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

We report the negative and bounded results of a preregistered research programme on small-model local adaptation under a constrained compute and memory envelope, executed on a fast environment and replicated on two Raspberry Pi classes. The question throughout was whether a small adaptable model can improve on a defined local target while retaining previously acquired capability, against a frozen-model control measured under the same conditions. On two very different apparatuses — a human-activity-recognition classifier adapting its final layer, and a linear-in-parameters estimator adapting on a nonlinear electronic circuit — a measurable adaptation gain was found and then failed to transfer: across resamplings of one target's windows, across a second local target, across adaptation doses, across the separation between operating regimes, and across plants. Every candidate construction examined at the pre-result stage — three tournament nominees, one battery construction, five environmental constructions — was rejected before any adapted outcome was observed, on independence, support, or identifiability grounds, and a frozen census for a third plant admitted none of thirteen candidates. Two of the programme's own instruments were found to select against their objectives and were retired. The physical hardware envelope was never the limiting variable, and the frozen code produced byte-identical decision receipts on both Pi classes. We state what each result shows on its own frozen contract, what it does not show, and why every finding remains a single-apparatus result rather than a general one.