Coverage, Uncertainty, and Evaluation in Learned World Models: Identification Limits and Validity Conditions

Claims that hallucinations in learned world models are predictable and preventable requiredistinct assumptions about statistical identification, distribution shift, uncertainty measurement,and evaluation. This paper develops a framework for separating these questions, using a recentgenerative-world-model study and TD-MPC2 as case studies. A two-point argument identifiesthe limits of inference on unobserved contexts without implying that realized error is monotonein data density. A rollout bound separates estimation error on the sampling distribution,occupancy-ratio amplification, and deployment mass singular with respect to that distribution;complementary Wasserstein and pathwise formulations clarify geometric approximation andpersistence of errors. We distinguish sampling-seed spread from calibrated epistemic uncertainty,establish the invariance of within-task AUROC to task-constant normalization, and characterizelimitations of round-trip, flow-instability, and action-shuffle diagnostics. We provide explicitassumptions for importance-weighted validation and conformal false-alarm control, togetherwith a separate certificate for accepted-prediction risk. For implicit models, corrected planningbounds, a scale-equivariant disagreement penalty, and local SimNorm gradient bounds delimitwhat short-horizon planning and latent normalization establish. Executed synthetic diagnosticsillustrate aggregation effects, disappearing terminal discrepancies, and failure of calibrationunder unsupported shift. They are not replications of the benchmark studies. A controlledexperimental protocol specifies the evidence still needed to distinguish coverage, representation,optimization, and capacity effects.