ClaimNov evaluates novelty at the claim level in AI-scientist proposals

ClaimNov extracts typed claims from proposed research ideas, retrieves related claims from a caller-supplied prior-work pool and produces claim-level novelty evidence. Across four AI-research datasets totaling 1,060 proposals, the authors report 97% detection of disguised papers at a 1.5% false-alarm rate and a reduction in one LLM judge’s false alarms from 11.8% to 1.5%. The public quickstart requires OpenAI and Anthropic API keys; the reported evaluation concerns AI research rather than biomedical or phage tasks.