AI scientific discovery pipelines struggle to judge when an idea is truly new. Existing novelty checks often reject non-novel work but also flag later-published ideas as false alarms. These checks cannot pinpoint which parts of a proposal are non-novel. The paper introduces ClaimNov, a claim-level novelty module. ClaimNov catches 97% of disguised papers with a 1.5% false-alarm rate. It reduces LLM judge false alarms from 11.8% to 1.5%. ClaimNov keeps proposals that build on established contributions. It rejects only 3% of ICLR-accepted proposals with disputed contributions. It keeps 91 of 100 novel proposals with one spliced prior contribution. Code and data are publicly available.

• 1,060 proposals in four datasets
• 97% of disguised papers caught
• 1.5% false-alarm rate
• 11.8% to 1.5% false alarms cut by ClaimNov