[
  {
    "name": "Transition Dataset",
    "type": "Dataset Pattern",
    "focus": "One-step dynamics",
    "best_for": [
      "Dynamics learning",
      "Control",
      "Simple environments"
    ],
    "required": [
      "observation_t",
      "action_t",
      "observation_t+1"
    ],
    "optional": [
      "reward",
      "done",
      "goal",
      "timestamp"
    ],
    "risk": "Frame-level random splits can leak nearly identical states across train and test."
  },
  {
    "name": "Episode Dataset",
    "type": "Dataset Pattern",
    "focus": "Sequential dynamics",
    "best_for": [
      "Rollouts",
      "Memory",
      "Long-horizon modeling"
    ],
    "required": [
      "episode_id",
      "observations[]",
      "actions[]"
    ],
    "optional": [
      "rewards[]",
      "terminal",
      "instruction",
      "metadata"
    ],
    "risk": "Broken episode boundaries can destroy temporal structure."
  },
  {
    "name": "Multimodal Episode",
    "type": "Dataset Pattern",
    "focus": "Cross-modal world state",
    "best_for": [
      "Robotics",
      "Embodied AI",
      "Physical AI"
    ],
    "required": [
      "video",
      "actions",
      "robot_state"
    ],
    "optional": [
      "depth",
      "audio",
      "language",
      "force",
      "pose"
    ],
    "risk": "Unsynchronized modalities create false transition errors."
  },
  {
    "name": "One-Step Benchmark",
    "type": "Benchmark Pattern",
    "focus": "Immediate next-state prediction",
    "best_for": [
      "Fast iteration",
      "Architecture debugging"
    ],
    "required": [
      "held-out transitions",
      "prediction metric"
    ],
    "optional": [
      "uncertainty",
      "per-task breakdown"
    ],
    "risk": "Can hide catastrophic long-horizon drift."
  },
  {
    "name": "Multi-Horizon Benchmark",
    "type": "Benchmark Pattern",
    "focus": "Error growth over time",
    "best_for": [
      "Rollouts",
      "Planning",
      "Simulation"
    ],
    "required": [
      "horizons",
      "rollout protocol",
      "per-horizon metric"
    ],
    "optional": [
      "drift ratio",
      "consistency score"
    ],
    "risk": "Averaging across horizons can conceal where failure begins."
  },
  {
    "name": "Action Fidelity Benchmark",
    "type": "Benchmark Pattern",
    "focus": "Correct action consequences",
    "best_for": [
      "Robotics",
      "Interactive worlds",
      "Agents"
    ],
    "required": [
      "paired actions",
      "ground-truth transitions"
    ],
    "optional": [
      "counterfactual actions",
      "action sensitivity"
    ],
    "risk": "Visual similarity alone does not prove correct causal response."
  },
  {
    "name": "OOD Generalization Split",
    "type": "Split Strategy",
    "focus": "Distribution shift",
    "best_for": [
      "Deployment",
      "Robotics",
      "Robustness"
    ],
    "required": [
      "held-out environments or tasks"
    ],
    "optional": [
      "held-out objects",
      "held-out embodiments"
    ],
    "risk": "Random splits can dramatically overestimate generalization."
  },
  {
    "name": "Planning Utility Benchmark",
    "type": "Benchmark Pattern",
    "focus": "Downstream decision quality",
    "best_for": [
      "Agents",
      "MPC",
      "Model-based RL"
    ],
    "required": [
      "planner",
      "task success metric",
      "world-model rollouts"
    ],
    "optional": [
      "regret",
      "return",
      "sample efficiency"
    ],
    "risk": "Prediction scores may not correlate with better decisions."
  },
  {
    "name": "Control Benchmark",
    "type": "Benchmark Pattern",
    "focus": "Closed-loop task performance",
    "best_for": [
      "Robotics",
      "Autonomous systems"
    ],
    "required": [
      "environment",
      "policy/controller",
      "success metric"
    ],
    "optional": [
      "safety violations",
      "energy",
      "path efficiency"
    ],
    "risk": "Offline prediction quality may not transfer to closed-loop control."
  },
  {
    "name": "Efficiency Benchmark",
    "type": "Benchmark Pattern",
    "focus": "Operational cost",
    "best_for": [
      "Real-time systems",
      "Large search spaces"
    ],
    "required": [
      "latency",
      "memory",
      "throughput"
    ],
    "optional": [
      "energy",
      "rollouts/sec",
      "cost"
    ],
    "risk": "A strong model may be unusable if rollouts are too slow for planning."
  }
]