Validate the Dream Before You Trust Its Verdict: Admissibility for World-Model Simulators

World Models (WMs) are increasingly used as 'test oracles' in robotics and autonomous driving, where they simulate the consequences of an agent's actions to provide a safety or success verdict. However, there is a fundamental 'trust inversion' compared to classical simulation. Traditional simulators are engineered and validated against physics, making them trusted components that evaluate untrusted policies. In contrast, generative WMs are learned artifacts that lack ground-truth physics, meaning the simulator itself is unverified. Relying on these models for safety-critical decisions without a formal accreditation process is inherently risky.

To address this, the authors propose a five-level 'admissibility ladder' (L0-L4) that a WM must climb before its verdicts can be trusted as evidence. This framework adapts established safety-critical practices like Verification, Validation & Accreditation (VV&A) and Safety of the Intended Functionality (SOTIF) to the generative setting:

• L0 (Generation Quality): Focuses on visual and temporal realism (e.g., FVD scores). This is a minimum viability gate but provides no evidence of action-following capability.
• L1 (Action-Robust): Requires that the simulated world responds systematically to different agent actions, ensuring the model is not just replaying static footage.
• L2 (Envelope-Declared): The model must define its training envelope and implement out-of-distribution detection to refuse inputs where it cannot reliably predict outcomes.
• L3 (Failure-Attributable): The system must be able to distinguish between failures caused by the policy under test and failures caused by the simulator itself.
• L4 (Verdict-Transfer-Validated): The highest level, requiring empirical evidence that performance in the simulator correlates with performance in the real world within the declared envelope.

The authors demonstrate that visual quality (L0) is a poor proxy for action-robustness (L1-L2). By applying their framework to two driving WMs, they show that models ranking high in visual generation often perform poorly in action-following. This suggests that current industry reliance on visual fidelity metrics is insufficient for safety assurance. The proposed ladder provides a principled, embodiment-agnostic path for researchers to move beyond 'plausible-looking' simulations toward verifiable, evidence-based testing.