Hallucinated references are the bane of academic paper review — a citation that reads perfectly and points at nothing. RefChecker automates the check: it extracts references with GROBID or an LLM, then verifies each one programmatically against authoritative sources — CrossRef, PubMed, Google Books — instead of trusting a model’s word for it.
Verification returns a four-level confidence classification per reference, falling back to LLM-based ambiguity resolution only when the authoritative sources can’t settle a match. Papers process individually or in batches through a Tkinter GUI.
The architecture is deliberately modular: extraction, verification sources, and reporting are plugins, so anyone can add a new source of truth or extraction strategy without touching the core. Built in Python, developed with Claude Code driving an agentic workflow against a personal git system for code and context management.