Skip to content

The Crown

Our project is an adversarial auditor that verifies clinical notes against two independent sources of truth: the actual visit transcript and the patient's structured FHIR chart.

Code Video Cerebral Valley

Core trick

Checks each note claim against the visit transcript and FHIR chart, and sends disputed claims to independent AI judges whose disagreements are kept, not averaged.

Tags

healthagents

Description

Our project is an adversarial auditor that verifies clinical notes against two independent sources of truth: the actual visit transcript and the patient's structured FHIR chart. It flags any claim that is not supported by either source and sends disputed claims to a panel of independent AI judges that surface disagreement instead of averaging it away. The result is a scored, evidence-cited safety report for every note. This matters because ambient scribing tools can automatically turn spoken conversations into permanent medical records, often without anyone checking the note word for word against what was actually said. A fabricated detail, a missed disclosure, or an overstated finding can then remain in a patient's chart without anyone noticing. When we tested our system on real, unmodified encounters, we found genuine errors that had gone unflagged, including a physical exam finding documented as performed when only imaging was discussed, a patient's age listed one year off from her actual birth date, a lab value labeled as "fasting" even though that word never appeared in the conversation, and a fabricated demographic detail with no supporting source. These examples showed us that an independent, evidence-based second pass can catch the kind of silent documentation errors that can put patient safety at risk before they ever become part of the medical record.
zCranking/Veritas-Hackathon-Project★ 00 forksHTMLMITCreated Jul 18, 2026 · last push Oct 12, 2026

README excerpt

# Veritas Charting

**An adversarial red-team harness that catches when an AI-generated clinical note
silently drifts from what was actually said in a doctor-patient visit.**

[![Live demo](https://img.shields.io/badge/live%20demo-open-2f6fdb)](https://zcranking.github.io/Veritas-Hackathon-Project/)
[![Python 3.10+](https://img.shields.io/badge/python-3.10%2B-3776ab)](https://www.python.org/)
[![Model](https://img.shields.io/badge/model-claude--sonnet--4--6-8a63d2)](https://docs.claude.com/)
[![License: MIT](https://img.shields.io/badge/license-MIT-green)](LICENSE)

<!-- ═══════════════════════════════════════════════════════════════
     SCREENSHOT 1 — HERO  (the single most important image)
     What to capture: the results view on the COVID-19 admission case,
     with a HIGH-severity confirmed finding visible in the right panel.
     Full browser width, light mode, ~1600px wide.
     Save as: assets/hero.png
     ══════════════════════════════════════════════════════════════ -->

![Veritas Charting — flagged discrepancies traced to source](assets/hero.png)

> Ambient scribes turn a visit into a SOAP note. The dangerous failure isn't a
> garbled note — it's a **fluent, plausible** one that quietly asserts a
> medication the patient stopped, or hardens a hedged *"I don't think so"* into
> a clean *"denies."* This harness treats the note-writer as a system under test
> and hunts for exactly those drifts.

**[▶ Try the live demo](https://zcranking.github.io/Veritas-Hackathon-

Team