Where a measured local-adaptation gain did not transfer: negative results from a preregistered small-model adaptation programme on constrained hardware

Abstract

This report collects the negative and bounded results of a preregistered research program. The program asked one question: can a small model, adapted on a device with limited compute and memory, beat a frozen copy of itself on a local target while keeping what it already knew? We tested this on two very different setups. The first was an activity classifier that adapts its final layer. The second was a linear estimator that adapts on a nonlinear electronic circuit. Each setup ran in a fast environment and was replicated on Raspberry Pi hardware: Setup A on a Raspberry Pi 5, Setup B on a Raspberry Pi 4 and a Raspberry Pi 5. In both setups we measured a real adaptation gain, and in both it did not carry over to the next test. In Setup A, the passing result was unstable across splits of the same target's data, and on a second target the recipe was unsupported, though not refuted. In Setup B's dose sweep, no dose met both the gain and the retention condition. A later sweep found a detectable gain of +3.79 dB with no detectable retention cost at the smallest regime separation, of unresolved usefulness, and it was gone by the third of four separations. A test on a second physical plant never ran, because the inherited estimator failed a memory check before any result. We also examined several candidate setups before looking at any adaptation result. These included three tournament candidates, one battery setup, and five environmental setups. All were rejected because the data were not independent enough, the evidence base was too small, or the effect could not be identified clearly. A separate search for a third physical plant checked thirteen candidates and admitted none. Two of our own measurement tools turned out to select against their own purpose. Both were retired. The hardware envelope of the Raspberry Pi never limited any result, and for Setup B the frozen code produced byte-identical decision records on both Pi models. For each result we state what it shows under its own frozen contract, what it does not show, and why every finding still rests on a single setup rather than a general rule. Files: the typeset paper (Costa_2026_Adaptive_Edge_Transfer_Limits_v1.2.pdf), the same manuscript as open Markdown (Costa_2026_Adaptive_Edge_Transfer_Limits_v1.2.md), and one reproducibility supplement (Costa_2026_Adaptive_Edge_Transfer_Limits_v1.2_supplement.zip). Version 1.2 resets the report as a typeset paper with plainer prose, new figures and tables, a list of terms, and a reproducibility supplement. It also corrects passages that the committed records contradict or had superseded; the version history at the end of the paper lists each one. Nothing is re-scored.