Fires Precisely, Not Early: A Pre-Registered Geodesic Trust Signal for World-Model Rollouts

Learned world models accumulate error under imagined rollouts, but an agent needs a defensible rule for deciding when a rollout is no longer trustworthy. We evaluate this question in MiniGrid environments whose complete state graphs, true dynamics, geodesic state distances, and planning outcomes are exactly computable. Our Displacement Horizon (DH) combines expected geodesic displacement over reachable predicted states with mass on states from which truth is unreachable; all thresholds are pre-registered and fixed before the decisive experiments. Against ensemble disagreement, DH has a substantially cleaner operating point: on Empty-16$\times$16 it flags 100 failed plans with no false positives, whereas the ensemble baseline has 0.19 precision. The result does not make DH a deployable winner: it is oracle-informed during evaluation, ranking results are equivalent within the small-$N$ noise floor, and signal-gated replanning shows no detectable benefit. We contribute an exact-ground-truth testbed, a pre-registered comparison protocol, and a negative actuator result that limits the method's claim.