Where Individual-Effect Information Survives: A Diagnostic Ladder for World Models

World models are widely used for planning and control, but it remains unclear whether they can answer counterfactual questions of the form ``what would have happened to this individual under a different action.'' Such a question sits at the third rung of Pearl's ladder and requires abduction, inferring from the history what distinguishes this individual from the others. Whether abduction is worth performing at all is settled by the environment before any model is trained, and this paper makes that determination measurable. The abduction gap is the share of the variance of the individual treatment effect that the history reveals and the branch-point observation does not, and it comes with an explicit floor below which no value of it is interpretable. Three findings follow. First, six of our nine environments do not admit the question. They fail one of five conditions we state on the environment and the estimator, and the failures are of four kinds, two of which belong to the measurement. Second, the answer depends more on how the measurement is set up than on which environment is measured. The parameterisation of the estimand alone takes the gap from 0 to 0.63, which exceeds the spread across environments. We propose reporting the gap as a certificate carrying its floor and its settings, and the calibration of a model as a certificate of slope, order and error. Third, substituting a trained model's own quantities into the decomposition of the ladder locates where the individual effect is lost. In lithium-ion degradation the environment admits 0.300 and the latent state of a recurrent state-space model trained on reconstruction alone retains 0.304. Its imagination rollout returns almost none of it, while a readout head on the frozen latent state recovers most of it. The individual effect is abduced and then discarded in rollout, and where it is discarded is measurable.