Tekmeria: Server-Issued Citation Handles and a Deterministic Audit for Biomedical LLM Answers

Abstract

Language models that answer biomedical questions with citations can cite papers that do not exist, misquote the ones that do, and cite retracted work. Most checks run after the answer is written and must judge whether an imperfect bibliographic match is the right work. We describe tekmeria, an open-source Model Context Protocol (MCP) server for PubMed, Europe PMC, ClinicalTrials.gov and openFDA. Every passage it returns carries a handle whose check code is a truncated keyed MAC over the passage's locator, so the installation that issued a handle can tell a genuine citation from an invented one without a bibliographic search, whichever model or host wrote the answer. A deterministic audit separates malformed, forged, unavailable and withdrawn citations from misquotes, flags retracted, preprint and outdated sources, and passes an answer only when every sentence is cited and every citation is valid. A conformance evaluation plants citations in 34 categories, drawn from seven real records, with ground truth from a reference oracle that the same author wrote separately from the audit. All 1,600 audits and all 257 whole answers come out as planted, across eleven seeds. Injected faults (37 of 38 caught) and a mutation analysis (71 of 78 mutants killed) measure what the evaluation can detect. Two live runs with a frontier model form a pilot. The audit checks provenance and quotation, not whether a passage supports its claim.