Characterize the ordering of harms from different skill-bank perturbations

Characterize the relative severity of harmful skill rewrites and meaning-preserving rewrites across perturbation axes beyond the ordering predicted by the minimal linear-Gaussian construction.

Background

The minimal linear-Gaussian construction predicts that a harmful rewrite, modeled as reversing the surface-form channel, causes greater degradation than a meaning-preserving rewrite. The authors explicitly state that this ordering was not tested empirically and that finer orderings among the perturbation axes remain unresolved.

References

We do not test this ordering, and flag it as a limitation of the instance: it isolates what a two-channel account can capture and leaves finer orderings among the axes open.

— Evolving Together, Failing Apart: The Skill-Policy Co-evolution Trap in Self-Improving Agents  (2610.20845 - Feng et al., 14 Oct 2026) in Appendix D.3, Remark 4 ("What the construction does not explain")