ReproCheck is an AI research reproducibility investigator that traces important claims from a research paper against evidence available in its implementation repository.
Instead of simply generating a summary, ReproCheck creates a Claim → Evidence analysis, classifying claims as Supported, Partially supported, Not verified, or Conflicting evidence. It also evaluates dataset reproducibility, identifies reproducibility blockers, highlights paper/code inconsistencies, and calculates evidence coverage.
ReproCheck uses the open-weight Qwen3-4B-Instruct-2507 model through Hugging Face. The system is designed to be conservative: missing evidence is treated as unverified rather than being presented as proof of reproduction.
Built with: Next.js, TypeScript, Tailwind CSS, Hugging Face Inference Providers, and Qwen.