Adjudicate which probe caliber defines module identity

Determine whether attribution rosters or loss-decomposition working sets should define module identity for the analyzed language tasks and their loser module.

Background

The paper uses multiple localization and probe calibers that produce materially different neuron sets. For the adjective task, the attribution roster is highly fluid, whereas the loss-decomposition working set is comparatively stable and gradually converges toward the attribution roster.

Because these measures support different interpretations of module membership, the paper does not claim that one is authoritative. Resolving the issue is necessary for interpreting module identity, roster stability, and the apparent weak learning of the loser task.

References

Which caliber defines module identity is an open question (Appendix~B, F11; Appendix~C).

— Pre-carved Niches: The Formation Dynamics of Modular Task Partitions in Early LLM Training  (2609.01170 - Li et al., 1 Sep 2026) in Section 6.1, “Gradient-level deprivation”; Appendix B, F11; Appendix C, R2-H1-1