Code and per-draw results for: Evaluating few-shot model updating in construction-materials machine learning: support–query separation at the unit of deployment

Analysis scripts, sampling manifests and per-draw results for a re-analysis of three public construction-materials datasets (D16 interlayer bond strength of 3D-printed mortar; BOxCrete multi-age concrete compressive strength; Crack22 crack segmentation in 3D earthen printing). The archive contains the fixed-query, matched-budget comparison of support samples drawn from the scored physical units versus other units, the few-shot calibration experiments under specimen-, object- and whole-object protocols, the Crack22 split protocols, fine-tuning arms and cluster bootstraps, and the script that draws all data figures. The datasets themselves are not redistributed; README.md gives their sources and the expected folder layout, maps every table and figure of the article to its result file and script, and lists the software versions used.