Veritas

Cryptographic proof-of-personhood with AI-powered quality scoring to eliminate fraud from clinical research.

Live Demo • Features • How It Works • Tech Stack • Getting Started

---

The Problem

The global clinical trials market is worth over $80 billion. A single failed Phase III trial costs around $350 million. An estimated 4% of trial volunteers are "professional patients" who fabricate or exaggerate symptoms for compensation, and a Nature investigation found that up to one quarter of clinical trials in some fields may be problematic or entirely fabricated.

A Harvard research team caught ten fake HIV patients in their trial. People were wearing wigs on video calls to avoid being recognized from previous sign-ups. A separate Alzheimer's study lost 40% of its data to participants who didn't have Alzheimer's.

The best fraud prevention tool most researchers have today is an attention check question: "Select Strongly Agree to prove you're reading this."

The Solution

Veritas combines World ID proof-of-personhood with a multi-dimensional AI quality scoring pipeline to give researchers clean, trustworthy data.

• Each participant cryptographically proves they are a unique human before answering a single question
• Every response is scored across coherence, effort, consistency, and specificity
• Researchers see per-enrollment quality breakdowns in a live analytics dashboard

---

Features

For Researchers

• Study Builder: create surveys with 5 question types (scale, multiple choice, checkbox, short text, long text), conditional logic, and dependency chains
• AI Question Analysis: automatic specificity scoring compares your questions against validated clinical instruments (PHQ-9, GAD-7, BPI) and suggests improvements
• Reverse-Scored Pair Detection: automatically identifies psychometric reverse pairs when a study is published, enabling contradiction detection without manual tagging
• Live Analytics Dashboard: 6 specialized tabs:
  • Overview: quality distribution, enrollment trends, dimension score gauges
  • Integrity: per-enrollment scores with drill-down into individual responses
  • Linguistic: text analysis of response content and patterns
  • Behavior: response timing analysis and suspicious speed flags
  • Questions: per-question metrics, response rates, and quality breakdown
  • Geographic: participant distribution
• Study Lifecycle: Draft → Active → Closed with one-click transitions

For Participants

• World ID Verification: prove you're a unique human with a single scan, no personal information stored
• Clean Survey Experience: one question at a time with progress tracking and automatic time measurement
• Real-Time Validity Feedback: nudges when an answer doesn't address the question, with the option to keep your response
• Compensation Tracking: clear visibility into study compensation and completion status

---

How It Works

Quality Scoring Pipeline

When a participant submits responses, Veritas runs a multi-stage scoring pipeline:

Submission → LLM Quality Scoring → Similarity Analysis → Structured Analysis → Final Verdict

Stage 1: LLM Quality Scoring (per response)

Dimension	Weight	What it measures
Coherence	30%	Is the answer logically meaningful and on-topic?
Effort	25%	Does it show genuine engagement beyond minimal effort?
Consistency	30%	Does it align with the participant's other responses?
Specificity	15%	Does it contain concrete, personal detail?

Stage 2: RAG Similarity Analysis (optional, for text responses)

• Embeds each response using text-embedding-3-small
• Queries nearest neighbors in Pinecone for the same question
• GPT-4o scores semantic similarity to detect copied or templated answers

Stage 3: Structured Analysis (pure math, no LLM)

• Response time scoring: calculates expected reading + answering time per question type, flags suspiciously fast completions
• Reverse-pair contradiction detection: checks if both items in a reverse-scored pair received the same polarity (both high or both low)

Final Score: 50% average LLM scores + 50% structured analysis. Enrollments scoring below 0.5 are automatically flagged.

World ID Integration

Participant → World ID Orb Scan → ZK Proof Generated → Backend Verification → Enrollment Created

• Each study has a unique action string (study_enrollment_{id})
• The nullifier hash is globally unique per person per action
• A person can enroll in multiple studies but never the same study twice
• Enforced cryptographically

---

Tech Stack

Layer	Technology
Framework	Next.js 14 (App Router), React 19, TypeScript
Styling	Tailwind CSS 4.2, shadcn/ui, custom glass morphism
Database	PostgreSQL via Supabase (raw SQL with pg Pool)
Auth	NextAuth.js (Credentials + World ID providers)
Identity	World ID (@worldcoin/idkit) — zero-knowledge proof-of-personhood
AI Scoring	OpenAI GPT-4o (quality scoring), GPT-5.4-mini (validation, reverse pairs)
Embeddings	OpenAI text-embedding-3-small + Pinecone vector DB
Visualization	Recharts, Three.js + React Three Fiber, custom GLSL shaders
Deployment	Vercel

---

Project Structure

src/
├── app/
│   ├── api/
│   │   ├── studies/           # Study CRUD, analytics, reverse pairs, specificity
│   │   ├── enrollments/       # Response submission + scoring pipeline
│   │   ├── validate-response/ # Real-time validity checking
│   │   ├── verify-proof/      # World ID proof verification
│   │   └── participant/       # Participant auth
│   ├── dashboard/             # Researcher dashboard + study detail
│   ├── study/                 # Participant-facing enrollment + survey
│   └── page.tsx               # Landing page
│
├── components/
│   ├── analytics/             # 6 dashboard tabs (overview, integrity, etc.)
│   ├── ui/                    # shadcn/ui primitives
│   └── *.tsx                  # Visual effects (globe, shaders, aurora)
│
├── lib/
│   ├── db.ts                  # PostgreSQL connection pool
│   ├── auth.ts                # NextAuth configuration
│   ├── scorer.ts              # LLM quality scoring pipeline
│   ├── quality.ts             # Structured analysis (timing, reverse pairs, specificity)
│   ├── similarity.ts          # RAG similarity via Pinecone
│   └── embeddings.ts          # OpenAI embedding wrapper
│
└── types/
    └── index.ts               # Shared type definitions

---

Getting Started

Prerequisites

• Node.js 18+
• PostgreSQL database (Supabase recommended)
• World ID developer account
• OpenAI API key

Environment Variables

# Database
DATABASE_URL="postgresql://postgres:[PASSWORD]@db.[REF].supabase.co:5432/postgres"

# Auth
NEXTAUTH_SECRET="your-secret-here"
NEXTAUTH_URL="http://localhost:3000"

# World ID
NEXT_PUBLIC_WORLD_APP_ID="app_..."
WORLD_RP_ID="rp_..."
RP_SIGNING_KEY="0x..."

# OpenAI
OPENAI_API_KEY="sk-..."

# Pinecone (optional — similarity analysis disabled if unset)
PINECONE_API_KEY="..."
PINECONE_INDEX="veritas-responses"

Install & Run

npm install
npm run dev

Open http://localhost:3000.

Demo Accounts

For quick testing without World ID:

Role	Email	Password
Researcher	researcher@demo.veritas	demo123
Participant	participant@demo.veritas	demo123

---

Database

All database operations use raw SQL via the pg driver. The prisma/schema.prisma file exists as a schema reference only. Prisma ORM is not used. Migrations are applied directly via Supabase.

Core Tables

Table	Purpose
Researcher	Researcher accounts (email/password or World ID)
Study	Survey definitions with status lifecycle
Question	Questions with type, config, conditional dependencies
Participant	Unique participants identified by World ID nullifier
Enrollment	Links participant to study with verification status
Response	Individual answers with time tracking
QualityScore	Per-response AI quality scores (coherence, effort, etc.)
ReversePair	Auto-detected psychometric reverse-scored pairs

---

Built with World ID and OpenAI