PaperStack-G: An Evidence-Grounded Research Intelligence System Linking Funding Awards to Scientific Findings
Year: 2026
Type: preprint
Abstract: Connecting funding records to scientific results requires evidence across award records and publications. PaperStack-G Lite integrates provenance-labelled award–publication links, retrieved passages, extracted findings and explicit evidence gaps in a literature system. A reported funding association does not establish that an award financed a result. We evaluate the integration on 100 UKRI AI-topic awards, using 56 synthetic test questions whose references the author verified before any system answered. Conditions share a corpus, answer model and context budget; primary comparisons are Holm-corrected. Explicit links increased complete-chain retrieval recall from 0.53 (95% CI 0.39–0.67) to 0.91 (0.80–0.97), each over 47 eligible questions. Under the corrected primary Codex judge, the complete system improved citation support over the linked baseline by 0.054 (95% CI 0.029–0.081; 43 paired questions). Correct-or-partial answer improvement and the missing-evidence ablation were inconclusive at this sample size, as were exploratory findings and statement-check ablations. Correcting omitted judge context substantially reduced the apparent citation advantage. Four linked-baseline answers remain unjudged after prompt-hash mismatches; the completed Opus second-judge pass preserves the pre-registered comparison decisions. These results support an evaluated software/demo contribution, not a demonstrated methods advance. Larger independently annotated evaluations and another answer model are needed to test generalisation.
Author: Mamun, S M Abdulla Al
Institution: University of Manchester
Language: English
Subfield: Library and Information Sciences
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7236753164
