Generalization of calibration-reshape patterns across checkpoints

Determine whether the calibration-reshape patterns observed under the State-Reconciliation Intervention and Call-Reconciliation Intervention generalize from the six evaluated model checkpoints to all 26 checkpoints in the main evaluation panel.

Background

The paper evaluates inference-time calibration plasticity using the State-Reconciliation Intervention (SRI) and Call-Reconciliation Intervention (CRI), but applies these probes to only 6 of the 26 checkpoints included in the main model panel. The reported effects vary substantially across model families and intervention mechanisms, including both positive and negative changes in trajectory accuracy. Consequently, the observed reshape patterns may not represent the behavior of the unevaluated checkpoints, leaving their broader generality unresolved.