Abstract
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. This paper makes two things measurable, whether abduction is worth performing in an environment and where a trained model loses its answer. The abducible share is the share of the variance of the individual treatment effect that the history reveals and the branch-point observation does not. It comes with an explicit estimation floor below which no value of it is interpretable. Six of our twelve environments cannot support the measurement, since each fails one of five conditions we state on the environment and the estimator, and the parameterisation of the estimand alone takes the abducible share from 0.06 to 0.79. We propose reporting the abducible share as a certificate carrying its floor and its settings, and the calibration of a model as a certificate of calibration slope, order and error. Substituting a trained model’s own quantities into the same decomposition locates where the individual effect is lost. In a reservoir computer, the latent state of a recurrent state-space model (RSSM) retains 0.195 of an abducible share of 0.213, whereas its imagination rollout returns 0.125. In a two-qubit processor the imagined state keeps this information over twenty imagined steps while the decoded output loses it, which places the loss on the output side of the rollout.