Toward Verifiable Medical Imaging World Models: An Engineering Framework for Evaluating Simulation Fidelity, Roll-Out...

Medical world models simulate how a patient's imaged state evolves over time and, in some systems, under clinical interventions. Their value comes from multi-step, action-conditioned roll-outs, yet evaluation is still dominated by perceptual or single-step metrics. We treat assessment of these models as a verification problem, distinct from later evidence of clinical benefit, and organize it around three axes. Simulation fidelity asks whether predicted states are correct at the level claimed by the model. Roll-out stability asks whether that correctness is maintained over multiple steps; we summarize this using an effective planning horizon, defined as the last step before a stated loss quantile first exceeds a clinical tolerance under the observed action policy. Clinical plausibility and reliability ask whether simulated responses are counterfactually consistent, calibrated, and adequate across patient subgroups. We also propose a claim-proportionate reporting standard and a reproducible procedure for estimating the horizon from retrospective data. Applying the framework to a purposively selected sample of nine imaging world models, we find that reported evidence is uneven across capability tiers. Under the proposed standard, every model claiming action-conditioned prediction or above omits at least one applicable item: none reports an explicit counterfactual-consistency check or a clinically meaningful subgroup analysis of the type specified in Item 7, and none reports an effective planning horizon at a stated clinical tolerance. Because the effective planning horizon is estimated on-policy, it should not be interpreted as evidence of off-policy counterfactual reliability.