Where Do the Gains Come From? An Audit of Dendritic Context-Gated Protection for Continual Learning

Cortical pyramidal neurons integrate contextual signals in dendritic compartments, and networks with active dendrites inherit this context selectivity. In preliminary, unpublished experiments we proposed DeLP (Dendritic Localized Protection), which freezes dendritic segments for non-current contexts and weights an elastic penalty by each neuron's measured usage under previous tasks; on Permuted-MNIST it appeared to dominate every baseline in both forgetting and accuracy. This paper is the audit we should have run before publication. Prompted by an anomaly — a baseline whose penalty was insensitive to a 100-fold change of its only hyperparameter — we found two defects in our MAS implementation, three deviations from the SI reference, a fixed-fraction pruning schedule that exhausted PackNet's capacity, a missing plain-EWC arm, hyperparameters selected on the test set, and λ grids that left every baseline in its protection-free regime. After correction, DeLP keeps the best five-task accuracy (80.37%, next best 75.27, p = 0.0024) but no longer leads on forgetting, and at ten tasks corrected EWC and MAS exceed its accuracy. A seven-rung mechanism ladder localizes what remains: with baselines given the same hidden activation, the entire dendritic chain adds +0.37 points at five tasks (p = 0.49); the one mechanism that survives a parameter-budget cap is usage weighting (+1.79, p = 0.005 at twenty tasks), worth +4 to +6 points once both sides are gain-calibrated. A hidden-layer gain scalar alone moves accuracy by 16 points — more than the entire method — so the original comparison measured an uncalibrated gain, not a mechanism. Finally, preregistered experiments in a rule-switching task matched to the neuroscience show that compartmentalized routing itself confers no adaptation advantage larger than 30% — the sensitivity limit of 8 preregistered seeds — across three cue-conflict intensities, up to and including fully adversarial cues. We report the corrected accounting as the contribution: when benchmarking protection methods, the gain must be declared, the grid must bracket the optimum, and the baseline must be implemented as carefully as the method.Code: https://github.com/zhouranxi/delp-audit