Withdrawal Constrained Transfer for Revising Rollout Support in World Model Planning

Predictive world models organize candidate futures, yet execution experience can change which futures should retain deployment authority. This paper studies revision of a rollout admission rule rather than construction of a new planner or safety objective. Withdrawal Constrained Transfer begins from a rule that is no more permissive than a fixed predecessor and returns authority along an ordered family fixed before calibration outcomes are revealed. The key asymmetry is that removing an unsafe predecessor admission can only decrease relative unsafe acceptance, whereas restoring an unsafe admission can increase it. A paired accounting identity turns these effects into one bounded monotone loss, and a finite sample conformal correction determines how far accretion may proceed while keeping expected unsafe acceptance no greater than that of the predecessor under exchangeability. A PointMaze instantiation uses a frozen predictive world model and an execution derived transition graph to construct the revision family. Independent rollout evaluation reproduces the relative reduction after the rule is frozen. A separate shared candidate planner shows nonzero recovered authority together with fewer system contacts than the predecessor. The planner evidence remains empirical because planner selection changes which cases are tested. The resulting formulation isolates support revision as a statistical problem between learned prediction and execution.