Large Language Models as Consumer Lab-Report Interpreters: A Design Case Study of an AI Tool for Mold and Biotoxin Testing

People increasingly obtain their laboratory results before, and sometimes instead of, a clinician's interpretation, and a growing share turn to general-purpose large language models (LLMs) to make sense of them. This paper examines a narrower and less studied phenomenon: purpose-built, domain-specific LLM interpreters designed for a single clinical area. We use as our case a free, consumer-facing mold toxin urine test/blood test interpreter that expresses mold and biotoxin test reports, including Environmental Relative Moldiness Index (ERMI) and HERTSMI-2 building scores, urinary mycotoxin panels, and the inflammatory blood markers associated with Chronic Inflammatory Response Syndrome (CIRS). The domain is analytically useful precisely because it is difficult: several of these tests lack consensus clinical reference ranges, results are easily confounded, and public-health authorities caution against over-interpretation. We analyze the tool's design choices, scope narrowing, an explicitly educational rather than diagnostic framing, a no-storage privacy model, redaction guidance, and handling of prompt injection in uploaded documents, as concrete risk-mitigation strategies, and we situate them against the empirical literature on LLM accuracy and hallucination in clinical interpretation and against the United States regulatory boundary between clinical decision support and a regulated medical device. We argue that domain-specific framing, transparent uncertainty, and mandatory human referral are the design principles most likely to make consumer lab interpreters beneficial rather than harmful, and we identify open problems, including the risk that a fluent interpreter lends unearned authority to contested testing paradigms. We close with design and policy recommendations. The authors disclose a direct interest in the tool studied; the analysis is conceptual and does not constitute an empirical evaluation of output accuracy.