[
  {
    "name": "Temporal Synchronization",
    "category": "Alignment",
    "question": "Are sensor streams, actions and events aligned on the correct timeline?",
    "metrics": [
      "timestamp error",
      "cross-modal lag",
      "event alignment accuracy"
    ],
    "inputs": [
      "sensor timestamps",
      "action logs",
      "event markers"
    ],
    "failure": "Small timing offsets can create false cause-and-effect relationships."
  },
  {
    "name": "Spatial Calibration",
    "category": "Alignment",
    "question": "Do observations agree in a common coordinate frame?",
    "metrics": [
      "reprojection error",
      "pose error",
      "extrinsic drift"
    ],
    "inputs": [
      "intrinsics",
      "extrinsics",
      "reference targets"
    ],
    "failure": "Calibration drift corrupts geometry and state estimation."
  },
  {
    "name": "Geometry Fidelity",
    "category": "Reconstruction",
    "question": "How accurately is real-world geometry preserved?",
    "metrics": [
      "surface distance",
      "depth error",
      "mesh completeness",
      "scale error"
    ],
    "inputs": [
      "reference geometry",
      "reconstruction",
      "depth"
    ],
    "failure": "Visual realism does not guarantee metric accuracy."
  },
  {
    "name": "Object Persistence",
    "category": "Reconstruction",
    "question": "Are objects and agents tracked consistently over time?",
    "metrics": [
      "ID switches",
      "track fragmentation",
      "trajectory error"
    ],
    "inputs": [
      "tracks",
      "ground-truth IDs",
      "trajectories"
    ],
    "failure": "Identity errors break long-horizon interaction histories."
  },
  {
    "name": "Semantic Accuracy",
    "category": "Semantics",
    "question": "Are objects, relations, tasks and affordances represented correctly?",
    "metrics": [
      "object accuracy",
      "relation F1",
      "affordance accuracy"
    ],
    "inputs": [
      "labels",
      "scene graph",
      "task annotations"
    ],
    "failure": "Correct geometry can still be unusable when semantics are wrong."
  },
  {
    "name": "Action Alignment",
    "category": "Interaction",
    "question": "Are actions paired with the exact state transitions they produced?",
    "metrics": [
      "action-state lag",
      "transition pairing accuracy",
      "causal consistency"
    ],
    "inputs": [
      "state_t",
      "action_t",
      "state_t+1"
    ],
    "failure": "Misaligned actions damage control and action-conditioned world models."
  },
  {
    "name": "State Completeness",
    "category": "Representation",
    "question": "Does the world state preserve variables required by downstream tasks?",
    "metrics": [
      "state coverage",
      "missing-variable rate",
      "task sufficiency"
    ],
    "inputs": [
      "compiled state",
      "task requirements"
    ],
    "failure": "Overcompression may remove variables needed for planning."
  },
  {
    "name": "Representation Efficiency",
    "category": "Representation",
    "question": "How compact is the representation relative to preserved information?",
    "metrics": [
      "compression ratio",
      "bytes per step",
      "encoding latency"
    ],
    "inputs": [
      "raw episode",
      "compiled representation"
    ],
    "failure": "An accurate representation may still be too costly to scale."
  },
  {
    "name": "Uncertainty Calibration",
    "category": "Reliability",
    "question": "Does confidence reflect actual compilation quality?",
    "metrics": [
      "calibration error",
      "coverage",
      "confidence-error correlation"
    ],
    "inputs": [
      "confidence scores",
      "ground-truth error"
    ],
    "failure": "Overconfident world states are dangerous for autonomous systems."
  },
  {
    "name": "Real2Sim Fidelity",
    "category": "Simulation",
    "question": "Does simulation reproduce relevant real-world behavior?",
    "metrics": [
      "trajectory divergence",
      "contact consistency",
      "sensor similarity",
      "physics error"
    ],
    "inputs": [
      "real trajectories",
      "sim trajectories",
      "contact events"
    ],
    "failure": "Visual similarity alone does not guarantee useful physical behavior."
  },
  {
    "name": "Downstream World-Model Utility",
    "category": "Downstream",
    "question": "Does compiled data improve world-model learning and rollout quality?",
    "metrics": [
      "prediction error",
      "rollout drift",
      "sample efficiency"
    ],
    "inputs": [
      "compiled dataset",
      "world model",
      "evaluation tasks"
    ],
    "failure": "Clean representations may still fail to improve predictive modeling."
  },
  {
    "name": "Planning & Control Utility",
    "category": "Downstream",
    "question": "Does world compilation improve decisions, planning or control?",
    "metrics": [
      "task success",
      "planning regret",
      "control return",
      "policy transfer"
    ],
    "inputs": [
      "planner",
      "controller",
      "compiled world",
      "task environment"
    ],
    "failure": "The strongest test is whether better compilation produces better actions."
  }
]