Evaluating the Physicality and Layer-Wise Decodability of Causal Concept Steering in Scientific Foundation Models

Physics-foundation models provide efficient surrogates for physical systems governed by ordinary and partial differential equations, but it remains unclear whether their internal representations encode general physical principles. Recent work shows that Walrus can be activation-steered to induce phenomena such as vorticity. We ask whether the resulting rollout reflects an initial perturbation, a modified but self-consistent version of physics, or a cosmetic shift. In this work, we evaluate activation steering of vorticity in Walrus on \texttt{shear\_flow} and evaluate the physicality of these causally steered representations. Using finite-difference metrics of momentum residual and divergence, we reveal that steering acts more as a continuous forcing term rather than an initial condition shift, with full vorticity response requiring repeated injection rather than a one-time intervention. This supports our hypothesis of a modified but self-consistent version of physics. Furthermore, we identify a practical low-strength range in which continuous positive steering with $\alpha \leq 0.4$ induces vorticity while divergence and residual remain roughly twice their unsteered values and observe that mild negative steering yields small, frame- and metric-dependent reductions in these diagnostics. Finally, we test whether a reference study’s 18 vortex and 10 laminar Reynolds-Schmidt parameter-pair categories used to construct the steering direction are linearly decodable across Walrus's processor layers with a layer-wise probe. Using held-out complete Reynolds-Schmidt parameter pairs, no spatially averaged layer representation exceeds chance performance. Thus, these regime categories are not linearly decodable under the tested evaluation. Code and data will be made publicly available upon acceptance.