Contract-Driven Modular Skill Deltas: An Audit of What Conditional Amputation on a Declared Path Does and Does Not Establish
This paper is a verification, specification, and artifact-audit case study for modular parameter-ownership contracts, not a proposal for a new neural architecture. Building on standard zero-initialized residual and low-rank adapters (ReZero, LoRA-B), we formalize a Base-Skill-Delta contract (C1-C3 under execution state S) and enforce two distinct mechanisms: a manual clause-to-claim licensing map for formal statements, and an executable verification gate (verify_all_numbers.py, 421 checks, 0 failures) that recomputes every reported statistic in the TeX source directly from raw per-seed JSON artifacts. We strictly separate the contract-native system from a sibling 43-expert routed ensemble. The routed ensemble outperforms a dense monolith by +8.31 p.p. at equal expert-update counts (+19.24 p.p. under an expert-only parameter-step proxy), while the contract-native architecture trails a matched monolith by 1.07 p.p. (B03). Applying the audit gate across 27 experimental artifacts exposed evaluation defects that standard reporting hides: we retract an earlier residual-amputation regularity (r = 0.8887 to 0.2883) after isolating an empty-gold scoring bug, distinguish output head-zeroing from latent weight erasure, and document an unattributed 6.75 p.p. gap between two nominally identical configurations (A22 and A24-B) together with checkpoint-replay limits
Authors
- Ilya Gepolov
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24427549
- Primary Topic
- Domain Adaptation and Few-Shot Learning
- Type
- preprint