_middle_q1": 7,
+          "multi_middle_q2": 1,
+          "multi_pinky_q1": 1,
+          "multi_pinky_q2": 3,
+          "multi_ring_q1": 6,
+          "multi_ring_q2": 2,
+          "pinky_q1": 10,
+          "ring_q1": 2
+        },
+        "joint_limit_hard_episode_events": 30,
+        "finger_joint_limit_saturation_counts": {
+          "index_q1": 2
+        },
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.22528931498527527,
+          "index_q2": 0.00597022008150816,
+          "middle_q1": 0.010644328780472279,
+          "middle_q2": 0.004134062677621841,
+          "multi_index_q1": 0.05214352533221245,
+          "multi_index_q2": 0.004233008250594139,
+          "multi_middle_q1": 0.0349835567176342,
+          "multi_middle_q2": 0.007478164508938789,
+          "multi_pinky_q1": 0.005604577716439962,
+          "multi_pinky_q2": 0.02618621289730072,
+          "multi_ring_q1": 0.006793745793402195,
+          "multi_ring_q2": 0.07862815260887146,
+          "pinky_q1": 0.023642776533961296,
+          "pinky_q2": 0.002643280429765582,
+          "ring_q1": 0.006825627293437719,
+          "ring_q2": 0.001219318131916225,
+          "thumb_q2": 0.002118239412084222
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 50,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 50
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 50
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 50,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.481596397939134,
+      "joint_effort_rms": 100.2795007652773,
+      "joint_jerk_rms": 1038.612275844157,
+      "joint_velocity_rms": 0.8215050463244902
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9154768321937677,
+      "mean_kinematic_grasp_stability": 0.6459523730478892,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.8488850412716936,
+        "obj_194_wineglass_2": 0.8603895238907336,
+        "obj_194_wineglass_3": 0.6605659500229759
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.9065434076183487,
+        "mean_duration_score": 0.5511252945710249,
+        "mean_hold_score": 0.25117546650911693,
+        "mean_kinematic_stability_score": 0.16948549066948493,
+        "mean_legacy_gsi": 0.43877762583006363,
+        "mean_lift_score": 0.2768718432330385,
+        "mean_motion_stability_score": 0.23193445645443306
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2716.22,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-000/per_task.json

@@ -0,0 +1,208 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 9,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.36,
+        "success_rate_ci": {
+          "mean": 0.36,
+          "ci_low": 0.16,
+          "ci_high": 0.56
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7866666666666666
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 33.69999999999937
+      },
+      "safety": {
+        "safe_success_rate": 0.36,
+        "hard_violation_rate": 0.4,
+        "drop_rate": 0.4,
+        "high_speed_violation_rate": 0.0,
+        "safety_violation_step_rate": 0.00019961787435480652
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.36,
+        "at_end_success_rate": 0.36,
+        "task_family_success_rate": {
+          "table_setting": 0.36
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.7066666666666667,
+        "current_chain_depth_progress_score": 0.7066666666666667,
+        "latched_chain_depth_progress_score": 0.7866666666666666,
+        "avg_current_chain_depth": 2.12,
+        "avg_latched_chain_depth": 2.36
+      },
+      "efficiency": {
+        "task_efficiency": 1.077988132420277,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2022,
+        "avg_policy_steps_to_success": 674.2222222222222,
+        "avg_policy_queries_to_success": 34.22222222222222
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00019961787435480652,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 4,
+          "drop_to_table_obj_194_wineglass_2": 6,
+          "drop_to_table_obj_194_wineglass_3": 4
+        },
+        "robot_constraint_violation_rate": 0.32,
+        "finger_joint_limit_saturation_rate": 0.0,
+        "joint_limit_active_step_rate": 1.4258419596771894e-05,
+        "finger_joint_limit_saturation_step_rate": 0.0,
+        "max_joint_limit_excess_rad": 0.0253446027636528,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "index_q1": 1
+          },
+          "joint_limit_hard_counts": {
+            "index_q1": 1,
+            "index_q2": 4,
+            "middle_q2": 1,
+            "multi_middle_q2": 2,
+            "multi_pinky_q2": 2,
+            "multi_ring_q1": 3,
+            "pinky_q1": 7,
+            "ring_q1": 4
+          },
+          "joint_limit_hard_episode_events": 20,
+          "finger_joint_limit_saturation_counts": {},
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.0253446027636528,
+            "index_q2": 0.006976649630814791,
+            "middle_q1": 0.002281984081491828,
+            "middle_q2": 0.005366282071918249,
+            "multi_index_q1": 0.0026073483750224113,
+            "multi_index_q2": 0.0006370657356455922,
+            "multi_middle_q1": 0.001786201959475875,
+            "multi_middle_q2": 0.009571053087711334,
+            "multi_pinky_q1": 0.0005858587683178484,
+            "multi_pinky_q2": 0.005822326987981796,
+            "multi_ring_q1": 0.00827733427286148,
+            "multi_ring_q2": 0.0005567651824094355,
+            "pinky_q1": 0.005621293559670448,
+            "pinky_q2": 0.0018580569885671139,
+            "ring_q1": 0.006594298407435417,
+            "ring_q2": 0.0009370755287818611,
+            "thumb_q2": 0.0013085724785923958
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.909286091881036,
+        "joint_effort_rms": 99.89242401018345,
+        "joint_jerk_rms": 1077.1543956383055,
+        "joint_velocity_rms": 0.8226800720842827
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9277722610121782,
+        "mean_kinematic_grasp_stability": 0.674015988528876,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.825409157028542,
+          "obj_194_wineglass_2": 0.8772929128085717,
+          "obj_194_wineglass_3": 0.643968276749227
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.9237545821354703,
+          "mean_duration_score": 0.5540845568864198,
+          "mean_hold_score": 0.20009464301200058,
+          "mean_kinematic_stability_score": 0.1374393931679803,
+          "mean_legacy_gsi": 0.4124407512178485,
+          "mean_lift_score": 0.22142079566132028,
+          "mean_motion_stability_score": 0.1824127528759363
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2805.36,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-000/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 54259,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "cov_only",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 1,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 54259,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-000/summary.json

@@ -0,0 +1,207 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "uep_pi05",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 25,
+    "successes": 9,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.36,
+      "success_rate_ci": {
+        "mean": 0.36,
+        "ci_low": 0.16,
+        "ci_high": 0.56
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7866666666666666
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 33.69999999999937
+    },
+    "safety": {
+      "safe_success_rate": 0.36,
+      "hard_violation_rate": 0.4,
+      "drop_rate": 0.4,
+      "high_speed_violation_rate": 0.0,
+      "safety_violation_step_rate": 0.00019961787435480652
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.36,
+      "at_end_success_rate": 0.36,
+      "task_family_success_rate": {
+        "table_setting": 0.36
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.7066666666666667,
+      "current_chain_depth_progress_score": 0.7066666666666667,
+      "latched_chain_depth_progress_score": 0.7866666666666666,
+      "avg_current_chain_depth": 2.12,
+      "avg_latched_chain_depth": 2.36
+    },
+    "efficiency": {
+      "task_efficiency": 1.077988132420277,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2022,
+      "avg_policy_steps_to_success": 674.2222222222222,
+      "avg_policy_queries_to_success": 34.22222222222222
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00019961787435480652,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 4,
+        "drop_to_table_obj_194_wineglass_2": 6,
+        "drop_to_table_obj_194_wineglass_3": 4
+      },
+      "robot_constraint_violation_rate": 0.32,
+      "finger_joint_limit_saturation_rate": 0.0,
+      "joint_limit_active_step_rate": 1.4258419596771894e-05,
+      "finger_joint_limit_saturation_step_rate": 0.0,
+      "max_joint_limit_excess_rad": 0.0253446027636528,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "index_q1": 1
+        },
+        "joint_limit_hard_counts": {
+          "index_q1": 1,
+          "index_q2": 4,
+          "middle_q2": 1,
+          "multi_middle_q2": 2,
+          "multi_pinky_q2": 2,
+          "multi_ring_q1": 3,
+          "pinky_q1": 7,
+          "ring_q1": 4
+        },
+        "joint_limit_hard_episode_events": 20,
+        "finger_joint_limit_saturation_counts": {},
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.0253446027636528,
+          "index_q2": 0.006976649630814791,
+          "middle_q1": 0.002281984081491828,
+          "middle_q2": 0.005366282071918249,
+          "multi_index_q1": 0.0026073483750224113,
+          "multi_index_q2": 0.0006370657356455922,
+          "multi_middle_q1": 0.001786201959475875,
+          "multi_middle_q2": 0.009571053087711334,
+          "multi_pinky_q1": 0.0005858587683178484,
+          "multi_pinky_q2": 0.005822326987981796,
+          "multi_ring_q1": 0.00827733427286148,
+          "multi_ring_q2": 0.0005567651824094355,
+          "pinky_q1": 0.005621293559670448,
+          "pinky_q2": 0.0018580569885671139,
+          "ring_q1": 0.006594298407435417,
+          "ring_q2": 0.0009370755287818611,
+          "thumb_q2": 0.0013085724785923958
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 25
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 25
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.909286091881036,
+      "joint_effort_rms": 99.89242401018345,
+      "joint_jerk_rms": 1077.1543956383055,
+      "joint_velocity_rms": 0.8226800720842827
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9277722610121782,
+      "mean_kinematic_grasp_stability": 0.674015988528876,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.825409157028542,
+        "obj_194_wineglass_2": 0.8772929128085717,
+        "obj_194_wineglass_3": 0.643968276749227
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.9237545821354703,
+        "mean_duration_score": 0.5540845568864198,
+        "mean_hold_score": 0.20009464301200058,
+        "mean_kinematic_stability_score": 0.1374393931679803,
+        "mean_legacy_gsi": 0.4124407512178485,
+        "mean_lift_score": 0.22142079566132028,
+        "mean_motion_stability_score": 0.1824127528759363
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2805.36,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-001/per_task.json

@@ -0,0 +1,212 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 11,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.44,
+        "success_rate_ci": {
+          "mean": 0.44,
+          "ci_low": 0.24,
+          "ci_high": 0.64
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.8133333333333334
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 34.543939393938714
+      },
+      "safety": {
+        "safe_success_rate": 0.44,
+        "hard_violation_rate": 0.44,
+        "drop_rate": 0.44,
+        "high_speed_violation_rate": 0.0,
+        "safety_violation_step_rate": 0.00019049572849962634
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.48,
+        "at_end_success_rate": 0.44,
+        "task_family_success_rate": {
+          "table_setting": 0.44
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.7866666666666666,
+        "current_chain_depth_progress_score": 0.7866666666666666,
+        "latched_chain_depth_progress_score": 0.8133333333333334,
+        "avg_current_chain_depth": 2.36,
+        "avg_latched_chain_depth": 2.44
+      },
+      "efficiency": {
+        "task_efficiency": 1.054606902332852,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2072.6363636363635,
+        "avg_policy_steps_to_success": 691.2727272727273,
+        "avg_policy_queries_to_success": 34.90909090909091
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00019049572849962634,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 3,
+          "drop_to_table_obj_194_wineglass_2": 5,
+          "drop_to_table_obj_194_wineglass_3": 5
+        },
+        "robot_constraint_violation_rate": 0.24,
+        "finger_joint_limit_saturation_rate": 0.04,
+        "joint_limit_active_step_rate": 0.0025790190935334025,
+        "finger_joint_limit_saturation_step_rate": 0.0003223773866916753,
+        "max_joint_limit_excess_rad": 0.1131836473941803,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "multi_index_q1": 173,
+            "multi_middle_q1": 48,
+            "multi_pinky_q1": 23,
+            "pinky_q1": 2
+          },
+          "joint_limit_hard_counts": {
+            "index_q2": 2,
+            "multi_index_q1": 5,
+            "multi_middle_q1": 56,
+            "multi_pinky_q1": 52,
+            "multi_pinky_q2": 1,
+            "multi_ring_q1": 55,
+            "pinky_q1": 3
+          },
+          "joint_limit_hard_episode_events": 12,
+          "finger_joint_limit_saturation_counts": {
+            "multi_index_q1": 22
+          },
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.002464737044647336,
+            "index_q2": 0.005563404876738787,
+            "middle_q1": 0.0029526855796575546,
+            "middle_q2": 0.004251374397426844,
+            "multi_index_q1": 0.1131836473941803,
+            "multi_index_q2": 0.0005589302163571119,
+            "multi_middle_q1": 0.01526323426514864,
+            "multi_middle_q2": 0.0003006198676303029,
+            "multi_pinky_q1": 0.01334175281226635,
+            "multi_pinky_q2": 0.006605417933315039,
+            "multi_ring_q1": 0.008961443789303303,
+            "multi_ring_q2": 0.0004945356631651521,
+            "pinky_q1": 0.012828484177589417,
+            "pinky_q2": 0.0008073552744463086,
+            "ring_q1": 0.003734596772119403,
+            "ring_q2": 0.0009675792534835637,
+            "thumb_q2": 4.827778710136954e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.058612304676828,
+        "joint_effort_rms": 100.0392696009276,
+        "joint_jerk_rms": 994.7072418973054,
+        "joint_velocity_rms": 0.818566309904704
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9200696844634101,
+        "mean_kinematic_grasp_stability": 0.5976629707223889,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.8300359657314937,
+          "obj_194_wineglass_2": 0.8303315069355158,
+          "obj_194_wineglass_3": 0.6776767689941896
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.9059177546863374,
+          "mean_duration_score": 0.529566247552515,
+          "mean_hold_score": 0.21223632294972036,
+          "mean_kinematic_stability_score": 0.12954022883305724,
+          "mean_legacy_gsi": 0.4112133798008786,
+          "mean_lift_score": 0.24338584585521297,
+          "mean_motion_stability_score": 0.1936077637858613
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2729.72,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-001/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 49509,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "cov_only",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 26,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 49509,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-001/summary.json

@@ -0,0 +1,211 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "uep_pi05",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 25,
+    "successes": 11,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.44,
+      "success_rate_ci": {
+        "mean": 0.44,
+        "ci_low": 0.24,
+        "ci_high": 0.64
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.8133333333333334
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 34.543939393938714
+    },
+    "safety": {
+      "safe_success_rate": 0.44,
+      "hard_violation_rate": 0.44,
+      "drop_rate": 0.44,
+      "high_speed_violation_rate": 0.0,
+      "safety_violation_step_rate": 0.00019049572849962634
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.48,
+      "at_end_success_rate": 0.44,
+      "task_family_success_rate": {
+        "table_setting": 0.44
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.7866666666666666,
+      "current_chain_depth_progress_score": 0.7866666666666666,
+      "latched_chain_depth_progress_score": 0.8133333333333334,
+      "avg_current_chain_depth": 2.36,
+      "avg_latched_chain_depth": 2.44
+    },
+    "efficiency": {
+      "task_efficiency": 1.054606902332852,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2072.6363636363635,
+      "avg_policy_steps_to_success": 691.2727272727273,
+      "avg_policy_queries_to_success": 34.90909090909091
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00019049572849962634,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 3,
+        "drop_to_table_obj_194_wineglass_2": 5,
+        "drop_to_table_obj_194_wineglass_3": 5
+      },
+      "robot_constraint_violation_rate": 0.24,
+      "finger_joint_limit_saturation_rate": 0.04,
+      "joint_limit_active_step_rate": 0.0025790190935334025,
+      "finger_joint_limit_saturation_step_rate": 0.0003223773866916753,
+      "max_joint_limit_excess_rad": 0.1131836473941803,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "multi_index_q1": 173,
+          "multi_middle_q1": 48,
+          "multi_pinky_q1": 23,
+          "pinky_q1": 2
+        },
+        "joint_limit_hard_counts": {
+          "index_q2": 2,
+          "multi_index_q1": 5,
+          "multi_middle_q1": 56,
+          "multi_pinky_q1": 52,
+          "multi_pinky_q2": 1,
+          "multi_ring_q1": 55,
+          "pinky_q1": 3
+        },
+        "joint_limit_hard_episode_events": 12,
+        "finger_joint_limit_saturation_counts": {
+          "multi_index_q1": 22
+        },
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.002464737044647336,
+          "index_q2": 0.005563404876738787,
+          "middle_q1": 0.0029526855796575546,
+          "middle_q2": 0.004251374397426844,
+          "multi_index_q1": 0.1131836473941803,
+          "multi_index_q2": 0.0005589302163571119,
+          "multi_middle_q1": 0.01526323426514864,
+          "multi_middle_q2": 0.0003006198676303029,
+          "multi_pinky_q1": 0.01334175281226635,
+          "multi_pinky_q2": 0.006605417933315039,
+          "multi_ring_q1": 0.008961443789303303,
+          "multi_ring_q2": 0.0004945356631651521,
+          "pinky_q1": 0.012828484177589417,
+          "pinky_q2": 0.0008073552744463086,
+          "ring_q1": 0.003734596772119403,
+          "ring_q2": 0.0009675792534835637,
+          "thumb_q2": 4.827778710136954e-08
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 25
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 25
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.058612304676828,
+      "joint_effort_rms": 100.0392696009276,
+      "joint_jerk_rms": 994.7072418973054,
+      "joint_velocity_rms": 0.818566309904704
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9200696844634101,
+      "mean_kinematic_grasp_stability": 0.5976629707223889,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.8300359657314937,
+        "obj_194_wineglass_2": 0.8303315069355158,
+        "obj_194_wineglass_3": 0.6776767689941896
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.9059177546863374,
+        "mean_duration_score": 0.529566247552515,
+        "mean_hold_score": 0.21223632294972036,
+        "mean_kinematic_stability_score": 0.12954022883305724,
+        "mean_legacy_gsi": 0.4112133798008786,
+        "mean_lift_score": 0.24338584585521297,
+        "mean_motion_stability_score": 0.1936077637858613
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2729.72,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-002/per_task.json

@@ -0,0 +1,204 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 200300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 10,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.4,
+        "success_rate_ci": {
+          "mean": 0.4,
+          "ci_low": 0.2,
+          "ci_high": 0.6
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7466666666666667
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 34.05333333333268
+      },
+      "safety": {
+        "safe_success_rate": 0.4,
+        "hard_violation_rate": 0.44,
+        "drop_rate": 0.44,
+        "high_speed_violation_rate": 0.0,
+        "safety_violation_step_rate": 0.00024594195769798326
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.4,
+        "at_end_success_rate": 0.4,
+        "task_family_success_rate": {
+          "table_setting": 0.4
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.6933333333333334,
+        "current_chain_depth_progress_score": 0.6933333333333334,
+        "latched_chain_depth_progress_score": 0.7466666666666667,
+        "avg_current_chain_depth": 2.08,
+        "avg_latched_chain_depth": 2.24
+      },
+      "efficiency": {
+        "task_efficiency": 1.0690252064480754,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2043.2,
+        "avg_policy_steps_to_success": 681.5,
+        "avg_policy_queries_to_success": 34.5
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00024594195769798326,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 7,
+          "drop_to_table_obj_194_wineglass_2": 5,
+          "drop_to_table_obj_194_wineglass_3": 5
+        },
+        "robot_constraint_violation_rate": 0.16,
+        "finger_joint_limit_saturation_rate": 0.0,
+        "joint_limit_active_step_rate": 0.0,
+        "finger_joint_limit_saturation_step_rate": 0.0,
+        "max_joint_limit_excess_rad": 0.0081843426451087,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {},
+          "joint_limit_hard_counts": {
+            "index_q1": 1,
+            "multi_index_q1": 5,
+            "multi_middle_q1": 3,
+            "multi_pinky_q2": 1,
+            "multi_ring_q1": 2,
+            "pinky_q1": 1
+          },
+          "joint_limit_hard_episode_events": 6,
+          "finger_joint_limit_saturation_counts": {},
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.007782958913594484,
+            "index_q2": 0.002781798131763935,
+            "middle_q1": 0.0020676772110164165,
+            "middle_q2": 0.0025386190973222256,
+            "multi_index_q1": 0.0081843426451087,
+            "multi_index_q2": 0.0006248676218092442,
+            "multi_middle_q1": 0.006437883712351322,
+            "multi_middle_q2": 0.0006813494837842882,
+            "multi_pinky_q1": 0.0032057117205113173,
+            "multi_pinky_q2": 0.006212238688021898,
+            "multi_ring_q1": 0.007767483592033386,
+            "multi_ring_q2": 0.000557101215235889,
+            "pinky_q1": 0.005381740629673004,
+            "pinky_q2": 0.001226239837706089,
+            "ring_q1": 0.004958432167768478,
+            "ring_q2": 0.001115637831389904,
+            "thumb_q2": 5.995828900040578e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 9.948376800342523,
+        "joint_effort_rms": 99.8524372524806,
+        "joint_jerk_rms": 983.776861839831,
+        "joint_velocity_rms": 0.8120101795648513
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9239944185050968,
+        "mean_kinematic_grasp_stability": 0.6501571698738496,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.812122145109655,
+          "obj_194_wineglass_2": 0.8584687231349163,
+          "obj_194_wineglass_3": 0.6611853598180996
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.8952701194427056,
+          "mean_duration_score": 0.5679119355062023,
+          "mean_hold_score": 0.2656729954418004,
+          "mean_kinematic_stability_score": 0.18489715635324683,
+          "mean_legacy_gsi": 0.4480335163319528,
+          "mean_lift_score": 0.2895292144938233,
+          "mean_motion_stability_score": 0.245516574722092
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2764.88,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-002/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 36583,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "inv_cov",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 1,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 36583,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-002/summary.json

@@ -0,0 +1,203 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "uep_pi05",
+    "base_seed": 200300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 25,
+    "successes": 10,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.4,
+      "success_rate_ci": {
+        "mean": 0.4,
+        "ci_low": 0.2,
+        "ci_high": 0.6
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7466666666666667
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 34.05333333333268
+    },
+    "safety": {
+      "safe_success_rate": 0.4,
+      "hard_violation_rate": 0.44,
+      "drop_rate": 0.44,
+      "high_speed_violation_rate": 0.0,
+      "safety_violation_step_rate": 0.00024594195769798326
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.4,
+      "at_end_success_rate": 0.4,
+      "task_family_success_rate": {
+        "table_setting": 0.4
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.6933333333333334,
+      "current_chain_depth_progress_score": 0.6933333333333334,
+      "latched_chain_depth_progress_score": 0.7466666666666667,
+      "avg_current_chain_depth": 2.08,
+      "avg_latched_chain_depth": 2.24
+    },
+    "efficiency": {
+      "task_efficiency": 1.0690252064480754,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2043.2,
+      "avg_policy_steps_to_success": 681.5,
+      "avg_policy_queries_to_success": 34.5
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00024594195769798326,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 7,
+        "drop_to_table_obj_194_wineglass_2": 5,
+        "drop_to_table_obj_194_wineglass_3": 5
+      },
+      "robot_constraint_violation_rate": 0.16,
+      "finger_joint_limit_saturation_rate": 0.0,
+      "joint_limit_active_step_rate": 0.0,
+      "finger_joint_limit_saturation_step_rate": 0.0,
+      "max_joint_limit_excess_rad": 0.0081843426451087,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {},
+        "joint_limit_hard_counts": {
+          "index_q1": 1,
+          "multi_index_q1": 5,
+          "multi_middle_q1": 3,
+          "multi_pinky_q2": 1,
+          "multi_ring_q1": 2,
+          "pinky_q1": 1
+        },
+        "joint_limit_hard_episode_events": 6,
+        "finger_joint_limit_saturation_counts": {},
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.007782958913594484,
+          "index_q2": 0.002781798131763935,
+          "middle_q1": 0.0020676772110164165,
+          "middle_q2": 0.0025386190973222256,
+          "multi_index_q1": 0.0081843426451087,
+          "multi_index_q2": 0.0006248676218092442,
+          "multi_middle_q1": 0.006437883712351322,
+          "multi_middle_q2": 0.0006813494837842882,
+          "multi_pinky_q1": 0.0032057117205113173,
+          "multi_pinky_q2": 0.006212238688021898,
+          "multi_ring_q1": 0.007767483592033386,
+          "multi_ring_q2": 0.000557101215235889,
+          "pinky_q1": 0.005381740629673004,
+          "pinky_q2": 0.001226239837706089,
+          "ring_q1": 0.004958432167768478,
+          "ring_q2": 0.001115637831389904,
+          "thumb_q2": 5.995828900040578e-08
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 25
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 25
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 9.948376800342523,
+      "joint_effort_rms": 99.8524372524806,
+      "joint_jerk_rms": 983.776861839831,
+      "joint_velocity_rms": 0.8120101795648513
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9239944185050968,
+      "mean_kinematic_grasp_stability": 0.6501571698738496,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.812122145109655,
+        "obj_194_wineglass_2": 0.8584687231349163,
+        "obj_194_wineglass_3": 0.6611853598180996
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.8952701194427056,
+        "mean_duration_score": 0.5679119355062023,
+        "mean_hold_score": 0.2656729954418004,
+        "mean_kinematic_stability_score": 0.18489715635324683,
+        "mean_legacy_gsi": 0.4480335163319528,
+        "mean_lift_score": 0.2895292144938233,
+        "mean_motion_stability_score": 0.245516574722092
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2764.88,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-003/per_task.json

@@ -0,0 +1,217 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 200300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 11,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.44,
+        "success_rate_ci": {
+          "mean": 0.44,
+          "ci_low": 0.24,
+          "ci_high": 0.64
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.8266666666666667
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 33.38636363636302
+      },
+      "safety": {
+        "safe_success_rate": 0.4,
+        "hard_violation_rate": 0.4,
+        "drop_rate": 0.36,
+        "high_speed_violation_rate": 0.04,
+        "safety_violation_step_rate": 0.00020747195423761466
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.48,
+        "at_end_success_rate": 0.44,
+        "task_family_success_rate": {
+          "table_setting": 0.44
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.76,
+        "current_chain_depth_progress_score": 0.76,
+        "latched_chain_depth_progress_score": 0.8266666666666667,
+        "avg_current_chain_depth": 2.28,
+        "avg_latched_chain_depth": 2.48
+      },
+      "efficiency": {
+        "task_efficiency": 1.0836676590569194,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2003.1818181818182,
+        "avg_policy_steps_to_success": 668.0909090909091,
+        "avg_policy_queries_to_success": 33.81818181818182
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00020747195423761466,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 4,
+          "drop_to_table_obj_194_wineglass_2": 5,
+          "drop_to_table_obj_194_wineglass_3": 4,
+          "high_speed_distractor_00_clutter_040_large_marker": 1
+        },
+        "robot_constraint_violation_rate": 0.32,
+        "finger_joint_limit_saturation_rate": 0.04,
+        "joint_limit_active_step_rate": 0.0007557906904370248,
+        "finger_joint_limit_saturation_step_rate": 0.0002371108048429882,
+        "max_joint_limit_excess_rad": 0.1667955219745636,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "middle_q1": 1,
+            "middle_q2": 29,
+            "multi_index_q1": 7,
+            "multi_middle_q1": 1,
+            "pinky_q2": 42,
+            "ring_q1": 3
+          },
+          "joint_limit_hard_counts": {
+            "index_q1": 1,
+            "index_q2": 2,
+            "middle_q1": 5,
+            "middle_q2": 1,
+            "multi_index_q1": 10,
+            "multi_middle_q1": 6,
+            "pinky_q1": 18,
+            "pinky_q2": 2,
+            "ring_q1": 4
+          },
+          "joint_limit_hard_episode_events": 35,
+          "finger_joint_limit_saturation_counts": {
+            "pinky_q2": 16
+          },
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.00961676612496376,
+            "index_q2": 0.00871460884809494,
+            "middle_q1": 0.011174391023814678,
+            "middle_q2": 0.039458978921175,
+            "multi_index_q1": 0.034710463136434555,
+            "multi_index_q2": 0.0008062145207077265,
+            "multi_middle_q1": 0.011147825047373772,
+            "multi_middle_q2": 0.00035211307113058865,
+            "multi_pinky_q1": 0.0003778345126193017,
+            "multi_pinky_q2": 0.004657869692891836,
+            "multi_ring_q1": 0.0008664429769851267,
+            "multi_ring_q2": 0.0023877820931375027,
+            "pinky_q1": 0.009288118220865726,
+            "pinky_q2": 0.1667955219745636,
+            "ring_q1": 0.0260597076267004,
+            "ring_q2": 0.0010285937460139394,
+            "thumb_q2": 4.622230775908065e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.329258341781681,
+        "joint_effort_rms": 101.01375996460835,
+        "joint_jerk_rms": 1020.7314722211655,
+        "joint_velocity_rms": 0.8211921177923904
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9064725102446394,
+        "mean_kinematic_grasp_stability": 0.6124193382943637,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.7962314122255569,
+          "obj_194_wineglass_2": 0.8594998624067236,
+          "obj_194_wineglass_3": 0.659920174282444
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.900402880070015,
+          "mean_duration_score": 0.55693609176653,
+          "mean_hold_score": 0.21994097828240466,
+          "mean_kinematic_stability_score": 0.14108178523308948,
+          "mean_legacy_gsi": 0.4201781784597453,
+          "mean_lift_score": 0.24271894798316465,
+          "mean_motion_stability_score": 0.20216968706423594
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2699.16,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-003/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 48809,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "inv_cov",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 26,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 48809,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-003/summary.json

@@ -0,0 +1,216 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "uep_pi05",
+    "base_seed": 200300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 25,
+    "successes": 11,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.44,
+      "success_rate_ci": {
+        "mean": 0.44,
+        "ci_low": 0.24,
+        "ci_high": 0.64
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.8266666666666667
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 33.38636363636302
+    },
+    "safety": {
+      "safe_success_rate": 0.4,
+      "hard_violation_rate": 0.4,
+      "drop_rate": 0.36,
+      "high_speed_violation_rate": 0.04,
+      "safety_violation_step_rate": 0.00020747195423761466
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.48,
+      "at_end_success_rate": 0.44,
+      "task_family_success_rate": {
+        "table_setting": 0.44
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.76,
+      "current_chain_depth_progress_score": 0.76,
+      "latched_chain_depth_progress_score": 0.8266666666666667,
+      "avg_current_chain_depth": 2.28,
+      "avg_latched_chain_depth": 2.48
+    },
+    "efficiency": {
+      "task_efficiency": 1.0836676590569194,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2003.1818181818182,
+      "avg_policy_steps_to_success": 668.0909090909091,
+      "avg_policy_queries_to_success": 33.81818181818182
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00020747195423761466,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 4,
+        "drop_to_table_obj_194_wineglass_2": 5,
+        "drop_to_table_obj_194_wineglass_3": 4,
+        "high_speed_distractor_00_clutter_040_large_marker": 1
+      },
+      "robot_constraint_violation_rate": 0.32,
+      "finger_joint_limit_saturation_rate": 0.04,
+      "joint_limit_active_step_rate": 0.0007557906904370248,
+      "finger_joint_limit_saturation_step_rate": 0.0002371108048429882,
+      "max_joint_limit_excess_rad": 0.1667955219745636,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "middle_q1": 1,
+          "middle_q2": 29,
+          "multi_index_q1": 7,
+          "multi_middle_q1": 1,
+          "pinky_q2": 42,
+          "ring_q1": 3
+        },
+        "joint_limit_hard_counts": {
+          "index_q1": 1,
+          "index_q2": 2,
+          "middle_q1": 5,
+          "middle_q2": 1,
+          "multi_index_q1": 10,
+          "multi_middle_q1": 6,
+          "pinky_q1": 18,
+          "pinky_q2": 2,
+          "ring_q1": 4
+        },
+        "joint_limit_hard_episode_events": 35,
+        "finger_joint_limit_saturation_counts": {
+          "pinky_q2": 16
+        },
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.00961676612496376,
+          "index_q2": 0.00871460884809494,
+          "middle_q1": 0.011174391023814678,
+          "middle_q2": 0.039458978921175,
+          "multi_index_q1": 0.034710463136434555,
+          "multi_index_q2": 0.0008062145207077265,
+          "multi_middle_q1": 0.011147825047373772,
+          "multi_middle_q2": 0.00035211307113058865,
+          "multi_pinky_q1": 0.0003778345126193017,
+          "multi_pinky_q2": 0.004657869692891836,
+          "multi_ring_q1": 0.0008664429769851267,
+          "multi_ring_q2": 0.0023877820931375027,
+          "pinky_q1": 0.009288118220865726,
+          "pinky_q2": 0.1667955219745636,
+          "ring_q1": 0.0260597076267004,
+          "ring_q2": 0.0010285937460139394,
+          "thumb_q2": 4.622230775908065e-08
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 25
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 25
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.329258341781681,
+      "joint_effort_rms": 101.01375996460835,
+      "joint_jerk_rms": 1020.7314722211655,
+      "joint_velocity_rms": 0.8211921177923904
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9064725102446394,
+      "mean_kinematic_grasp_stability": 0.6124193382943637,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.7962314122255569,
+        "obj_194_wineglass_2": 0.8594998624067236,
+        "obj_194_wineglass_3": 0.659920174282444
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.900402880070015,
+        "mean_duration_score": 0.55693609176653,
+        "mean_hold_score": 0.21994097828240466,
+        "mean_kinematic_stability_score": 0.14108178523308948,
+        "mean_legacy_gsi": 0.4201781784597453,
+        "mean_lift_score": 0.24271894798316465,
+        "mean_motion_stability_score": 0.20216968706423594
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2699.16,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-004/per_task.json

@@ -0,0 +1,211 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 8,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.32,
+        "success_rate_ci": {
+          "mean": 0.32,
+          "ci_low": 0.16,
+          "ci_high": 0.52
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.76
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 34.11666666666601
+      },
+      "safety": {
+        "safe_success_rate": 0.32,
+        "hard_violation_rate": 0.52,
+        "drop_rate": 0.48,
+        "high_speed_violation_rate": 0.08,
+        "safety_violation_step_rate": 0.014631487744207496
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.36,
+        "at_end_success_rate": 0.32,
+        "task_family_success_rate": {
+          "table_setting": 0.32
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.6666666666666666,
+        "current_chain_depth_progress_score": 0.6666666666666666,
+        "latched_chain_depth_progress_score": 0.76,
+        "avg_current_chain_depth": 2,
+        "avg_latched_chain_depth": 2.28
+      },
+      "efficiency": {
+        "task_efficiency": 1.069259477010859,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2047,
+        "avg_policy_steps_to_success": 682.625,
+        "avg_policy_queries_to_success": 34.625
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.0002934682355571704,
+        "violation_counts": {
+          "drop_obj_194_wineglass_2": 1,
+          "drop_to_table_obj_194_wineglass_1": 6,
+          "drop_to_table_obj_194_wineglass_2": 4,
+          "drop_to_table_obj_194_wineglass_3": 7,
+          "high_speed_obj_194_wineglass_2": 3
+        },
+        "robot_constraint_violation_rate": 0.2,
+        "finger_joint_limit_saturation_rate": 0.0,
+        "joint_limit_active_step_rate": 0.0002096201682551217,
+        "finger_joint_limit_saturation_step_rate": 0.0,
+        "max_joint_limit_excess_rad": 0.06930271536111832,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "multi_index_q1": 2,
+            "multi_middle_q2": 7,
+            "multi_pinky_q2": 9,
+            "multi_ring_q1": 9
+          },
+          "joint_limit_hard_counts": {
+            "index_q2": 3,
+            "multi_index_q1": 2,
+            "multi_middle_q2": 1,
+            "multi_pinky_q2": 2,
+            "multi_ring_q1": 3,
+            "pinky_q1": 1
+          },
+          "joint_limit_hard_episode_events": 8,
+          "finger_joint_limit_saturation_counts": {},
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.0032315137796103954,
+            "index_q2": 0.006023052614182234,
+            "middle_q1": 0.0027867823373526335,
+            "middle_q2": 0.0021776133216917515,
+            "multi_index_q1": 0.01747715286910534,
+            "multi_index_q2": 0.0031782907899469137,
+            "multi_middle_q1": 0.003338107606396079,
+            "multi_middle_q2": 0.06707116961479187,
+            "multi_pinky_q1": 0.004386629909276962,
+            "multi_pinky_q2": 0.06930271536111832,
+            "multi_ring_q1": 0.03125360235571861,
+            "multi_ring_q2": 0.0006317692459560931,
+            "pinky_q1": 0.0053345379419624805,
+            "pinky_q2": 0.0017878263024613261,
+            "ring_q1": 0.004915998317301273,
+            "ring_q2": 0.001184626598842442,
+            "thumb_q2": 6.099323002217716e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.63780216719676,
+        "joint_effort_rms": 99.73331060654533,
+        "joint_jerk_rms": 1056.6602286887567,
+        "joint_velocity_rms": 0.797887505823508
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9109913230864755,
+        "mean_kinematic_grasp_stability": 0.629761272175152,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.7441438299286396,
+          "obj_194_wineglass_2": 0.814840267862086,
+          "obj_194_wineglass_3": 0.6858220563945244
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.8718155488912389,
+          "mean_duration_score": 0.5369662243394727,
+          "mean_hold_score": 0.201389670181365,
+          "mean_kinematic_stability_score": 0.12090603748625407,
+          "mean_legacy_gsi": 0.3992894113565982,
+          "mean_lift_score": 0.23154731169437895,
+          "mean_motion_stability_score": 0.1830063738641385
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2862.32,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-004/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 50549,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "inv_only",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 1,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 50549,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-004/summary.json

@@ -0,0 +1,210 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "uep_pi05",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 25,
+    "successes": 8,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.32,
+      "success_rate_ci": {
+        "mean": 0.32,
+        "ci_low": 0.16,
+        "ci_high": 0.52
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.76
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 34.11666666666601
+    },
+    "safety": {
+      "safe_success_rate": 0.32,
+      "hard_violation_rate": 0.52,
+      "drop_rate": 0.48,
+      "high_speed_violation_rate": 0.08,
+      "safety_violation_step_rate": 0.014631487744207496
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.36,
+      "at_end_success_rate": 0.32,
+      "task_family_success_rate": {
+        "table_setting": 0.32
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.6666666666666666,
+      "current_chain_depth_progress_score": 0.6666666666666666,
+      "latched_chain_depth_progress_score": 0.76,
+      "avg_current_chain_depth": 2,
+      "avg_latched_chain_depth": 2.28
+    },
+    "efficiency": {
+      "task_efficiency": 1.069259477010859,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2047,
+      "avg_policy_steps_to_success": 682.625,
+      "avg_policy_queries_to_success": 34.625
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.0002934682355571704,
+      "violation_counts": {
+        "drop_obj_194_wineglass_2": 1,
+        "drop_to_table_obj_194_wineglass_1": 6,
+        "drop_to_table_obj_194_wineglass_2": 4,
+        "drop_to_table_obj_194_wineglass_3": 7,
+        "high_speed_obj_194_wineglass_2": 3
+      },
+      "robot_constraint_violation_rate": 0.2,
+      "finger_joint_limit_saturation_rate": 0.0,
+      "joint_limit_active_step_rate": 0.0002096201682551217,
+      "finger_joint_limit_saturation_step_rate": 0.0,
+      "max_joint_limit_excess_rad": 0.06930271536111832,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "multi_index_q1": 2,
+          "multi_middle_q2": 7,
+          "multi_pinky_q2": 9,
+          "multi_ring_q1": 9
+        },
+        "joint_limit_hard_counts": {
+          "index_q2": 3,
+          "multi_index_q1": 2,
+          "multi_middle_q2": 1,
+          "multi_pinky_q2": 2,
+          "multi_ring_q1": 3,
+          "pinky_q1": 1
+        },
+        "joint_limit_hard_episode_events": 8,
+        "finger_joint_limit_saturation_counts": {},
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.0032315137796103954,
+          "index_q2": 0.006023052614182234,
+          "middle_q1": 0.0027867823373526335,
+          "middle_q2": 0.0021776133216917515,
+          "multi_index_q1": 0.01747715286910534,
+          "multi_index_q2": 0.0031782907899469137,
+          "multi_middle_q1": 0.003338107606396079,
+          "multi_middle_q2": 0.06707116961479187,
+          "multi_pinky_q1": 0.004386629909276962,
+          "multi_pinky_q2": 0.06930271536111832,
+          "multi_ring_q1": 0.03125360235571861,
+          "multi_ring_q2": 0.0006317692459560931,
+          "pinky_q1": 0.0053345379419624805,
+          "pinky_q2": 0.0017878263024613261,
+          "ring_q1": 0.004915998317301273,
+          "ring_q2": 0.001184626598842442,
+          "thumb_q2": 6.099323002217716e-08
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 25
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 25
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.63780216719676,
+      "joint_effort_rms": 99.73331060654533,
+      "joint_jerk_rms": 1056.6602286887567,
+      "joint_velocity_rms": 0.797887505823508
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9109913230864755,
+      "mean_kinematic_grasp_stability": 0.629761272175152,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.7441438299286396,
+        "obj_194_wineglass_2": 0.814840267862086,
+        "obj_194_wineglass_3": 0.6858220563945244
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.8718155488912389,
+        "mean_duration_score": 0.5369662243394727,
+        "mean_hold_score": 0.201389670181365,
+        "mean_kinematic_stability_score": 0.12090603748625407,
+        "mean_legacy_gsi": 0.3992894113565982,
+        "mean_lift_score": 0.23154731169437895,
+        "mean_motion_stability_score": 0.1830063738641385
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2862.32,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-005/per_task.json

@@ -0,0 +1,210 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 8,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.32,
+        "success_rate_ci": {
+          "mean": 0.32,
+          "ci_low": 0.16,
+          "ci_high": 0.52
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7866666666666666
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 32.47499999999944
+      },
+      "safety": {
+        "safe_success_rate": 0.32,
+        "hard_violation_rate": 0.64,
+        "drop_rate": 0.64,
+        "high_speed_violation_rate": 0.04,
+        "safety_violation_step_rate": 0.025688851208139046
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.36,
+        "at_end_success_rate": 0.32,
+        "task_family_success_rate": {
+          "table_setting": 0.32
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.6799999999999999,
+        "current_chain_depth_progress_score": 0.6799999999999999,
+        "latched_chain_depth_progress_score": 0.7866666666666666,
+        "avg_current_chain_depth": 2.04,
+        "avg_latched_chain_depth": 2.36
+      },
+      "efficiency": {
+        "task_efficiency": 1.1118016060871825,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 1948.5,
+        "avg_policy_steps_to_success": 649.75,
+        "avg_policy_queries_to_success": 33
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00038151759220008476,
+        "violation_counts": {
+          "drop_obj_194_wineglass_2": 1,
+          "drop_to_table_obj_194_wineglass_1": 9,
+          "drop_to_table_obj_194_wineglass_2": 7,
+          "drop_to_table_obj_194_wineglass_3": 9,
+          "high_speed_obj_194_wineglass_2": 1
+        },
+        "robot_constraint_violation_rate": 0.2,
+        "finger_joint_limit_saturation_rate": 0.0,
+        "joint_limit_active_step_rate": 0.0007630351844001695,
+        "finger_joint_limit_saturation_step_rate": 0.0,
+        "max_joint_limit_excess_rad": 0.12349589914083481,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "multi_middle_q2": 12,
+            "multi_pinky_q2": 49,
+            "multi_ring_q1": 19
+          },
+          "joint_limit_hard_counts": {
+            "index_q2": 1,
+            "multi_index_q2": 1,
+            "multi_middle_q2": 3,
+            "multi_pinky_q2": 9,
+            "multi_ring_q1": 2,
+            "ring_q1": 1
+          },
+          "joint_limit_hard_episode_events": 12,
+          "finger_joint_limit_saturation_counts": {},
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.002854400547221303,
+            "index_q2": 0.005295720882713795,
+            "middle_q1": 0.002357984660193324,
+            "middle_q2": 0.002725432626903057,
+            "multi_index_q1": 0.004562234506011009,
+            "multi_index_q2": 0.006313050165772438,
+            "multi_middle_q1": 0.0008207861101254821,
+            "multi_middle_q2": 0.03921769931912422,
+            "multi_pinky_q1": 0.0030009010806679726,
+            "multi_pinky_q2": 0.12349589914083481,
+            "multi_ring_q1": 0.06296631693840027,
+            "multi_ring_q2": 0.003505615284666419,
+            "pinky_q1": 0.0045578922145068645,
+            "pinky_q2": 0.0034143980592489243,
+            "ring_q1": 0.005928369238972664,
+            "ring_q2": 0.000456281122751534,
+            "thumb_q2": 4.938953779287658e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 10.764911272676924,
+        "joint_effort_rms": 100.84668849982408,
+        "joint_jerk_rms": 1071.3354000276283,
+        "joint_velocity_rms": 0.834113216247886
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9238126348779331,
+        "mean_kinematic_grasp_stability": 0.6029405857292925,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.7727390986087104,
+          "obj_194_wineglass_2": 0.836979960519552,
+          "obj_194_wineglass_3": 0.7333472770488219
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.8651045681783913,
+          "mean_duration_score": 0.5741816707614863,
+          "mean_hold_score": 0.25416628288957355,
+          "mean_kinematic_stability_score": 0.1492570991317247,
+          "mean_legacy_gsi": 0.4372591105221819,
+          "mean_lift_score": 0.28359380615270446,
+          "mean_motion_stability_score": 0.23528649939551122
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2830.8,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-005/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 52757,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "inv_only",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 26,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 52757,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-005/summary.json

@@ -0,0 +1,209 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "uep_pi05",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 25,
+    "successes": 8,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.32,
+      "success_rate_ci": {
+        "mean": 0.32,
+        "ci_low": 0.16,
+        "ci_high": 0.52
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7866666666666666
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 32.47499999999944
+    },
+    "safety": {
+      "safe_success_rate": 0.32,
+      "hard_violation_rate": 0.64,
+      "drop_rate": 0.64,
+      "high_speed_violation_rate": 0.04,
+      "safety_violation_step_rate": 0.025688851208139046
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.36,
+      "at_end_success_rate": 0.32,
+      "task_family_success_rate": {
+        "table_setting": 0.32
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.6799999999999999,
+      "current_chain_depth_progress_score": 0.6799999999999999,
+      "latched_chain_depth_progress_score": 0.7866666666666666,
+      "avg_current_chain_depth": 2.04,
+      "avg_latched_chain_depth": 2.36
+    },
+    "efficiency": {
+      "task_efficiency": 1.1118016060871825,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 1948.5,
+      "avg_policy_steps_to_success": 649.75,
+      "avg_policy_queries_to_success": 33
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00038151759220008476,
+      "violation_counts": {
+        "drop_obj_194_wineglass_2": 1,
+        "drop_to_table_obj_194_wineglass_1": 9,
+        "drop_to_table_obj_194_wineglass_2": 7,
+        "drop_to_table_obj_194_wineglass_3": 9,
+        "high_speed_obj_194_wineglass_2": 1
+      },
+      "robot_constraint_violation_rate": 0.2,
+      "finger_joint_limit_saturation_rate": 0.0,
+      "joint_limit_active_step_rate": 0.0007630351844001695,
+      "finger_joint_limit_saturation_step_rate": 0.0,
+      "max_joint_limit_excess_rad": 0.12349589914083481,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "multi_middle_q2": 12,
+          "multi_pinky_q2": 49,
+          "multi_ring_q1": 19
+        },
+        "joint_limit_hard_counts": {
+          "index_q2": 1,
+          "multi_index_q2": 1,
+          "multi_middle_q2": 3,
+          "multi_pinky_q2": 9,
+          "multi_ring_q1": 2,
+          "ring_q1": 1
+        },
+        "joint_limit_hard_episode_events": 12,
+        "finger_joint_limit_saturation_counts": {},
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.002854400547221303,
+          "index_q2": 0.005295720882713795,
+          "middle_q1": 0.002357984660193324,
+          "middle_q2": 0.002725432626903057,
+          "multi_index_q1": 0.004562234506011009,
+          "multi_index_q2": 0.006313050165772438,
+          "multi_middle_q1": 0.0008207861101254821,
+          "multi_middle_q2": 0.03921769931912422,
+          "multi_pinky_q1": 0.0030009010806679726,
+          "multi_pinky_q2": 0.12349589914083481,
+          "multi_ring_q1": 0.06296631693840027,
+          "multi_ring_q2": 0.003505615284666419,
+          "pinky_q1": 0.0045578922145068645,
+          "pinky_q2": 0.0034143980592489243,
+          "ring_q1": 0.005928369238972664,
+          "ring_q2": 0.000456281122751534,
+          "thumb_q2": 4.938953779287658e-08
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 25
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 25
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 10.764911272676924,
+      "joint_effort_rms": 100.84668849982408,
+      "joint_jerk_rms": 1071.3354000276283,
+      "joint_velocity_rms": 0.834113216247886
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9238126348779331,
+      "mean_kinematic_grasp_stability": 0.6029405857292925,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.7727390986087104,
+        "obj_194_wineglass_2": 0.836979960519552,
+        "obj_194_wineglass_3": 0.7333472770488219
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.8651045681783913,
+        "mean_duration_score": 0.5741816707614863,
+        "mean_hold_score": 0.25416628288957355,
+        "mean_kinematic_stability_score": 0.1492570991317247,
+        "mean_legacy_gsi": 0.4372591105221819,
+        "mean_lift_score": 0.28359380615270446,
+        "mean_motion_stability_score": 0.23528649939551122
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2830.8,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-006/per_task.json

@@ -0,0 +1,216 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 6,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.24,
+        "success_rate_ci": {
+          "mean": 0.24,
+          "ci_low": 0.08,
+          "ci_high": 0.4
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.7466666666666667
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 33.661111111110486
+      },
+      "safety": {
+        "safe_success_rate": 0.24,
+        "hard_violation_rate": 0.68,
+        "drop_rate": 0.68,
+        "high_speed_violation_rate": 0.04,
+        "safety_violation_step_rate": 0.00036589332176929746
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.24,
+        "at_end_success_rate": 0.24,
+        "task_family_success_rate": {
+          "table_setting": 0.24
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.6266666666666666,
+        "current_chain_depth_progress_score": 0.6266666666666666,
+        "latched_chain_depth_progress_score": 0.7466666666666667,
+        "avg_current_chain_depth": 1.88,
+        "avg_latched_chain_depth": 2.24
+      },
+      "efficiency": {
+        "task_efficiency": 1.0731137093300667,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 2019.6666666666667,
+        "avg_policy_steps_to_success": 673.3333333333334,
+        "avg_policy_queries_to_success": 34.166666666666664
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.00036589332176929746,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 10,
+          "drop_to_table_obj_194_wineglass_2": 8,
+          "drop_to_table_obj_194_wineglass_3": 8,
+          "high_speed_obj_194_wineglass_2": 1
+        },
+        "robot_constraint_violation_rate": 0.24,
+        "finger_joint_limit_saturation_rate": 0.0,
+        "joint_limit_active_step_rate": 0.0005014093668690373,
+        "finger_joint_limit_saturation_step_rate": 0.0,
+        "max_joint_limit_excess_rad": 0.09870506823062897,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "index_q1": 15,
+            "middle_q1": 1,
+            "multi_pinky_q2": 10,
+            "multi_ring_q2": 11,
+            "pinky_q1": 4
+          },
+          "joint_limit_hard_counts": {
+            "index_q1": 3,
+            "index_q2": 2,
+            "middle_q1": 1,
+            "multi_index_q1": 1,
+            "multi_middle_q1": 1,
+            "multi_middle_q2": 1,
+            "multi_pinky_q2": 3,
+            "multi_ring_q1": 4,
+            "multi_ring_q2": 2,
+            "pinky_q1": 10,
+            "ring_q1": 2
+          },
+          "joint_limit_hard_episode_events": 23,
+          "finger_joint_limit_saturation_counts": {},
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.09870506823062897,
+            "index_q2": 0.00597022008150816,
+            "middle_q1": 0.010644328780472279,
+            "middle_q2": 0.004134062677621841,
+            "multi_index_q1": 0.005395730957388878,
+            "multi_index_q2": 0.004083765670657158,
+            "multi_middle_q1": 0.005551327019929886,
+            "multi_middle_q2": 0.007478164508938789,
+            "multi_pinky_q1": 0.0021698717027902603,
+            "multi_pinky_q2": 0.02618621289730072,
+            "multi_ring_q1": 0.00624956376850605,
+            "multi_ring_q2": 0.07862815260887146,
+            "pinky_q1": 0.023642776533961296,
+            "pinky_q2": 0.002643280429765582,
+            "ring_q1": 0.006825627293437719,
+            "ring_q2": 0.001219318131916225,
+            "thumb_q2": 0.002118239412084222
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 11.16668768604001,
+        "joint_effort_rms": 98.73497029459057,
+        "joint_jerk_rms": 1109.4477517628764,
+        "joint_velocity_rms": 0.8017986012352084
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.9320194001606834,
+        "mean_kinematic_grasp_stability": 0.6733275812221824,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.8822220862315222,
+          "obj_194_wineglass_2": 0.8461792752707818,
+          "obj_194_wineglass_3": 0.6414149785706683
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.8894488377542646,
+          "mean_duration_score": 0.571103531634818,
+          "mean_hold_score": 0.2919579706553225,
+          "mean_kinematic_stability_score": 0.19340316378101752,
+          "mean_legacy_gsi": 0.4642539699990262,
+          "mean_lift_score": 0.321248856656651,
+          "mean_motion_stability_score": 0.273675559580632
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2951.68,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-006/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 43761,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "none",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 1,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 43761,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-006/summary.json

@@ -0,0 +1,215 @@
+{
+  "metric_schema_version": "core-v1.8-policy-step-query-split",
+  "metadata": {
+    "policy_name": "uep_pi05",
+    "base_seed": 100300000,
+    "seed_policy": "episode_seed = base_seed + episode_index - 1",
+    "randomization_recorded": true,
+    "n_episodes": 25,
+    "successes": 6,
+    "error_count": 0,
+    "evaluation_protocol": "reach_and_stop",
+    "early_stop": true,
+    "success_definition": "terminal.raw_condition holds for dwell_time_s",
+    "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+    "max_episode_steps": 3246,
+    "dwell_time_s": 0.5,
+    "field_semantics": {
+      "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+      "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+      "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+      "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+      "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+      "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+      "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+      "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+      "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+      "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+      "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+      "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+      "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+      "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+      "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+      "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+      "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+      "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+      "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+      "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+      "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+      "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+      "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+      "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+      "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+      "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+    }
+  },
+  "core_metrics": {
+    "completion": {
+      "success_rate": 0.24,
+      "success_rate_ci": {
+        "mean": 0.24,
+        "ci_low": 0.08,
+        "ci_high": 0.4
+      }
+    },
+    "progress": {
+      "latched_stage_completion_rate": 0.7466666666666667
+    },
+    "efficiency": {
+      "avg_time_to_success_s": 33.661111111110486
+    },
+    "safety": {
+      "safe_success_rate": 0.24,
+      "hard_violation_rate": 0.68,
+      "drop_rate": 0.68,
+      "high_speed_violation_rate": 0.04,
+      "safety_violation_step_rate": 0.00036589332176929746
+    },
+    "robustness": {}
+  },
+  "diagnostic_metrics": {
+    "completion": {
+      "ever_instant_success_rate": 0.24,
+      "at_end_success_rate": 0.24,
+      "task_family_success_rate": {
+        "table_setting": 0.24
+      }
+    },
+    "progress": {
+      "current_stage_completion_rate": 0.6266666666666666,
+      "current_chain_depth_progress_score": 0.6266666666666666,
+      "latched_chain_depth_progress_score": 0.7466666666666667,
+      "avg_current_chain_depth": 1.88,
+      "avg_latched_chain_depth": 2.24
+    },
+    "efficiency": {
+      "task_efficiency": 1.0731137093300667,
+      "expert_time_s": 36.06666666666666,
+      "expert_time_step": 2164,
+      "avg_steps_to_success": 2019.6666666666667,
+      "avg_policy_steps_to_success": 673.3333333333334,
+      "avg_policy_queries_to_success": 34.166666666666664
+    },
+    "safety": {
+      "safety_violation_events_per_step": 0.00036589332176929746,
+      "violation_counts": {
+        "drop_to_table_obj_194_wineglass_1": 10,
+        "drop_to_table_obj_194_wineglass_2": 8,
+        "drop_to_table_obj_194_wineglass_3": 8,
+        "high_speed_obj_194_wineglass_2": 1
+      },
+      "robot_constraint_violation_rate": 0.24,
+      "finger_joint_limit_saturation_rate": 0.0,
+      "joint_limit_active_step_rate": 0.0005014093668690373,
+      "finger_joint_limit_saturation_step_rate": 0.0,
+      "max_joint_limit_excess_rad": 0.09870506823062897,
+      "robot_constraint_diagnostics": {
+        "joint_limit_soft_counts": {
+          "index_q1": 15,
+          "middle_q1": 1,
+          "multi_pinky_q2": 10,
+          "multi_ring_q2": 11,
+          "pinky_q1": 4
+        },
+        "joint_limit_hard_counts": {
+          "index_q1": 3,
+          "index_q2": 2,
+          "middle_q1": 1,
+          "multi_index_q1": 1,
+          "multi_middle_q1": 1,
+          "multi_middle_q2": 1,
+          "multi_pinky_q2": 3,
+          "multi_ring_q1": 4,
+          "multi_ring_q2": 2,
+          "pinky_q1": 10,
+          "ring_q1": 2
+        },
+        "joint_limit_hard_episode_events": 23,
+        "finger_joint_limit_saturation_counts": {},
+        "joint_limit_max_excess_by_joint": {
+          "index_q1": 0.09870506823062897,
+          "index_q2": 0.00597022008150816,
+          "middle_q1": 0.010644328780472279,
+          "middle_q2": 0.004134062677621841,
+          "multi_index_q1": 0.005395730957388878,
+          "multi_index_q2": 0.004083765670657158,
+          "multi_middle_q1": 0.005551327019929886,
+          "multi_middle_q2": 0.007478164508938789,
+          "multi_pinky_q1": 0.0021698717027902603,
+          "multi_pinky_q2": 0.02618621289730072,
+          "multi_ring_q1": 0.00624956376850605,
+          "multi_ring_q2": 0.07862815260887146,
+          "pinky_q1": 0.023642776533961296,
+          "pinky_q2": 0.002643280429765582,
+          "ring_q1": 0.006825627293437719,
+          "ring_q2": 0.001219318131916225,
+          "thumb_q2": 0.002118239412084222
+        },
+        "coverage": {
+          "hard_limit_source": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          },
+          "joint_role_manifest": {
+            "available": 25,
+            "unavailable": 0,
+            "unknown": 0
+          }
+        },
+        "hard_limit_sources": {
+          "runtime.joint_pos_limits": 25
+        },
+        "soft_limit_sources": {
+          "runtime.soft_joint_pos_limits": 25
+        },
+        "unknown_runtime_joint_names": [],
+        "baseline_checks": {
+          "post_home": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          },
+          "post_reset": {
+            "checked": 25,
+            "hard_violation_episodes": 0
+          }
+        }
+      }
+    },
+    "robot_motion": {
+      "joint_acceleration_rms": 11.16668768604001,
+      "joint_effort_rms": 98.73497029459057,
+      "joint_jerk_rms": 1109.4477517628764,
+      "joint_velocity_rms": 0.8017986012352084
+    },
+    "grasp": {
+      "kinematic_grasp_stability_index": 0.9320194001606834,
+      "mean_kinematic_grasp_stability": 0.6733275812221824,
+      "per_object_best_kinematic_grasp_stability": {
+        "obj_194_wineglass_1": 0.8822220862315222,
+        "obj_194_wineglass_2": 0.8461792752707818,
+        "obj_194_wineglass_3": 0.6414149785706683
+      },
+      "grasp_gsi_diagnostics": {
+        "mean_anti_slip_score": 0.8894488377542646,
+        "mean_duration_score": 0.571103531634818,
+        "mean_hold_score": 0.2919579706553225,
+        "mean_kinematic_stability_score": 0.19340316378101752,
+        "mean_legacy_gsi": 0.4642539699990262,
+        "mean_lift_score": 0.321248856656651,
+        "mean_motion_stability_score": 0.273675559580632
+      }
+    },
+    "tool": {
+      "tool_selection_accuracy": null,
+      "tool_switch_success_rate": null,
+      "tool_switch_total": 0,
+      "tool_switch_successful": 0
+    },
+    "runtime": {
+      "avg_steps": 2951.68,
+      "error_count": 0
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-007/per_task.json

@@ -0,0 +1,211 @@
+{
+  "scenes/03_wine_glass_plate_balance.yaml": {
+    "metric_schema_version": "core-v1.8-policy-step-query-split",
+    "metadata": {
+      "base_seed": 100300000,
+      "seed_policy": "episode_seed = base_seed + episode_index - 1",
+      "randomization_recorded": true,
+      "n_episodes": 25,
+      "successes": 15,
+      "error_count": 0,
+      "evaluation_protocol": "reach_and_stop",
+      "early_stop": true,
+      "success_definition": "terminal.raw_condition holds for dwell_time_s",
+      "measurement_window": "episode_start_to_stable_success_or_failure_budget",
+      "max_episode_steps": 3246,
+      "dwell_time_s": 0.5,
+      "field_semantics": {
+        "success_rate": "primary protocol success rate: stable_success for reach_and_stop; terminal at_end_success_budget for fixed_horizon",
+        "avg_time_to_success_s": "main efficiency metric: mean time_to_stable_success_s over reached-success episodes",
+        "task_efficiency": "benchmark-specific expert-normalized speed ratio: expert_time_s / time_to_stable_success_s over reached-success episodes",
+        "avg_steps_to_success": "diagnostic mean steps_to_stable_success over reached-success episodes, or None if no episode succeeds",
+        "avg_policy_steps_to_success": "diagnostic mean policy-rate control/action steps through stable success over reached-success episodes",
+        "avg_policy_queries_to_success": "diagnostic mean actual model inference queries through stable success over reached-success episodes; None when the first-success query count was not observed",
+        "safe_success_rate": "episode success rate with no task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "hard_violation_rate": "episode rate with any task-level hard safety violation (drop or tracked-object high speed; joint limits are diagnostic by default)",
+        "drop_rate": "episode rate with drop_off_table or drop_to_table_after_lift violation",
+        "high_speed_violation_rate": "episode rate with any tracked object exceeding the configured speed threshold",
+        "excessive_impact_rate": "deprecated contact-force impact metric; no longer populated by core safety",
+        "robot_constraint_violation_rate": "diagnostic episode rate with any asset hard-joint-limit observation; excluded from safe_success_rate unless include_in_core_safety is explicitly enabled",
+        "finger_joint_limit_saturation_rate": "episode rate with dexterous finger joints saturated beyond finger joint-limit thresholds; diagnostic by default and not included in safe_success_rate",
+        "joint_limit_active_step_rate": "diagnostic mean fraction of steps with any joint exceeding the soft diagnostic margin",
+        "finger_joint_limit_saturation_step_rate": "diagnostic mean fraction of steps with finger joint-limit saturation",
+        "max_joint_limit_excess_rad": "diagnostic maximum operating/soft-limit excess observed in radians; hard-limit source and role validation are in robot_constraint_diagnostics",
+        "safety_violation_rate_time": "compatibility diagnostic; with fixed physics_dt this is equivalent to safety_violation_step_rate and is omitted from grouped summary output",
+        "safety_violation_step_rate": "core safety metric: fraction of executed steps with task-level safety violations; joint-limit observations are excluded by default",
+        "safety_violation_rate_event": "deprecated compatibility alias of safety_violation_events_per_step; omitted from grouped summary output",
+        "safety_violation_events_per_step": "diagnostic event density; may exceed 1 when multiple violations happen in one step",
+        "episode_violation_rate": "compatibility alias of hard_violation_rate; omitted from grouped core output",
+        "latched_stage_completion_rate": "main progress metric; ever-reached stage completion rate at episode end",
+        "current_stage_completion_rate": "diagnostic terminal-state stage completion rate at episode end",
+        "current_normalized_progress_score": "chain-depth diagnostic; normalized progress score at episode end",
+        "current_chain_depth_progress_score": "alias of current_normalized_progress_score for ordered dependency-depth diagnostics",
+        "latched_chain_depth_progress_score": "alias of latched_normalized_progress_score for ever-reached dependency-depth diagnostics",
+        "robustness": "success rates grouped by benchmark protocol category: cov, inv, and inv+cov"
+      }
+    },
+    "core_metrics": {
+      "completion": {
+        "success_rate": 0.6,
+        "success_rate_ci": {
+          "mean": 0.6,
+          "ci_low": 0.4,
+          "ci_high": 0.8
+        }
+      },
+      "progress": {
+        "latched_stage_completion_rate": 0.8
+      },
+      "efficiency": {
+        "avg_time_to_success_s": 32.84333333333275
+      },
+      "safety": {
+        "safe_success_rate": 0.48,
+        "hard_violation_rate": 0.36,
+        "drop_rate": 0.32,
+        "high_speed_violation_rate": 0.04,
+        "safety_violation_step_rate": 0.0002257372740611748
+      },
+      "robustness": {}
+    },
+    "diagnostic_metrics": {
+      "completion": {
+        "ever_instant_success_rate": 0.6,
+        "at_end_success_rate": 0.6,
+        "task_family_success_rate": {
+          "table_setting": 0.6
+        }
+      },
+      "progress": {
+        "current_stage_completion_rate": 0.7733333333333333,
+        "current_chain_depth_progress_score": 0.7733333333333333,
+        "latched_chain_depth_progress_score": 0.8,
+        "avg_current_chain_depth": 2.32,
+        "avg_latched_chain_depth": 2.4
+      },
+      "efficiency": {
+        "task_efficiency": 1.1011607088308866,
+        "expert_time_s": 36.06666666666666,
+        "expert_time_step": 2164,
+        "avg_steps_to_success": 1970.6,
+        "avg_policy_steps_to_success": 657.2,
+        "avg_policy_queries_to_success": 33.2
+      },
+      "safety": {
+        "safety_violation_events_per_step": 0.0002257372740611748,
+        "violation_counts": {
+          "drop_to_table_obj_194_wineglass_1": 3,
+          "drop_to_table_obj_194_wineglass_2": 7,
+          "drop_to_table_obj_194_wineglass_3": 3,
+          "high_speed_distractor_01_clutter_061_foam_brick": 1
+        },
+        "robot_constraint_violation_rate": 0.2,
+        "finger_joint_limit_saturation_rate": 0.04,
+        "joint_limit_active_step_rate": 0.0005320950031441978,
+        "finger_joint_limit_saturation_step_rate": 3.224818200873926e-05,
+        "max_joint_limit_excess_rad": 0.22528931498527527,
+        "robot_constraint_diagnostics": {
+          "joint_limit_soft_counts": {
+            "index_q1": 23,
+            "multi_index_q1": 8,
+            "multi_middle_q1": 5
+          },
+          "joint_limit_hard_counts": {
+            "index_q1": 2,
+            "index_q2": 1,
+            "multi_index_q1": 5,
+            "multi_middle_q1": 6,
+            "multi_pinky_q1": 1,
+            "multi_ring_q1": 2
+          },
+          "joint_limit_hard_episode_events": 7,
+          "finger_joint_limit_saturation_counts": {
+            "index_q1": 2
+          },
+          "joint_limit_max_excess_by_joint": {
+            "index_q1": 0.22528931498527527,
+            "index_q2": 0.005092688836157322,
+            "middle_q1": 0.0026535438373684883,
+            "middle_q2": 0.0033318917267024517,
+            "multi_index_q1": 0.05214352533221245,
+            "multi_index_q2": 0.004233008250594139,
+            "multi_middle_q1": 0.0349835567176342,
+            "multi_middle_q2": 0.0006088213413022459,
+            "multi_pinky_q1": 0.005604577716439962,
+            "multi_pinky_q2": 0.0003086419019382447,
+            "multi_ring_q1": 0.006793745793402195,
+            "multi_ring_q2": 0.0005573886446654797,
+            "pinky_q1": 0.004016959108412266,
+            "pinky_q2": 0.0005209397641010582,
+            "ring_q1": 0.003422384150326252,
+            "ring_q2": 0.0009658187627792358,
+            "thumb_q2": 4.865187364089252e-08
+          },
+          "coverage": {
+            "hard_limit_source": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            },
+            "joint_role_manifest": {
+              "available": 25,
+              "unavailable": 0,
+              "unknown": 0
+            }
+          },
+          "hard_limit_sources": {
+            "runtime.joint_pos_limits": 25
+          },
+          "soft_limit_sources": {
+            "runtime.soft_joint_pos_limits": 25
+          },
+          "unknown_runtime_joint_names": [],
+          "baseline_checks": {
+            "post_home": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            },
+            "post_reset": {
+              "checked": 25,
+              "hard_violation_episodes": 0
+            }
+          }
+        }
+      },
+      "robot_motion": {
+        "joint_acceleration_rms": 9.796505109838256,
+        "joint_effort_rms": 101.82403123596403,
+        "joint_jerk_rms": 967.7767999254373,
+        "joint_velocity_rms": 0.8412114914137719
+      },
+      "grasp": {
+        "kinematic_grasp_stability_index": 0.8989342642268521,
+        "mean_kinematic_grasp_stability": 0.6185771648735962,
+        "per_object_best_kinematic_grasp_stability": {
+          "obj_194_wineglass_1": 0.815547996311865,
+          "obj_194_wineglass_2": 0.8745997725106853,
+          "obj_194_wineglass_3": 0.6797169214752836
+        },
+        "grasp_gsi_diagnostics": {
+          "mean_anti_slip_score": 0.9236379774824328,
+          "mean_duration_score": 0.5311470575072317,
+          "mean_hold_score": 0.21039296236291133,
+          "mean_kinematic_stability_score": 0.14556781755795234,
+          "mean_legacy_gsi": 0.4133012816611011,
+          "mean_lift_score": 0.23249482980942598,
+          "mean_motion_stability_score": 0.1901933533282341
+        }
+      },
+      "tool": {
+        "tool_selection_accuracy": null,
+        "tool_switch_success_rate": null,
+        "tool_switch_total": 0,
+        "tool_switch_successful": 0
+      },
+      "runtime": {
+        "avg_steps": 2480.76,
+        "error_count": 0
+      }
+    }
+  }
+}

MXI/per_task_eval/03_wine_glass_plate_balance/native_shards/source-007/run-protocol.json

@@ -0,0 +1,47 @@
+{
+  "schema_version": 2,
+  "transport": "direct_uep_ws",
+  "backend": "pytorch",
+  "record_profile": "preview",
+  "training_sampling": null,
+  "existing_server_port": 43049,
+  "rollout_split": "diagnostic",
+  "record_dir": null,
+  "rollout_dataset_dir": null,
+  "rollout_chunk_size": 20,
+  "task": "03_wine_glass_plate_balance",
+  "task_cli": "scenes/03_wine_glass_plate_balance.yaml",
+  "robot_key": "multi_xarm7_with_ability",
+  "channel": "none",
+  "generalization_split": "unseen",
+  "anchor_dir": "workspace/datasets/Bench2Dex/teleopdata/dataset/03_wine_glass_plate_balance/origin-generalization",
+  "expected_anchor_count": null,
+  "episodes": 25,
+  "start_episode": 26,
+  "seed": 100000000,
+  "inv_cov_seed": 200000000,
+  "episode_steps": 1082,
+  "warmup_steps": 60,
+  "eval_action_horizon": 20,
+  "uep_num_steps": 10,
+  "uep_action_dim": 26,
+  "uep_state_dim": 26,
+  "uep_config": "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json",
+  "uep_checkpoint": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt",
+  "uep_assets": "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets",
+  "uep_port": 43049,
+  "sim_wrapper": "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh",
+  "driver": "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py",
+  "source_sha256": {
+    "workspace/tasks/bench2dex-multitask26-step245k-local-eval-20261111/configs/train-frozen.json": "daa293b4b8763a4aa468ad3a1fe658882f9b11851d246323a8378e37cfca032b",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/checkpoints/step_00353000.pt": "08d6049fa2230c9097ed56d4320d1825e66905a6e2aab488c5da22daaaf0f3e6",
+    "workspace/tasks/bench2dex-pi05-multitask26-8a800-20261107/data/assets/Bench2Dex_task03_multi_xarm7_with_ability.json": "f86229e899e4fe062e1187c009aa4155736e6118897a9b1ccd2cb2ca9e413e77",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/scripts/policy_server.py": "aac37e777d465df0185237ad5a3e81663a415d790042c708d845e5a42cf5d5da",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/main.py": "6fd6f73751b4ec645e2bcd7a6f934350d20107089fae7684e856bbb43078bd00",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_recorder.py": "83eaba6f37a35e17dd9b7ababa8d6585a0bb888b15c8e0e2cfc61ba99601868f",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_dataset.py": "9e8d0fa46caa67f528c49ccf66d93b1c793b3ce8b679c1f842712710ffc584d4",
+    "workspace/tasks/bench2dex-pi05-multitask26-resume147k-r3-20261110/source/tools/examples/bench2dex/rollout_sampling.py": "db93adb67732d2ecffd7a8e420f120c6b2b0db24304581888db2e2a5e284ef8d",
+    "workspace/tasks/bench2dex-pi05-priority-backlog-cloud-eval-8pro6000-20261112/native_bench/run_policy.py": "32d380738c9036d81a09966d6eb1b1bd5187172772b680d9b1427e20861c7b76",
+    "workspace/tasks/bench2dex-multitask26-latest-cloud-eval-8pro6000-20261113/sim-runtime.sh": "b99223e64dc09a4dc74a0a694d0024e35356049fdd2df5660fd1d64fb5affe88"
+  }
+}