Distinguishing Cross-Backbone Decision-Behavior Transfer from Frozen-Core Portability

DecPort investigates two related questions: whether input-specific probabilistic decision behavior can be transferred across heterogeneous frozen language-model backbones without target labels, and whether a particular learned frozen DecisionCore contributes reusable functionality after the backbone change. The study progresses from supervised shared-head transfer through label-free latent alignment and decision-space distillation to two accepted Open-Jev stages. Decision 0006 establishes reproducible, input-specific cross-backbone behavior distillation, while exposing a structural identifiability problem in the linear-core design. Decision 0007 addresses that confound with a nonlinear learned DecisionCore, a constrained rank-128 target adapter, a scale-matched random nonlinear core, mismatched-teacher and untrained controls, and a target-specific distilled baseline under preregistered SHIP/NO-SHIP criteria. The final gate returns NO-SHIP for the original reusable learned-DecisionCore hypothesis. At the same time, input-specific behavior transfer remains positive across all three tested target backbones, with Choice and Score carrying most of the effect while Noul remains problematic. The combined evidence supports label-free decision-behavior distillation within the tested setting, but does not support useful reusable learned-core portability.