From Complementarity to Selectability: Evaluating Hybrid Retrieval for Standard-Informed Furniture Carbon Screening

Concept-stage designers make carbon choices before complete product evidence exists, yet their answers must be computable and traceable to governing standards. Hybrid retrieval combines these functions: a knowledge graph can execute carbon calculations, while text retrieval recovers conditional standards guidance. We test whether complementarity, in which sources succeed on different cases, translates into selectability: query-time evidence may indicate which source to trust. A knowledge graph and a GraphRAG text index were evaluated on 100 furniture-carbon questions across five primary conditions, with oracle and routing analyses. Retrieval value depended on task class: correctness improved by 0.5600 (cluster-bootstrap 95% CI [0.4487, 0.6583]) over the closed-book condition for factual questions (factor lookups, product totals, and material lists), but by only 0.0160 (95% CI [−0.0986, 0.1420]) for standards interpretation. The stores reversed rank across categories. The observed correctness difference between concurrent fusion and the graph alone was 0.0007 (95% CI [−0.0532, 0.0628]). A cross-fitted oracle exposed recoverable headroom, with category routing examined through both a score-pool diagnostic and a separate live evaluation on the same benchmark. Concurrent fusion recovered little of the source-selection headroom. The evaluated outputs are material-stage screening estimates from author-compiled bills of materials. Designers and retrieval-system developers should prioritize structured access to exact factor fields, unit conversions, and linked inventory data for record-specific questions, using standards indexing to provide inspectable clause context for conditional interpretation.