Evidence-grounded decision support for cleantech innovation scouting: a system-oriented design-science study

Introduction Cleantech innovation scouting requires analysts to integrate heterogeneous evidence while preserving traceability. We present GRaph-Augmented INnovation Scouting (GRAINS), a system-oriented design-science artifact that connects governed evidence preparation, hybrid retrieval, optional graph expansion, and staged cite-or-abstain synthesis. Its contribution combines an inspectable evidence-to-answer architecture, an explicit rationale linking design choices to scouting requirements, and empirical observations of system behavior and evaluation sensitivity. Methods We evaluate GRAINS through an exploratory comparison on a source-derived, in-corpus benchmark of 50 broad-cleantech queries with relevance judgments from two large language model judge pipelines and a separate exploratory case-based audit of six battery-domain cases assessed by one researcher. Results Hybrid achieved Recall@100 of 0.951 and Precision@10 of 0.278 using the combined relevance judgments of the two judge pipelines, while none of the tested retrieval contrasts reached statistical significance after global Holm correction. Comparisons of top-10 ranking quality (nDCG@10) depended on the judge pipeline and whether query-source documents were excluded. Limiting the number of retrieved chunks per document increased document diversity with a modest reduction in nDCG@10; the evaluated graph-expanded configuration had lower observed early-ranking scores than Hybrid. Structural checks confirmed citation-identifier compliance in evaluated follow-up answers. The case audit illustrated how the system organizes supported actors, applications, and uncertainties, while revealing gaps in task-specific evidence. Discussion Together, these findings provide bounded system-design observations on the behavior and evaluation of evidence-grounded scouting, while independent human relevance assessment and analyst-workflow validation remain necessary for broader validation.