Agentic AI framework: reduces silent hallucination in healthcare

A multi-agent Agentic AI framework submitted on 16 Jun 2026 aims to curb two failure modes in clinical conversational agents: premature diagnostic handoff and silent clinical hallucinations. The architecture replaces "LLM-as-a-judge" routing with deterministic orchestration constraints and adds a neuro-symbolic OLDCARTS gate plus an epistemic uncertainty gate to intercept divergent outputs.

What does the framework do?

The paper describes a multi-agent framework that enforces structured information gathering and checks for epistemic disagreement before a diagnosis is delivered. It enforces OLDCARTS completeness (Onset, Location, Duration, Character, Aggravating/Alleviating factors, Radiation, Timing, and Severity) through a neuro-symbolic state-tracking gate that blocks diagnostic transitions until required dimensions are collected, and computes semantic entropy (H) across K = 5 independent diagnostic samples as an uncertainty-quantification gate.

The authors position these mechanisms as replacements for LLM-as-judge routing: deterministic orchestration constraints drive agent flow, the neuro-symbolic gate guarantees protocol completeness, and the epistemic gate flags divergent outputs prior to patient-facing delivery.

How was it tested and what were the results?

The system was evaluated on 150 test cases using simulated patient agents powered by the llama-3.1-70b-instruct model, and the full architecture achieved 49.3% diagnostic precision, an absolute improvement of 11.3 percentage points over an unconstrained baseline. The study also reports a statistically significant negative correlation, r = -0.181 with p < 0.05, between OLDCARTS completeness (σ) and semantic entropy (H), indicating more complete symptom collection was associated with lower diagnostic uncertainty.

Evaluation specifics from the submission include the simulation setup (simulated patient agents), the base generative model used (llama-3.1-70b-instruct), the test set size (150 cases), the uncertainty sampling parameter (K = 5), and the headline metrics: 49.3% precision and +11.3 percentage points versus baseline.

Why it matters

Clinical conversational agents can hand off prematurely or hallucinate confidently, risks that may reach patients unnoticed. This framework tackles both failure modes with two concrete, mechanistic controls: enforced protocol completeness and an epistemic entropy check across multiple diagnostic samplings. The reported negative correlation between OLDCARTS completeness and semantic entropy provides empirical support that forcing structured questioning reduces internal model disagreement, which directly targets a source of silent hallucination.

Those are pragmatic levers: enforcing a known clinical protocol (OLDCARTS) is an auditable constraint, and measuring semantic entropy across K = 5 samples is a quantifiable safety gate. Together they move beyond ad hoc human-judge routing toward reproducible, rule-governed intervention points inside agentic pipelines.

What to watch

The next validation steps are replication on clinical or prospectively collected datasets and external benchmarking against other safety architectures. Key signals will be whether the 49.3% diagnostic precision and the reported +11.3 percentage point improvement hold outside simulated agents and whether the negative correlation (r = -0.181, p < 0.05) between OLDCARTS completeness and semantic entropy replicates with real patient conversations.

Authors and provenance The paper, "Agentic AI-based Framework for Mitigating Premature Diagnostic Handoff and Silent Hallucination in Healthcare Applications," lists Divyansh Srivastava, Shreya Ghosh, Anshul Verma, and Rajkumar Buyya, and was submitted to arXiv on 16 Jun 2026.