AVM-1: Proof of Principle for an Independent Bibliographic Verification Module for LLM-Generated References

Abstract Large Language Models (LLMs) used for assisting research can produce plausible bibliographic references containing errors in title, author, year, or DOI, degrading paper quality or imposing heavy verification burdens. To address this issue, this paper introduces AVM-1, an independent Python program designed to interface between the LLM and the user to verify citations against Crossref-registered metadata. In the present implementation, the program receives a candidate reference, typically generated by an LLM, consisting of a title, a DOI, and optionally a publication year and an author name. It queries the Crossref REST API using the DOI as a lookup key to retrieve the registered metadata, then compares the supplied title, year, and author against this external record. As the DOI itself may be missing, fabricated, or wrong, the program assigns one of the following three non-interchangeable states. It outputs VERIFIED when no contradiction is detected, INCONSISTENT when the resolved record contradicts the supplied fields, and BLOCKED when no Crossref record can be resolved. Title comparison relies on a metric combining word-frequency cosine similarity and token-order sequence matching. Tests in Google Colab show that the module detects several bibliographic inconsistencies, although paraphrased titles exceeding the similarity threshold may remain undetected. The test set includes one unmodified case: a reference generated by an LLM and published by the author, whose DOI resolved to an unrelated article; the module reports it as INCONSISTENT. This work serves as a functional proof of principle rather than a statistical validation of general system reliability. To facilitate reproduction and practical testing, the accompanying Python source code features an interactive Google Colab widget mode. It allows users to verify custom references directly within defined operational limits.