One Degree of Freedom Moves the Frontier, the Other Four Do Not: A Preregistered Factorization of an Active-Dendrite Repair, Its Architecture-Bound Supply, and Five Failed Escapes
This record (version 1) deposits the manuscript and the complete reproduction package for a preregistered factorization of an active-dendrite continual-learning repair into its protection-budget degrees of freedom. Across 1,213 newly collected training runs plus 60 reused controls, five preregistered escapes were tested and four failed: the strength of the protection, its temporal decay schedule, a forward-pass-only intervention, and the context pathway leave the frontier unchanged (R1: 24 Holm-corrected cells, minimum p = 0.585). The one degree of freedom that works -- reallocating a fixed total anchor budget toward recent tasks under exact weight conservation -- moves the Pareto frontier on its home MLP architecture (T=20: +8.08 accuracy over the un-repaired baseline while improving retention over the repair itself, Holm p = 6.4e-5, four preregistered criteria met in full), with direction specificity confirmed by a preregistered reverse-weighting arm (worse on retention in all 20/20 paired seeds, |d_z| = 4.0). The supply of this freedom is architecture-bound: on a calibrated ResNet-18+GN backbone the same reallocation reaches neither of the preregistered verdict levels, after a first-round apparent transfer (+1.94, Holm p = 0.041 at n = 10) was overturned by the preregistered sample-extension clause (n = 20: +0.72, p = 0.34) -- a flip reported in full in the main text. On a convolutional backbone, a preregistered lambda sweep (six points, 240 runs) shows that the companion audit's non-transfer was partly a single-point artifact on the deepest trough of a non-monotonic dose curve, while the recalibrated effect remains an order of magnitude below its MLP value. Every experiment ran under a written preregistration with implementation-correctness gates whose failure voids the data; six gates fired across the program and none was resolved by relaxing a criterion; all six are tabulated in the paper. Two independent machine-checkable verifiers are included and run from the package root with the standard library alone: r_repro/verify_package.py (25 checks) recomputes every headline statistic from the raw per-run records, and paper/check_paper_numbers.py (17 checks) additionally re-derives the seed-direction counts, per-dataset splits, figure-caption numbers, every protocol-table constant, and the 1,213-run ledger, asserting agreement with the manuscript text. The package contains the LaTeX source and compiled PDF, the figure-generation script and all figures, the twelve local modules used by the R1-R3 runners, the raw per-run JSON records of all five experiments (results_r1, results_r2, results_r2t, results_r3, and the reused audit controls results_e9), and the preregistrations with dated amendments and collection reports. Code and records are also archived in the project repository linked under related identifiers.
Authors
- Ranxi Zhou
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24148620
- Primary Topic
- Domain Adaptation and Few-Shot Learning
- Type
- preprint