Segment Specialization in Active-Dendrite Continual Learning: An Audit, an Artifact, and a Causal Repair

Active-dendrite architectures route each input through a competition among dendritic segments, one of which is trained per task; usage-localized elastic weight consolidation (DeLP) protects what has been learned by scaling per-neuron Fisher anchors with segment usage. Both mechanisms rest on an implicit assumption that, to our knowledge, has never been measured: that the segment-task mapping constitutes a specialization that protection can leverage at sub-neuron granularity. We audit this assumption on Permuted-MNIST and Permuted-Fashion and report three findings. First, a granularity ladder of anchors (plain, neuron-resolved, segment-resolved with attribution-tracked reference states) shows that the entire benefit of usage localization is realized at neuron level: segment-resolved anchors are statistically indistinguishable from neuron-resolved anchors across two datasets, two task counts (T=10,20), and two elastic weights (all |Delta l| <= 0.24, n=10 paired tests n.s.), and remain so under FIFO anchor budgets (|Delta l| <= 0.44 across eight budget cells). Second, the audit explains why: task-dominant gating is absent - among active neurons the probability that a neuron's dominant segment is its own task's segment starts each phase at chance and is pushed below chance by that phase's own training (T=20: 0.05 -> 0.01), and Fisher mass is diffuse across segments (top-1 share 11-18%). The audit also uncovers a measurement artifact relevant to the active-dendrite literature: under production sampling budgets 56-63% of neurons are never activated, and naive argmax instrumentation over all neurons fabricates a strong "segment-0 dominance" (approx. 60%) that vanishes under active-only filtering. Third, because the missing specialization is load-bearing, restoring it causally improves the retention-plasticity trade-off: a constant bias on the own-segment dendritic input raises the learning term in every cell of a {dataset} x {depth} x {granularity} grid (Delta l > 0 in 20/20 dose cells, significant in 18/20 at n=10 paired), with final accuracy improving in step (17/20 simultaneously significant, and every grid cell significant on both terms at some dose; Delta l up to +4.6, t(9)=8.1); forgetting is reduced at T=10 (Delta BWT up to +1.7) and mildly traded back at T=20 (significant in 3/8 cells). Signed-drift instrumentation shows the network never counteracts the bias at the parameter level; the own-segment win rate saturates at 0.13-0.19 regardless of dose because any effective dose dominates the raw gate scale - partial restoration suffices. Segment-task specialization in this architecture family is thus not a weak tendency but an actively violated assumption - and a repairable one. This record is a timestamped preprint deposited to establish a citable public record independently of journal acceptance. Files included: main.pdf (full paper), LaTeX source, four figures, the Q8 data-closure report, the submission cover letter, and the figure-generation script. Companion code and raw per-seed records: https://github.com/zhouranxi/delp-anchor-budget . Related prior work by the same author (the method audited here): https://doi.org/10.5281/zenodo.23602281 . No dataset DOI is claimed in this record.