Route-level effects on gravity and Walker performance

Determine whether automatic calibration-risk routing between residual and direct target world-model candidates changes return or violation outcomes for the HalfCheetah gravity and Walker friction shifts relative to the corresponding unselected forced routes.

Background

The exact-routing counterfactual fits residual and direct candidates under matched data, initialization, and isolated random streams, then compares automatic deployment of the lower-calibration-risk family with forced deployment of the alternative. Automatic routing selects the direct family for all gravity runs and the residual family for all Walker runs.

The Ant contrast yields a statistically resolved return benefit, but the corresponding gravity and Walker return contrasts have confidence intervals containing zero. The violation contrasts also remain statistically unresolved for all three environments, so the causal effect of routing on these endpoints is not established for gravity or Walker.