Resolve the contact-role contribution ranking

Determine the relative ranking of predictive wrench supervision, Action–Wrench Critic valuation, and local adaptation wrench supervision in Facet-0, since the reported ablation results do not distinguish their individual contributions reliably.

Background

Facet-0 uses wrist-wrench information in three distinct roles: predicting future wrench consequences during semantic–contact alignment, valuing proposed action–wrench pairs with the Action–Wrench Critic, and supervising the auxiliary wrench head during local policy adaptation. The ablation study removes each role separately on RAM insertion and reports only small differences among the resulting success rates.

Because each ablation cell contains 20 trials and the individual removals differ by only one to two successes, the paper explicitly states that the ordering of these three roles is unresolved. A larger or otherwise more discriminating evaluation is therefore needed to determine which contact role contributes most and whether the roles have genuinely different effects.