Verifact

An open-source verification engine designed to assist critical thinking by cross-referencing public claims against documented records, accredited fact-checkers, news archives, and institutional evidence.

---

Mission & Principles

Misinformation rarely spreads because people are gullible. It spreads because social media is engineered for emotional velocity, while factual investigation is tedious and time-consuming. In the few seconds between encountering an alarming claim and sharing it, most people do not have time to cross-examine journalistic archives or consult official records.

Verifact is not a "ministry of truth" and does not claim infallible authority. We believe no single algorithm or entity should hand down dogmatic verdicts from on high.

Instead, Verifact functions as an empowering research assistant:

1. Input: A user submits a claim via text, web link, screenshot, or audio clip.
2. Triangulation: The engine retrieves documented evidence across accredited fact-checking registries, mainstream news archives, institutional records, and academic databases.
3. Synthesis: Contextual synthesis identifies points of consensus, documented refutations, or omitted context, presenting verifiable sources so users can judge the evidence for themselves.

---

The Verification Engine & Pipeline

  User Input (Text / URL / Screenshot / Audio Clip)
                       │
       ┌───────────────┴───────────────┐
       ▼                               ▼
  Image OCR                     In-Browser Speech
(Google Cloud Vision)         (Transformers.js / WASM)
       │                               │
       └───────────────┬───────────────┘
                       ▼
             Claim Pre-Processing
       (De-noise, normalize interrogatives)
                       │
       ┌───────────────┼───────────────┬───────────────┐
       ▼               ▼               ▼               ▼
  Fact-Check      News & Web     Institutional     Academic /
  Registries        Search          Records        Scientific
 (Google Fact    (Tavily / News)  (Google CSE)   (Europe PMC,
    Check)                                       OpenAlex, etc.)
       └───────────────┬───────────────┴───────────────┘
                       ▼
              Evidence Aggregator
                       │
                       ▼
             Bayesian Credibility Scoring
                       │
                       ▼
             Contextual Synthesis
           (Gemini 2.5 Flash via LLM)
                       │
                       ▼
        Transparent Report with Primary Citations

1. Ingestion & Pre-Processing

• Text & URLs: Evaluates raw text or extracts the primary narrative from article URLs.
• Screenshots: Passes user-uploaded images through OCR to extract on-image text and metadata without storing images permanently.
• Client-Side Speech-to-Text: Transcribes spoken claims directly inside the user's browser using @huggingface/transformers and an int8-quantized Whisper model running via WebAssembly (onnxruntime-web). The audio never leaves the client device.
• Claim Normalization: Isolates the central testable hypothesis from surrounding commentary and transforms rhetorical queries into testable assertions.

2. Multi-Layer Source Retrieval

Rather than relying on model training memory (which can hallucinate or become outdated), Verifact retrieves live external records in parallel:

• Layer 1: Accredited Fact-Checks: Queries the Google Fact Check Tools API for prior reviews published by certified IFCN signatories.
• Layer 2: News Coverage & Reporting: Searches reputable press indices and journalistic databases via Tavily Search API and NewsAPI.
• Layer 3: Institutional & Official Records: Searches official government gazettes, legislative registries, and public institutional archives via Google Custom Search Engine (cx).
• Layer 4: Academic & Biomedical Repositories: When claims involve medical, clinical, or scientific assertions, the engine cross-references open biomedical literature (Europe PMC, OpenAlex, ClinicalTrials.gov, and openFDA).

3. Bayesian Credibility Scoring & Evidence Status

The scoring engine combines evidence weights probabilistically across all responsive layers:

• Evidence Status: Maps findings to clear statuses:
  • Confirmat de documente / surse multiple (Corroborated by primary sources)
  • Contrazis de sursele oficiale / presă (Contradicted by documented facts)
  • Lipsit de context verificabil (Missing critical context)
  • Fără surse credibile identificate (No credible evidence found)
• Signal Coloration: False or unsubstantiated claims carry prominent signal red warnings (Informație falsă / Informație posibil falsă / neverificată).

4. Contextual Synthesis & Hallucination Prevention

• Gathered articles and official quotes are synthesized into a coherent dossier.
• A language model evaluates consensus, highlights missing context, and writes an executive explanation.
• Strict Citation Integrity: Every assertion in the synthesis must link directly to an existing source in the retrieved evidence set. The engine rejects ungrounded or hallucinated citations.

---

Tech Stack

• Frontend & Routing: Next.js 16 (App Router), React 18, Strict TypeScript (strict: true).
• Styling: Native CSS Modules (mobile-first, zero utility-framework overhead, shared design tokens).
• Database & Auth: Supabase (PostgreSQL with Row Level Security, email/password, and Google OAuth).
• Language Models & AI: OpenRouter (defaulting to Google Gemini 2.5 Flash, with fallbacks) or direct Google AI Studio.
• Client-Side ML: @huggingface/transformers with ONNX Runtime Web for local, in-browser speech transcription.
• Testing: Jest (ts-jest), Playwright (E2E), and ESLint 9.

---

Codebase Organization

src/
├── app/                  # Next.js App Router (pages and API route handlers)
│   ├── api/              # Serverless API routes (/api/verify, /api/ocr, etc.)
│   ├── (auth)/           # Authentication flows (login, callback, reset)
│   ├── rapoarte/         # Public and authenticated report views
│   └── ...               # Informational and educational pages
├── components/           # React components with adjacent CSS Modules
│   ├── ui/               # Base primitives (Button, Modal, Input, Callout)
│   ├── verify/           # Verification search input, status streams, report dossier
│   ├── dashboard/        # User history and saved checks
│   └── layout/           # Header, Footer, navigation
├── lib/                  # Application business logic
│   ├── verification/     # 4-layer search orchestrator, scoring algorithm, source filters
│   ├── ai/               # Gemini prompt pipelines, claim decomposition, evidence synthesis
│   ├── transcription/    # In-browser Whisper speech-to-text via Transformers.js
│   ├── supabase/         # Server and client Supabase clients
│   └── utils/            # Structured logging, circuit breakers, retries, and rate limits
└── types/                # Strict TypeScript interfaces and domain schemas

---

Guiding Principles & Privacy

• Transparent Verification: All evidence sources, search queries, and confidence breakdowns are visible to the user. No opaque black boxes.
• Strict Data Privacy: Screenshots uploaded for OCR are processed in memory and discarded; they are never kept on disk or sold. Audio transcription runs entirely inside the user's browser without uploading audio files to remote servers.
• Political Neutrality: Verifact does not take editorial stances or political positions. It surfaces what documented records show, cross-referenced from independent and accredited sources.
• Open Source: The code is licensed under the MIT License.

---

License

This project is open-source under the MIT License.