Section 39.7: Evaluating consistency, controllability, and horizon

"Agents fail at the bottleneck, not at the mean. Evaluate the minimum, not the average."

A World Model Needs More Than One Score

Problem First

Hand a planner a world model that scores 0.91 on every perceptual leaderboard, and it may still drive your robot into a wall the metric never saw coming: the number you trusted measured how pretty the future looked, not whether the future obeyed the robot. This section gives you a way to catch that gap, auditing any generative world model along three separate axes, consistency, controllability, and usable horizon, so you can decide whether a planner can trust it. Researchers and product teams love single numbers, but generative world models rarely fail in one dimension. A future can look consistent but ignore action, obey action for a few steps but drift later, or stay persistent while breaking task semantics. Evaluation must therefore expose the failure axis, not hide it. Model uncertainty and calibration interact with all three axes: a well-calibrated generator should signal when its own predictions become unreliable.

Core Model

A useful evaluation panel separates at least three core properties: $$\text{consistency}: o_t \rightarrow o_{t+1} \text{ stays semantically coherent},$$ $$\text{controllability}: a_t \text{ changes the future in the intended direction},$$ $$\text{usable horizon}: H^* = \max H \text{ such that the generated future remains decision-valid.}$$

"Decision-valid" means the generated future is close enough to what would actually happen that a planner choosing between actions based on it would make the same choice it would make against ground truth; once the generated frames drift far enough from that, the horizon has been exceeded even if the video still looks smooth.

Consistency is measured the same way the other two axes are: by running a fixed evaluation panel and scoring the output, typically with object-permanence checks (does an object that leaves the frame and returns keep the same identity and position) and geometry checks (does room or hallway layout stay fixed across frames) against a reference trace, rather than by inspecting a single clip for a plausible look. For embodied use, the usable horizon is usually the most revealing number. It measures not how long the model can keep drawing plausible frames, but how long the generated future remains trustworthy enough for policy learning, evaluation, or planning. Figure 39.7B shows why the three axes are combined with a minimum rather than an average: the weakest axis is the one that decides whether a planner should trust the model.

Named systems illustrate why these axes diverge in practice. Sora (OpenAI, 2024) demonstrated strong short-clip consistency but was not designed for action-conditional control (control where the model's next output depends on an explicit action input, rather than only on the preceding frames), so its controllability score for embodied planning tasks is effectively undefined. Genie 2 (Google DeepMind, 2024) accepted discrete action tokens and, according to the reported demonstrations, showed measurable controllability over rollouts typically in the 10-20 second range. Its usable horizon has been reported to degrade noticeably beyond roughly 15 seconds of branching in these demonstrations, though the exact cutoff depends on the scene and action sequence tested. NVIDIA Cosmos (2025) targets physical-AI pipelines and reports consistency metrics per domain. In practice it measures usable horizon by downstream policy-improvement results rather than perceptual scores alone. These examples show that the three axes do not move together and must be reported separately.

Of the three axes these systems expose, controllability is the one most often reported loosely, so it is worth pinning down before any cross-model comparison. Controllability deserves its own operational definition rather than a vague reference to "intended direction." In practice, measure it by running a fixed action script (a predetermined, repeatable sequence of commands, used so every model is tested against the exact same inputs), for example, turn left 30 degrees, move forward 2 meters, stop, from a held-out initial state, then compare the resulting observation against a ground-truth simulator or a human-labeled expected outcome at each step. The reference here is a conventional physics engine, not the learned simulator under test. A simple scalar is the fraction of action steps for which the world model's output falls within an acceptable tolerance of the reference. Systems that score high on consistency but low on this probe are generating plausible-looking futures that are essentially action-independent; this is called action-blind hallucination, and it is a critical failure mode for any planner that relies on the model to evaluate counterfactuals. In practice, a model can score 0.88 on consistency while its controllability probe sits at 0.42, meaning the generated future looks coherent but responds to actions at barely chance level.

If a model scores 0.88 on consistency but only 0.42 on controllability, what is the planner actually optimizing against?

Action-blind hallucination is particularly dangerous for physical robots because a planner that samples from such a model receives confident-looking predictions regardless of what action it selects. On a real robot, this breaks the feedback loop that keeps the plan grounded: the robot commits torques, the world changes in ways the model never reflected, and subsequent plans are built on an increasingly fictional premise. Unlike a video-game agent that can reset, a physical manipulator or mobile platform pays the full cost of that divergence in hardware stress, dropped objects, or collisions.

Why the control signal loses

Action-blind hallucination typically traces back to one identifiable mechanism, though other contributing factors are possible. Video generative models train predominantly on action-free video, so they learn a strong temporal prior and a comparatively weak action-conditioning signal. At inference time the model interpolates plausible next frames from its learned motion statistics rather than from the supplied action token. The resulting trajectory satisfies the model's frame-to-frame smoothness objective while ignoring the control input. The data imbalance drives this. A typical internet-scraped pretraining corpus holds on the order of billions of action-free frames, while the paired action-observation data used to fine-tune conditioning may cover only tens of thousands of transitions. The model has seen smooth motion roughly 100,000 times more often than it has seen what a specific action token causes, so the learned motion prior overpowers the control signal.

Because this imbalance corrupts an entire rollout rather than a single frame, the evidence you need to catch it lives in the trajectory, not in a summary number. The evaluation artifact should always include trajectories, not just aggregates. Horizon failure is often visible in one trace long before it meaningfully shifts an average score.

Minimal Probe

The following snippet computes the minimum of the three axes directly. That minimum is a better deployment signal than the average because the planner will fail where the weakest property fails.

# Combine consistency, controllability, and horizon into a conservative audit.
# The minimum axis is the bottleneck the deployment team must fix first.
metrics = {
    "consistency": 0.88,
    "controllability": 0.72,
    "usable_horizon": 0.54,
}
bottleneck = min(metrics, key=metrics.get)
print({"bottleneck": bottleneck, "audit_pass": min(metrics.values()) >= 0.7})

Expected behavior: The audit fails because the usable horizon is too short even though the short-term clip looks coherent and somewhat controllable. That is exactly the point of the panel: a planner or evaluator needs a long-enough trustworthy future, not merely an attractive first second.

Practical Recipe

1. Report each axis separately and report the bottleneck explicitly.
2. Keep at least one broken trace in every evaluation packet.
3. Test horizon under repeated actions and under branching counterfactual actions.
4. Do not compare models unless they are scored on the same initial-state panel, action scripts, and acceptance thresholds.

Good evaluation is not an afterthought to world-model research. It shapes what progress means. If the field rewards only photorealism, models optimize photorealism. If the field rewards task-grounded controllability and usable horizon, model design and data curation follow those incentives. A world model that passes the visual test but fails the planning test is not a simulator: it is a screensaver.

This is why reproducibility matters here. A world model fails on one initial state and looks excellent on another. Without saved seed panels and trace artifacts, evaluation collapses into storytelling; the right artifact makes the failure replayable.