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

This report collects the negative and bounded results of a preregistered research programme. The programme 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 two Raspberry Pi models. In both setups we measured a real adaptation gain. In both, that gain did not carry over to the next test. It did not survive a different split of the same target's data, a second local target, a different adaptation dose, a wider gap between operating regimes, or a different physical plant. 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 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. At a glance (every value is from the cited evidence packet in the paper): Setup A, activity classifier (final layer adapts) measured gain subject 10: seeds +0.253 / +0.147 / +0.040 (EXP-003); PASS with replay 2:1 (EXP-006) transfer check replay dose -> 4:1 fails by 19 windows on one seed (bounded: one subject) second target -> gain 2 of 81, gate 3 (unsupported, not refuted) resampled splits -> 5 / 6 / 2 of 159, gate 5 (split-unstable, DEC-016) Setup B, adaptive estimator on a nonlinear circuit measured gain largest +0.197 dB (EXP-019); +3.79 dB at separation 1 (EXP-020) transfer check adaptation dose -> gain and cost never hold together (DEC-025, one circuit) regime separation-> peak at the smallest step, gone by step 3 (DEC-026) second plant -> arm does not port; EXP-022 never opened (DEC-031 to DEC-034) Rejected before any result: 3 nominees, 1 battery, 5 environmental constructions; third-plant census 13 -> 0 Instruments retired after measurement: support gate (DEC-018), headroom gate (DEC-027) Physical ladder: Raspberry Pi envelope never binding; Pi 4 and Pi 5 receipts byte-identical The PDF contains Figures 1 to 4: a study map, a two-setup comparison, the transfer-limit ladder, and a one-page summary of what the record supports and what it does not. Every figure value is copied from the cited packets. Files: manuscript.pdf (version 1.1) · manuscript.md (source) · figures.tar.gz (the four figures as SVG) · build.sh (reproduces the PDF). Version 1.1 corrects the author affiliation, rewrites the prose in plainer English without changing any bound, and adds the figures. Results, numbers and citations are unchanged from version 1.0 (record 22541660).