AI Hallucination Detection Dashboard
Trust,
Measured.
A governance harness that scores AI trust, detects six failure modes, and routes risk to human review
Scroll
The Problem
Integrating AI into legacy systems presents significant challenges regarding data traceability, governance, and the risk of disrupting existing data pipelines. Technical and non-technical teams often struggle to establish repeatable evaluation processes that can observe and measure AI behavior without exposing protected data or impacting production systems. There is a critical need for a framework that clearly defines what is being tested, how it is evaluated, and how AI decision-making can be made transparent.
Dual-Layer
Testing
The project involved developing a self-contained AI hallucination demo designed to test a documentation retrieval tool using synthetic CMS Medicare Part B carrier claims data. This approach allowed for high-volume, regulated data structures to be used without compromising patient privacy.
The system utilizes a dual-layer testing structure: the first layer defines operational domains (Factual, Financial, and Coding), while the second targets specific failure modes such as contradictions and hallucinations. The architecture includes a "Trust Engine" that automates an evaluation sequence, fetching evidence, classifying failures, and running a scoring formula while incorporating Human-in-the-Loop (HITL) oversight for high-stakes outputs. This framework is grounded in established governance guardrails, including the NIST AI RMF, the EU AI Act, and HIPAA.
The Trust Score
TS=max(0,100−(S×Wc))
The framework evaluates AI performance through a deductive Trust Score formula: TS=max(0,100−(S×Wc)).
Weight Category (Wc): Calculated by combining Reversibility Tiers (how easily a fact can be recovered from the source) and Regulatory Exposure Tiers (the legal impact of the failure, such as HIPAA violations).
Severity (S): A composite score (S=V+E+I+T) based on four sub-metrics: Verifiability (ease of checking against source), Evidence Support (grounding in the evidence package), User Impact (effect on decision-making), and Traceability Impact (integrity of the audit trail).
Failure Modes: The system specifically monitors six categories: Evidence Traceability, Retrieval Accuracy, Unsupported Entity, Unsupported Reasoning, Evidence Omission, and Contradiction Detection.
Evidence Traceability
7.0
Retrieval Accuracy
6.5
Unsupported Entity
8.5
Unsupported Reasoning
7.5
Evidence Omission
4.0
Contradiction Detection
7.0
Outcomes
The project successfully demonstrated a complete, dependency-free evaluation loop that provides proportional scoring based on the severity of AI failures. The framework effectively mapped technical failure modes to specific regulatory obligations, such as the EU AI Act's transparency requirements and HIPAA's privacy protections. Additionally, the system established a traceable audit log that records every AI response, evaluation outcome, and reviewer action, ensuring finalized decision ownership and accountability.
Next Steps
- Live Model Integration: Transition from using scripted/prewritten responses to integrating a live AI model into the evaluation architecture.
- Expansion of Failure Modes: Move beyond the initial six fixed failure modes to determine if real-world AI agents produce failures outside these categories.
- Broader Schema Testing: Extend testing to other claim types and data structures beyond the Medicare Part B CMS Carrier SAF schema.
- Enhanced Scoring and Operations: Implement dynamic sub-measure scoring and develop tiered review operations for more complex human-in-the-loop workflows