Resolve the gradient-roster caliber discrepancy

Determine which gradient-based neuron ordering—task-conditioned gradient norm or task-conditioned gradient magnitude—provides the appropriate gradient-roster caliber for module and physics-entry conclusions.

Background

The paper distinguishes an ordering by absolute task-conditioned gradient norm from an ordering by absolute task-conditioned gradient. Their per-neuron correlation is only moderate, and their top-983 neuron sets differ substantially.

Consequently, conclusions about gradient rosters and physics-module entry depend on the choice of ordering. The paper retains caliber labels and postpones selecting a unique ordering.

References

A third, gradient-roster caliber question remains open (Appendix~B, F10): the orderings by $|\mathrm{grad_norm_task}|$ and by $|\mathrm{grad_task}|$ are not equivalent (per-neuron Pearson 0.59--0.71; top-983 sets differ).

— Pre-carved Niches: The Formation Dynamics of Modular Task Partitions in Early LLM Training  (2609.01170 - Li et al., 1 Sep 2026) in Appendix A, “Measurement Discipline for Formation Dynamics”