Traceable Is Not Supported: A Layered Metabolite–Disease Evidence Agent and What Its Validator Missed

Scientific agents use large language models (LLMs) to search databases and summarise evidence. To prevent hallucinations, developers often check whether the names and numbers in the summary can be traced to specific retrieved records. But traceability is not the same as correctness. We built MetaboCausal, an agent that collects evidence on whether a metabolite affects the risk of a disease. It also finds the effector genes that change the metabolite's level, and the existing drugs, compounds and clinical trials for these genes. In the agent, code builds an evidence card, and an LLM only writes a short summary, which a validator checks. The design blocked obvious hallucinations. When it could not read an input, it said so instead of guessing, and it blocked invented numbers. But one error got through: the summary called GCSH and GLDC effector genes. The names were real and in the records, but nothing there showed that they change the metabolite's level. When we removed the database with the two names, the validator caught the error. Adding this database made the validator weaker. Then, we had LLM agents attack the validator in two rounds to find its gaps. In the first round, 51 test cases found eight defects. When the second round re-tested them, seven were confirmed and one was rejected. In the second round, we also tested three parts not covered by the first round, and confirmed 16 defects there. We fixed these 16 defects, including the most serious one. Three defects from the first round and one from the second share one cause: the validator checks that a value is in the records, but not how the sentence uses it. Therefore, we recommend that validators check each value and how it is used in the sentence. Validators should also be audited again whenever an evidence layer is added.