[
  {
    "name": "Multi-Step Prediction Loss",
    "phase": "Training",
    "category": "Long-Horizon Learning",
    "difficulty": "Advanced",
    "goal": "Reduce compounding rollout error",
    "summary": "Train the world model over several imagined future steps instead of only a one-step target.",
    "signals": [
      "multi-step loss",
      "rollout consistency",
      "horizon weighting"
    ],
    "when": "Use when one-step predictions look strong but longer rollouts drift quickly."
  },
  {
    "name": "Latent Overshooting",
    "phase": "Training",
    "category": "Latent Dynamics",
    "difficulty": "Advanced",
    "goal": "Stabilize latent dynamics across multiple steps",
    "summary": "Compare predicted latent states against future encoded targets several steps ahead.",
    "signals": [
      "latent loss",
      "overshooting horizon",
      "representation stability"
    ],
    "when": "Useful for latent-state models and model-based reinforcement learning."
  },
  {
    "name": "Scheduled Sampling",
    "phase": "Training",
    "category": "Exposure Bias",
    "difficulty": "Intermediate",
    "goal": "Reduce train-deployment mismatch",
    "summary": "Gradually train on model-generated states instead of always conditioning on ground truth.",
    "signals": [
      "teacher forcing ratio",
      "self-conditioned steps"
    ],
    "when": "Useful when rollout behavior degrades because the model never trained on its own errors."
  },
  {
    "name": "Action-Conditioned Consistency",
    "phase": "Training",
    "category": "Control",
    "difficulty": "Advanced",
    "goal": "Preserve causal action effects",
    "summary": "Encourage different actions to produce meaningfully different predicted futures.",
    "signals": [
      "action sensitivity",
      "counterfactual consistency"
    ],
    "when": "Important for robotics, agents and interactive world models."
  },
  {
    "name": "Joint Reconstruction + Prediction",
    "phase": "Training",
    "category": "Representation Learning",
    "difficulty": "Intermediate",
    "goal": "Balance state fidelity and predictive usefulness",
    "summary": "Combine reconstruction with future-state objectives so the latent representation remains informative and predictive.",
    "signals": [
      "reconstruction loss",
      "prediction loss",
      "latent regularization"
    ],
    "when": "Useful when the representation becomes too compressed or too visually focused."
  },
  {
    "name": "Rollout Error by Horizon",
    "phase": "Evaluation",
    "category": "Long-Horizon Evaluation",
    "difficulty": "Core",
    "goal": "Measure error accumulation",
    "summary": "Evaluate prediction error at each future horizon rather than reporting only an average.",
    "signals": [
      "error@1",
      "error@5",
      "error@10",
      "drift curve"
    ],
    "when": "Essential for any model used beyond one-step forecasting."
  },
  {
    "name": "Action Fidelity Test",
    "phase": "Evaluation",
    "category": "Control",
    "difficulty": "Advanced",
    "goal": "Verify that action changes are reflected in predicted futures",
    "summary": "Compare predicted consequences under different actions and validate against real transitions.",
    "signals": [
      "action effect error",
      "counterfactual separation"
    ],
    "when": "Critical for robotics, control and interactive simulation."
  },
  {
    "name": "Uncertainty Calibration",
    "phase": "Evaluation",
    "category": "Reliability",
    "difficulty": "Advanced",
    "goal": "Measure whether confidence tracks prediction quality",
    "summary": "Evaluate if uncertain predictions receive lower confidence than reliable predictions.",
    "signals": [
      "calibration error",
      "coverage",
      "ensemble variance"
    ],
    "when": "Important for safety-sensitive or out-of-distribution settings."
  },
  {
    "name": "Distribution Shift Evaluation",
    "phase": "Evaluation",
    "category": "Generalization",
    "difficulty": "Advanced",
    "goal": "Measure robustness beyond familiar training conditions",
    "summary": "Test new environments, dynamics, objects, actions or visual conditions not seen during training.",
    "signals": [
      "OOD error",
      "relative degradation",
      "recovery"
    ],
    "when": "Necessary before deployment into diverse real environments."
  },
  {
    "name": "Planning Utility",
    "phase": "Evaluation",
    "category": "Decision Making",
    "difficulty": "Advanced",
    "goal": "Measure whether the world model improves decisions",
    "summary": "Evaluate downstream planning or control performance using the learned world model.",
    "signals": [
      "task success",
      "return",
      "planning regret",
      "sample efficiency"
    ],
    "when": "The strongest evaluation when the model is intended for agents or control."
  },
  {
    "name": "Compute & Rollout Efficiency",
    "phase": "Evaluation",
    "category": "Efficiency",
    "difficulty": "Core",
    "goal": "Quantify operational cost",
    "summary": "Measure latency, memory and throughput for one-step and multi-step simulation.",
    "signals": [
      "ms/step",
      "tokens/s",
      "VRAM",
      "rollouts/s"
    ],
    "when": "Important for real-time planning and large candidate-search spaces."
  },
  {
    "name": "Model Error vs. Task Error",
    "phase": "Evaluation",
    "category": "Task Relevance",
    "difficulty": "Advanced",
    "goal": "Separate prediction quality from downstream usefulness",
    "summary": "Compare predictive error with the actual effect on planning, control or task success.",
    "signals": [
      "prediction loss",
      "task success",
      "control regret"
    ],
    "when": "Useful when low reconstruction loss does not translate into better decisions."
  }
]