{
  "episodes": 100,
  "eef_decisions": 4081,
  "subgoal_segments": 1288,
  "task_families": {
    "transfer": 52,
    "unresolved": 24,
    "rotate_object": 1,
    "surface_move": 4,
    "wipe": 2,
    "press": 4,
    "multi_step": 8,
    "articulated": 4,
    "scoop": 1
  },
  "instruction_available": 77,
  "missing_instructions": 23,
  "recorded_outcomes": {
    "success": 81,
    "failure": 19
  },
  "split_scope": "Episode IDs are disjoint (80/10/10). Scenes, operators, tasks, and recording dates are not held out; this sample is not a generalization benchmark.",
  "annotation_status": "Generated weak semantic labels; no per-decision object verification. Recorded outcome metadata does not verify generated subgoal completion.",
  "causal_boundary": "Current/past observations are inputs; derived actions and rationales are posthoc labels. Feedback/progress are available only after the action segment. Generated rationale text references the labeled action and must remain on the output side.",
  "task_family_note": "24 episodes have unresolved generated task-family categorization; 23 episodes lack task instructions. These are different categories."
}