Independent effects of confidence weighting and deterministic predicates

Determine whether the confidence critic and deterministic predicate filters independently improve return or violation outcomes in the predecessor mechanism matrix for calibration-routed Model-Corrected World Model training.

Background

The predecessor mechanism matrix removes or retains individual components, including confidence weighting and deterministic predicates, while comparing the resulting systems with a frozen gap-routed predecessor. The experiments are intended to distinguish the contributions of these bounded-trust mechanisms from other components such as validation checkpoint restoration, violation costs, and sparse residual modeling.

The reported contrasts for separate confidence and predicate removals do not resolve a control endpoint. Consequently, the paper supports the use of these mechanisms as part of the complete system but does not establish their independent causal benefits.

References

The predecessor matrix supports validation checkpoint restoration on Ant and Walker, avoiding a fixed residual route on Ant, and a return--violation tradeoff from the violation cost; separate confidence and predicate removals remain unresolved.

— Calibration-risk routing for controlled world-model adaptation  (2610.01001 - Zhang et al., 1 Oct 2026) in Section 5.3, “Mechanisms, Capacity, and Inspectability”