Hallucination and commission auditor

Finds claims the source never supports.

In audits Layer 2 · Measure
Takes inSource, output
The agentHallucination auditor
Hands backUnsupported-claim list

Checked against clinicians

Same.

Where a person decides

Same.

Works on

Scribes, Q&A, chat, summaries.

Delivered inside a Quality Audit today. Reviewed .

The problem it solves

AI can write something the source never said. Fluent text hides it, and reviewers skim.

How it works

01

Split into claims

The agent breaks the note into individual claims.

02

Verify each claim

Each claim is checked against the transcript or chart.

03

Mark unsupported claims

Claims with no support are flagged, with the nearest source text shown.

An example, illustrative

Before

Note: 'Lungs clear on auscultation.' Transcript: no exam mentioned.

With the agent

Flagged: unsupported. No source for this exam finding.

What it needs from you

Source material and the output to check.

Its limits

Some true claims come from clinician knowledge not in the source, so flags go to a person.

Quick answers

What does the hallucination auditor do?

Verify each output claim against source.

How is the hallucination auditor checked against clinicians?

Same.

What is a hallucination in this context?

Content in the output that the source doesn't support.

Is every flagged claim wrong?

No. Some are valid but unsourced. Clinicians decide, and the rate of false flags is reported.

What can the hallucination auditor not do?

Same. No agent makes a clinical decision.

Use the hallucination auditor on your product