Recompile, Don't Transport: Empirical Validation of Data-Driven Personalization Updates on Edge-Class Hardware

Adapter-based personalization (e.g., LoRA) does not survive base-model change. We test this
directly: across 12 users from the Enron corpus and two base models from the same family
(Qwen2.5-1.5B-Instruct → Qwen2.5-3B-Instruct), adapter transport fails universally with tensor-
shape mismatch (12/12 users). Recompiling a fresh adapter from the same user data on the new
base fully recovers personalization at 250 training examples and exceeds the original adapter at 500
examples (fingerprint distance 0.83 vs 0.86; authorship probability 0.51 vs 0.44). Recompilation
takes 47–163 seconds on an A100 and is operationally viable on edge-class hardware. We argue
the durable asset for on-device personalization is therefore the source data and a frozen per-user
evaluation set, not the adapter — a requirement mechanically incompatible with closed-weight
deployment. This is the second paper in a research program arguing that durable personalization
is one of several structural reasons the on-device LLM future will be open-weight.