DB-Curator: Rubric-Guided Evidence Acquisition for Scientific Database Construction

Constructing chemical databases is time-consuming because evidence linking compounds to exposure, biological origin or reaction conditions is dispersed across a vast literature. We present DB-Curator, a deep-research agentic framework whose user-approved rubric guides both evidence acquisition and row-level decisions. Topic-level retrieval discovers compounds, while a Researcher--Critique loop tracks criterion states and targets unresolved requirements. Across five compound-curation tasks, DB-Curator evaluated 1,743 candidates and retained 1,000 rows, each with at least one attributed passage found in its cited paper. Passages were located for 8,343/8,493 citations in these retained outputs, versus 0/991 in retrieval-free baseline outputs without citation-based selection. On 1,887 supplied candidates, DB-Curator produced 970 system-Confirmed outcomes, versus 375 under zero-shot labeling. Three-repeat comparisons showed that adaptive evidence construction supplied most of the yield gain, followed by smaller, bidirectional revisions during verification. These results support rubric-guided evidence acquisition for chemical database curation.