Most disease-linked changes in DNA sit outside genes, where their effects are hard to read. An AI can sound confident about them—and still invent the evidence, so this system is designed to stop and say when the evidence is not enough.

More than 90 percent of disease-associated variants fall in regulatory regions outside genes, yet interpreting what they do remains a central problem in genomic medicine. Large language models can hallucinate changes in DNA-binding proteins, fabricate experimental support, and assign importance to signals that are statistically negligible.

ARGUS separates biological computation from language-model reasoning, then lets a planner choose evidence sources according to the uncertainty that remains. The central rule is simple: auditable, rule-based code classifies the biological evidence, while the language model may choose the next investigation or write the report, but cannot assign the verdict.

This works like a careful detective: each observation changes what ARGUS does next, rather than sending it through a fixed checklist. ARGUS has two stages: one makes deterministic predictions, and the other subjects each prediction to evidence-based scrutiny. So the system does not treat its first prediction as the answer; it opens an investigation around that prediction.

The system was evaluated on one variant at a cancer-risk region, with four DNA-binding proteins investigated from the same starting point. Every observation came from real local database queries; none of the results were simulated. For one protein, both DNA versions received nearly maximal predicted binding, triggering a saturation warning: the prediction could not distinguish the two versions.

The planner therefore checked experimental evidence, which found significant allele-specific binding and contradicted the model’s prediction that there was no difference. ARGUS called this a rescued case—a possible model false negative—and stopped once the evidence settled the hypothesis.

The same planner produced structurally different investigation paths because the intermediate observations differed, even though the starting variant and available tools were the same. One case resolved with a single tool call, while another required all three tools because the evidence stayed incomplete: one source had no data, another contradicted the claim, and a third offered only partial support.

The important point is that each next decision depended on the previous verdict, not on a predetermined sequence. The evaluation shows how the system behaves and how its investigation paths diverge, but it is not a predictive-accuracy benchmark or a controlled measurement of reduced hallucination.

The starting predictions were precomputed, and the investigation loop was not evaluated as a complete system that turns natural language directly into biological conclusions. ARGUS shows that using a language model for biological interpretation does not inevitably require accepting hallucinated conclusions.

Because each verdict is separated from the language model’s reasoning and tied to a specific observation, the conclusions are actively tested rather than merely plausible. For someone facing an uncertain genetic clue, that means a clearer record of what the evidence supports—and where it still does not.

ARGUS makes an AI investigate one claim at a time, changing course when each piece of evidence demands it. That can make genetic clues more traceable and help prevent uncertain interpretations from being mistaken for medical facts.