Nalin Dhiman, Siddhrath Panwar
We test whether motion direction in a frozen, independently pretrained FLYVIS checkpoint remains linearly accessible when retinal apparent scale is held out. The downstream decoder is supervised on direction labels from the other scales, so the experiment measures transfer of an existing representation rather than autonomous six-way classification by FLYVIS. The analysis unit is a unique rendered condition; moving bars and translating targets form the 192-condition scale-varying set, while geometrically scale-neutral moving edges provide a direction control. Across eight scales, the frozen representation gives leave-one-scale-out accuracy 0.885 (Wilson 95% interval 0.833-0.923). Transfer is nearly perfect across the six interior held-out scales (0.986) but reaches chance at the upper boundary [Formula: see text] (0.167), revealing strong but range-limited scale stability. Capacity and sampling-matched direction-trained temporal models reach 0.983-0.986 with hexagonal, collision-free square-projected, or self-only neighbourhood operators. This establishes raw task solvability and shows that sampling geometry alone does not explain success; unlike FLYVIS, these controls are optimized directly for direction labels. Direct interventions further show that T4/T5 activity contributes to preservation of the intact linear code: after T4/T5 silencing, a refitted decoder retains 0.891 accuracy while the fixed intact decoder falls to 0.193, indicating distributed information in reorganized coordinates. Learned-parameter shuffles reveal realization-dependent code disruption and dynamical fragility. Together, the results establish a reusable cross-scale benchmark, strong interior-scale transfer in a frozen connectome-constrained model, a sharp extrapolation boundary, and specific perturbational sensitivities. Optical-flow pretraining, connectome-constrained dynamics, and learned parameters define complementary mechanistic factors for controlled factorial study.