The difference between "mentioned" and "answered"

The story describes a product to assist nurses with their work. Nurses have ninety minutes and a form to complete, inquiring about falls from the past year. The AI scribe mentioned in the story simply notes that the nurse mentioned having a couple of wobbles, but never verifies if these constitute a fall. This creates a gap the nurse must then fill, often through a phone call or follow-up visit.

The product aims to close this gap by reliably determining if a mentioned item actually satisfies the form's requirement. The developers built an evaluation harness with 47 labelled cases to test the system's accuracy. A false negative (marking an insufficient answer as sufficient) costs a single extra question, while a false positive (accepting an insufficient answer) results in a blank legal field.

The system employs a rule-based approach, defaulting to insufficient and only marking sufficient when confident. The architecture uses Google ADK 2.6.2 on Cloud Run, Gemini 3.6 Flash on Vertex AI, and Firestore for session state. The system transcribes audio, routes it to an adjudicator, and coaches based on the outcome. The design ensures that each item is separately scoreable and prevents one wrong verdict from corrupting others.

The agent tracks coverage against a human-authored form and quotes the relevant span without suggesting answers or identifying interviewees. It focuses on learning and complying with professional practices, rather than generating content or making assumptions about interviewees.