spb/wpai0001_uqo

1\section{Conclusion}\label{sec:conclusion}2We set out to test whether a causally identified subsystem can be transplanted between language models3without training. In a micro-world built so that the question is identifiable, it can: one orthogonally4mapped residual row, fitted on task-free data, moves a rule from a donor to a recipient that never learned it,5with the full battery of controls at chance and a preregistered confirmatory test passed on the primary6contrast. The same experiments draw the boundary: regression-fitted maps annihilate what the recipient7cannot predict, donor-free variants cache rather than compute, and conventional learning is unbeatable when8labels are available. On open-weight models the same protocol finds task vectors that are causally sufficient within a model for antonyms and capitals, transfers them between models above null controls in some pairs and directions under a frozen rule, and fails for tasks that no model represents as a vector; the honest summary is partial transfer of task identity, not acquisition of a capability. Everything needed to reproduce9or refute these statements---generators, models, hooks, registry, manifests and analysis scripts---is in the10repository.11