Technical Summary: DrugClaw and DrugAudit

Problem Statement

Drug-information question answering (QA) operates in a high-stakes environment where hallucinated facts can directly mislead clinical decision-making. While Large Language Models (LLMs) demonstrate strong general biomedical knowledge, their parametric knowledge degrades rapidly outside high-resource disease areas, and they lack native audit trails to regulatory or peer-reviewed sources. Furthermore, existing tool-augmented agents often fail by skipping primary regulatory sources in favor of the first available URL or by fabricating plausible-looking numerics without citations.

A critical gap exists in the evaluation layer: current biomedical QA scoring typically grades surface answer correctness but fails to distinguish between citations to upstream primary sources (e.g., FDA Labels, FAERS) and downstream aggregator wrappers of those same sources. In regulated drug-information settings, provenance is not merely a presentation choice but a fundamental evaluation axis.

Methodology

The DrugClaw System

DrugClaw is a multi-agent retrieval-augmented system designed to query a registry of drug and pharmacovigilance skills via a reflection-driven state-machine workflow. Its architecture is built upon three design goals:

1. Primary-source traceability: Every emitted claim must carry an identifier resolvable against a primary regulatory or peer-reviewed record.
2. Calibrated abstention: The system must refuse to answer when underlying skills return no supporting evidence rather than fabricating facts.
3. Adaptive retrieval depth: The retrieve-and-reason loop terminates once evidence is sufficient but deepens automatically when evidence is thin.

Architecture and Workflow:
The system operates as a discrete dynamical system over a state $S_t = (P_t, E_t, G_t, a_t, r_t)$, where $P_t$ is the query plan, $E_t$ is the evidence pool, $G_t$ is a claim-evidence subgraph, $a_t$ is the intermediate prose answer, and $r_t$ is a reflection sufficiency score. The workflow iterates through six specialized agents (Plan, Retrieval, Graph Builder, Re-Ranker, Response, Reflector) until $r_t \geq \tau_{suff}$ or a maximum iteration count is reached.

• Skill Registry: DrugClaw accesses a closed registry of 57 active skills drawn from over 70 curated resources (e.g., openFDA, ChEMBL, DrugBank, PharmGKB, SIDER). These are organized into a 15-subcategory taxonomy.
• Sandboxed Code Agent: To ensure safety and determinism, retrieval is mediated through a Code Agent. This agent translates subqueries into a restricted Python sublanguage ($L_{safe}$) that bans unbounded iteration and dynamic execution. Code is executed in a proxy-only sandbox with a hard timeout. If validation or runtime fails, the system falls back to a direct API call.
• Reflection Mechanism: The Reflector agent computes a sufficiency score $\rho_{t+1}$ based on coverage, completeness, source quality, consistency, and actionability. A regression penalty is applied if the new score is lower than the previous one, discouraging "thrashing" across iterations.
• Reasoning Modes: The system supports three modes:
  • Linear: A single-pass workflow prioritizing low latency.
  • Graph: A full iterative workflow with claim-evidence graph construction and reranking.
  • Web-only: A fallback mode bypassing the registry for direct web search (used only when the registry is insufficient).

The DrugAudit Benchmark

To evaluate these systems, the authors introduce DrugAudit, a 3,772-item authority-aware benchmark derived from nine source databases. Unlike standard benchmarks, DrugAudit includes gold-standard citations with field-level locators and URLs.

Authority-Aware Metric Panel:
The evaluation framework introduces four metrics to address structural biases in prior scoring:

1. Source Authority Match: Uses an equivalence-bucket rule and an asymmetric "upstream-of" relation. It awards credit if a system cites a primary upstream source (e.g., FAERS) even if the gold standard cites a downstream aggregator (e.g., SIDER), provided the upstream source is authoritative for that query.
2. Snippet Semantic Overlap: Replaces strict substring matching with token-Jaccard similarity over non-stop-word tokens.
3. Primary-source Rate: Calculates the mean weight of cited sources based on a hierarchy (1.0 for regulatory/peer-reviewed, 0.7 for curated KBs, 0.5 for aggregators).
4. Faithfulness Rate: Measures the fraction of predicted citations that are grounded in the gold source universe or answer text.

Evaluation Protocol:
The paper employs a dual-judge LLM-as-judge protocol using Llama-3.1-70B-Instruct and gpt-oss-120b. These judges are external to all candidate systems to eliminate self-preference bias. They score answer correctness based on factual equivalence rules (unit conversion, rounding tolerance, order-insensitivity) and calibrated refusal logic. Inter-judge agreement ($\kappa$) is reported as 0.88 (almost-perfect).

Key Contributions

1. DrugClaw: An authority-grounded multi-agent system that queries a closed registry of vetted regulatory and peer-reviewed sources, ensuring every claim is traceable to a primary record.
2. DrugAudit: A 3,772-item benchmark with an evaluation panel that explicitly scores upstream-of-gold source match, token-level semantic overlap, and citation faithfulness.
3. Dual-Judge Protocol: A scalable evaluation method using two independent, open-weight LLMs to achieve almost-perfect agreement ($\kappa = 0.88$) on answer correctness.
4. Empirical Results: Comprehensive comparison of seven systems, demonstrating that DrugClaw outperforms baselines on primary-source rates and evidence density.

Results

Evaluated on DrugAudit, MedQA (drug-related subset, 751 items), and PubMedQA (drug-related subset, 512 items), DrugClaw achieved top-1 performance across all headline metrics:

• Primary-Source Rate: DrugClaw-graph achieved 0.918, a +10.1 percentage point (pp) improvement over the next-best system (0.817).
• Faithfulness: DrugClaw-linear achieved 0.887, a +5.9 pp improvement over the next-best.
• Composite Evidence Index (EI): DrugClaw-graph scored 0.632 (Llama judge) and 0.623 (gpt-oss judge), leading all baselines.
• Closed-Form Accuracy: DrugClaw-graph achieved 0.920 on MedQA and 0.693 on PubMedQA.
• Calibrated Refusal: DrugClaw refused to answer on 41–48% of items. Among these refusals, 91–97% aligned with gold answers indicating "no data," whereas baselines showed only 1–5% alignment. This indicates the system effectively abstains when evidence is insufficient.

The results show that while closed-form accuracy leads were competitive (within bootstrap confidence intervals of baselines), the authority-aware metrics (Primary-source rate and Evidence Index) showed statistically significant separation from baselines.

Significance

The paper claims that DrugClaw and DrugAudit address the critical need for provenance in drug-information QA. By distinguishing between primary regulatory records and downstream aggregators, the system operationalizes the standards of evidence-based medicine reporting. The results suggest that a closed-registry approach with adaptive retrieval and calibrated refusal yields higher structural quality in citations and lower hallucination rates compared to open-web or purely parametric approaches. The dual-judge protocol and authority-aware metrics provide a robust framework for evaluating future retrieval-augmented systems in regulated domains.

The authors note limitations, including the difficulty of parsing multi-claim prose from graph mode by LLM judges and the coverage trade-off inherent in a closed registry (which may miss long-tail drugs not yet in the registry). However, they assert that the approach does not require retraining models and offers a concrete path toward more auditable clinical AI.