{
  "schema_version": "1.0.0",
  "title": "Occlusion-aware human-to-robot garment-fold comparison",
  "status": "WORKING",
  "decision": "PASS_IMAGE_SPACE",
  "display_asset": {
    "path": "h2r-garment-fold-occlusion-completion.mp4",
    "sha256": "915c21139997c1d2a3fe503dd6bc8272de199df295a571dab55ebb52fc92d844",
    "codec": "h264",
    "width": 1920,
    "height": 540,
    "fps": 24,
    "frames": 192,
    "duration_seconds": 8.0,
    "source_time_normalized": true,
    "presentation_only": true
  },
  "poster": {
    "path": "h2r-garment-fold-occlusion-completion-poster.jpg",
    "sha256": "d8135f7ddbd81efd97299d55877719787be93bcb85d80316b3cbcf8308ef4887"
  },
  "source": {
    "kind": "real_human_video",
    "sha256": "fe964197b31cb7d68bc657bc8b2788e14f6c0c213ae9e9b1fa8209c508673650",
    "width": 3840,
    "height": 2160,
    "fps": 25,
    "frames": 273,
    "duration_seconds": 10.92,
    "visible_occlusion": "A near-camera laptop or monitor covers a substantial middle-right workspace region during the middle and late task, and a near-camera finger covers part of the terminal frame."
  },
  "candidate": {
    "generation_model": "MiniMaxAI/MiniMax-H3 Ref2VA",
    "generation_model_version": "42ed227ee7df40d41602854ae760620d6eb651fe",
    "seed": 2026089401,
    "sha256": "a36e1eed819573f44724700ca7db81637a19ecec61c47d271823309aadf4680d",
    "target_robot": "dual_industrial_robot_arm",
    "target_end_effector": "parallel_jaw_gripper",
    "width": 1024,
    "height": 768,
    "fps": 24,
    "frames": 192,
    "duration_seconds": 8.0
  },
  "automatic_evaluation": {
    "result_file_sha256": "c29f5a42c3471c9d30cf6d7b4e07ae94268e2622ae72f9a1c707d3068fa89c64",
    "transfer_core": 98.33333333333331,
    "visual_quality": 100.0,
    "all_six_hard_gates_true": true,
    "maximum_critical_disagreement_0_to_4": 0.8,
    "judge_model_families": ["Qwen3-VL", "Gemma-4", "Ministral-3"],
    "constructed_negative_suite": {
      "case_count": 6,
      "all_non_abstaining_reject": true,
      "report_sha256": "7c424ebff2722cd325f399ccc26fc9fd256ec40abd167111c3f71e012f55eb38"
    }
  },
  "occlusion_completion_claim": "The candidate uses visible task context to generate an unoccluded, source-relative robot execution with a continuous garment identity, fold sequence, and terminal state. This is semantic task completion, not a claim that hidden source pixels have a unique recovered ground truth.",
  "physical_boundary_followup": {
    "path": "h2r-garment-fold-physical-boundary-followup.json",
    "status": "PARTIAL",
    "manifest_sha256": "75027355299e930719c7aba0da6bb9e23e9da8f4b4ca61b14527c590ce7415f0",
    "completed": "Fail-closed physical gates, scale and force non-identifiability experiments, eight anti-spoof attacks, human-agreement harness controls, and a frozen external evidence acquisition contract.",
    "candidate_physical_gates_passed": 0,
    "candidate_physical_gates_total": 9,
    "promoted": false
  },
  "claim_boundary": "PASS_IMAGE_SPACE and occlusion completion refer only to visible 2-D video evidence. The follow-up validates the physical-evidence harness and shows why monocular RGB cannot establish the missing measurements; it does not establish metric 3-D cloth geometry, force, collision safety, joint feasibility, simulation success, real-robot execution, or agreement with real human ratings."
}