What if the safest genomics AI is not the one that always gives an answer, but the one that knows when evidence cannot support a verdict? ARGUS builds that restraint directly into its investigation loop. Over 90% of disease-associated variants from genome-wide association studies fall in noncoding regulatory regions, yet their functional interpretation remains a central open problem in genomic medicine.

Large language models prompted to interpret such variants routinely hallucinate transcription factor binding changes, fabricate experimental support, and assign biological significance to statistically negligible signals. ARGUS strictly separates deterministic biological computation from LLM-mediated reasoning.

ARGUS uses what it calls deterministic science, agentic reasoning: all biological classification is performed by rule-based, auditable code. The LLM may select investigation actions or write the final report, but it never classifies evidence or assigns verdicts, constrained by pre-classified facts it cannot override.

ARGUS uses a hypothesis-directed investigation loop in which a planner autonomously selects evidence sources, a verifier deterministically interprets each observation, and intermediate results change the investigation path. The defining criterion for agency is that an observation changes what ARGUS does next.

ARGUS consists of two stages: a deterministic prediction stage and an agentic investigation stage that subjects each prediction to evidence-based scrutiny. The deterministic prediction stage is shared with the existing DeepVRegulome pipeline and orchestrated with LangGraph.

Figure one presents ARGUS as an observation-dependent investigation loop rather than a fixed sequence: a rule-based planner selects evidence, a deterministic verifier updates the claim status, and a stopping rule produces structured output. The claim-specific examples show FOXA1 reaching a rescued verdict after three logged events, while KLF6 is abstained after eight logged events, reflecting different intermediate observations.

The separate Atlas comparison and optional language-model narration remain outside the evaluated decision loop. The planner selects the next evidence source based on current uncertainty, querying ADASTRA first for all hypotheses, including null predictions.

If ADASTRA does not resolve the hypothesis, the planner selects JASPAR, and if JASPAR does not resolve it, ENCODE cCRE is queried. If no admissible evidence source remains, the planner returns ABSTAIN. An alternative LLM-mediated planner, Claude Sonnet 4.6, reads the current hypothesis state, remaining budget, and available tool descriptions, then proposes the next action with an explicit rationale.

The verifier deterministically interprets each tool’s raw output and updates the hypothesis state. For ADASTRA, its ten-branch decision tree distinguishes unavailable data, significant allele-specific binding, saturation, direction, and statistical power.

JASPAR evidence is computational and indirect: it can weaken or support a hypothesis but never independently confirm or falsify transcription factor binding. ENCODE cCRE provides regulatory context only. ARGUS was evaluated on variant rs6983267 at the 8q24 cancer risk locus.

Four transcription factors were investigated, producing four divergent trajectories from the same planner operating on real evidence. All observations came from actual local database queries, with no results simulated. For FOXA1, the DNABERT model predicts reference and alternate probabilities of zero point nine nine five, both above the zero point nine five saturation threshold.

The classifier assigns retained binding with a saturation warning, so the planner recognizes saturation and queries ADASTRA. Real ADASTRA data returns significant allele-specific binding for FOXA1, with the reference allele preferred, a false discovery rate of zero point zero three zero, fifteen ChIP-seq experiments, and four hundred seventy-five reads.

The verifier records a possible model false negative and assigns rescued because independent experimental data indicates a real allelic event the model cannot distinguish. The same planner, given the same variant and the same tool set, produces structurally different trajectories because intermediate observations differ.

FOXA1 resolves in one tool call because ADASTRA is definitive. KLF6 requires all three tools because ADASTRA has no data, JASPAR contradicts, and cCRE provides partial support. RAD21 and SP1 follow the same tool path as KLF6 but stop for different reasons: a powered nonsignificant experimental test, and the absence of direct evidence, respectively.

Table one follows ARGUS as the same planner investigates four transcription factors at rs6983267, adapting to different intermediate observations. FOXA1 is rescued in three steps after ADASTRA finds significant allele-specific binding, with an FDR of zero point zero three zero, while KLF6, RAD21, and SP1 each end in abstention after mixed, nonsignificant, unavailable, or non-adjudicated evidence across ADASTRA, JASPAR, and cCRE.

The table matters because it makes the planner’s evidence-dependent trajectories visible rather than reducing every case to a single prediction. The evaluation ran all hypotheses through both the fixed evidence priority of ADASTRA, JASPAR, and cCRE, and an LLM-mediated planner using Claude Sonnet 4.6 with a budget of three tool calls per hypothesis.

The LLM planner receives the current hypothesis state, remaining tools, capability descriptions, and remaining budget, then returns a structured action proposal with an explicit rationale. All biological classification remains deterministic regardless of which planner is used.

Table two compares fixed and LLM planners on five hypotheses using identical frozen evidence snapshots. Both planners rescued FOXA1 with one call to tool A, while the other four hypotheses were abstained on; the fixed planner used A, J, and C, whereas the LLM planner used A and J.

The authors note that LLM results were identical across three independent replicates, illustrating budget behavior without establishing a general planning advantage. The evaluation comprises five transcription-factor hypotheses at two loci.

It demonstrates system behavior and trajectory divergence, but it is not a predictive accuracy benchmark or a controlled measurement of hallucination reduction. The initial DVR outputs are precomputed, and the loop has not been evaluated as an end-to-end natural-language-to-inference system.

ARGUS shows how an LLM can choose what evidence to inspect while deterministic code decides what that evidence means. On a small real-data evaluation, the system produced different evidence paths and preserved abstention when the claim could not be resolved.