{
  "schema_version": 1,
  "status": "frozen_before_factory_closed_loop_outcomes",
  "created_at_utc": "2026-11-12T18:11:16.921403+00:00",
  "scope": "Nine published G1 Factory tasks; matched native randomized resets; current-checkpoint engineering comparison",
  "tasks": [
    {
      "id": "LMBoxLift",
      "instruction": "Walk to the table and pick up the box"
    },
    {
      "id": "LMBoxLiftFloor",
      "instruction": "Walk to the box and pick it up from the floor"
    },
    {
      "id": "LMBoxTableToShelfPnP",
      "instruction": "Move the box from the table to the shelf"
    },
    {
      "id": "LMDrillLift",
      "instruction": "Walk to the table and pick up the power drill"
    },
    {
      "id": "LMDrillLiftObs",
      "instruction": "Navigate around the shelf and pick up the power drill from the table"
    },
    {
      "id": "LMDrillPnP",
      "instruction": "Move the power drill from one table to another"
    },
    {
      "id": "LMPickDrillFromHolder",
      "instruction": "Pick up the power drill from the holder"
    },
    {
      "id": "LMPushButton",
      "instruction": "Walk to the console and press the red button"
    },
    {
      "id": "LMPushShelfForward",
      "instruction": "Push the tall wheeled shelf to the marked area"
    }
  ],
  "modes": [
    "act",
    "groot",
    "adaptive"
  ],
  "evaluation_seeds": [
    1000,
    1001,
    1002,
    1003,
    1004,
    1005,
    1006,
    1007,
    1008,
    1009
  ],
  "pilot_seeds": [
    42
  ],
  "pilot_results_excluded": true,
  "physics_device": "cpu",
  "control_hz": 50,
  "prediction_horizon": 40,
  "execute_steps": 40,
  "episode_seconds": 20,
  "runtime_selection": "Factory pinned gitlinks and their declared IsaacSim6.0.1 runtime; exact revisions recorded in source receipt. Model-card6.2 conflict disclosed.",
  "dataset_revision": "1c5b2fbfaa35f70451bb462426c1b8adec1b7ec6",
  "expert_revision": "e0f5a056257fc89b386b7584cd5978285c6a02de",
  "act": {
    "members": 4,
    "task_conditioning": "31 proprio +9 one-hot catalog task ID",
    "rgb_preprocess": "Native640x480\u2192320x240 exact integer2x2 area resize; same implementation in cache and inference",
    "split_per_task": {
      "train_pool": 90,
      "heldout_calibration": 10,
      "subbag_per_member": 72
    },
    "training_choice": "Cost pilot before production; production checkpoint choice based on held-out command prediction, before evaluation outcomes"
  },
  "router": {
    "quantile": "q95_per_task",
    "calibration_observations_per_task": 100,
    "refresh_chunks": 3,
    "release_confirmations": 2,
    "crossfade_steps": 5,
    "physical_bounds": {
      "wrist_rms_mm": 40.0,
      "wrist_rms_deg": 20.0,
      "hand_rms_rad": 0.18,
      "height_rms_m": 0.02,
      "navigation_xy_rms_mps": 0.05,
      "navigation_yaw_rms_radps": 0.1
    },
    "groups": "Both wrists through actual Factory USD FK, both hands, base height, navigation. Waist affects wrist FK. No elbow-angle discrepancy penalty."
  },
  "native_success": "Unmodified per-task native termination predicate; preserve component measurements and disclose proxy limitations.",
  "attempt_selection": "First infrastructure-valid completed episode for each task/seed/mode; never rerun task failures to select a better outcome.",
  "validation": [
    "Matched initial robot joints/root pose and relevant object/articulation states across modes",
    "Camera and native WBC consistent across modes",
    "Actual robot USD FK versus simulator wrist poses",
    "Exact recorded video frame alignment",
    "Expert RNG isolation across clients",
    "Frozen model, calibration, source, asset identities"
  ],
  "interpretation": [
    "10 matched reset seeds per task is a small engineering sample; not an exact reproduction of published100-episode figures.",
    "Expert trained on all900public demonstrations plus foundation pretraining; ACT810train90calibration uses taskID and smaller images. Not a controlled causal architecture comparison.",
    "Synchronous inference pauses simulation; no asynchronous latency-hiding, Jetson orWAN claim.",
    "Unknown policy disagreement between queries remains unknown, not zero.",
    "No threshold tuning on final evaluation outcomes.",
    "Native success may use geometric or hand-position proxies rather than physical grasp contact; report native predicate names and visual evidence."
  ]
}