VerityNgn Methodology Documentation

Version 1.0 | Last Updated: October 23, 2025

Executive Summary

VerityNgn is a multimodal AI-powered video verification system that analyzes YouTube videos to assess the truthfulness of claims made within them. The system combines cutting-edge multimodal LLM analysis, counter-intelligence techniques, and probabilistic reasoning to generate comprehensive truthfulness reports. Core Innovation: VerityNgn is the first system to combine:

Frame-by-frame multimodal video analysis (1 FPS sampling)
YouTube counter-intelligence (analyzing review videos for contradictory evidence)
Press release bias detection and penalty system
Probabilistic truthfulness assessment with evidence weighting

System Architecture
Multimodal Analysis Pipeline
Claims Extraction Methodology
Counter-Intelligence System
Probability Distribution Model
Evidence Classification & Weighting
Verification Algorithm
Report Generation
Limitations & Future Work

1. System Architecture

Overview

VerityNgn follows a pipeline architecture with six main stages:

┌─────────────────────────────────────────────────────┐
│  Stage 1: Video Download & Preprocessing            │
│  - YouTube URL validation                            │
│  - Video metadata extraction                         │
│  - Transcript extraction                             │
└────────────┬────────────────────────────────────────┘
             │
             ▼
┌─────────────────────────────────────────────────────┐
│  Stage 2: Multimodal Analysis                        │
│  - Frame-by-frame video analysis (1 FPS)            │
│  - Audio transcription with timestamps              │
│  - OCR for visual text extraction                   │
│  - Claims extraction (5-15 claims per video)        │
└────────────┬────────────────────────────────────────┘
             │
             ▼
┌─────────────────────────────────────────────────────┐
│  Stage 3: Counter-Intelligence Gathering            │
│  - YouTube review search                             │
│  - Contradictory evidence detection                  │
│  - Press release identification                      │
└────────────┬────────────────────────────────────────┘
             │
             ▼
┌─────────────────────────────────────────────────────┐
│  Stage 4: Claims Verification                        │
│  - Web search for supporting evidence               │
│  - Evidence classification & weighting              │
│  - Source credibility assessment                    │
└────────────┬────────────────────────────────────────┘
             │
             ▼
┌─────────────────────────────────────────────────────┐
│  Stage 5: Probability Calculation                    │
│  - Multi-factor probability distribution            │
│  - Counter-intelligence adjustments                  │
│  - Evidence quality boosts                           │
│  - Normalization & mapping to verdicts              │
└────────────┬────────────────────────────────────────┘
             │
             ▼
┌─────────────────────────────────────────────────────┐
│  Stage 6: Report Generation                          │
│  - HTML, Markdown, and JSON formats                 │
│  - Evidence compilation                              │
│  - Detailed explanations                             │
│  - Source attribution                                │
└─────────────────────────────────────────────────────┘

Technology Stack

Multimodal LLM: Google Gemini 2.5 Flash (via Vertex AI)
Context Window: 64K tokens for comprehensive analysis
Video Processing: yt-dlp for video download and metadata
Web Search: Google Custom Search API
Framework: LangGraph for workflow orchestration
Storage: Local filesystem or Google Cloud Storage

2. Multimodal Analysis Pipeline

Video Processing Strategy

VerityNgn uses an aggressive frame-sampling approach:

SAMPLING_RATE = 1 FPS  # 1 frame per second
TARGET_CLAIMS = 3 per minute of video
CONTEXT_WINDOW = 64K tokens (Gemini 2.5 Flash)

Frame-by-Frame Analysis

For each frame, the system analyzes:

Visual Content:
- On-screen text and graphics
- Demonstrations and visual evidence
- Charts, graphs, and data visualizations
- Product displays or examples
Audio Track:
- Spoken statements with timestamps
- Speaker identification
- Tone and delivery analysis
Metadata:
- Video title and description
- Channel information
- Upload date and view count
- Comments and engagement metrics

CRAAP Criteria Integration

All claims are evaluated using the CRAAP framework:

Currency: When was the information published?
Relevance: Is it relevant to the video’s main message?
Authority: What credentials or expertise is claimed?
Accuracy: Can it be verified with external sources?
Purpose: Is it promotional, educational, or persuasive?

Prompt Engineering

The system uses carefully crafted prompts to ensure high-quality claim extraction:

CRITICAL INSTRUCTIONS FOR MULTIMODAL VIDEO ANALYSIS:
- Analyze this video with MAXIMUM detail at 1 FRAME PER SECOND sampling rate
- Extract ALL factual claims, statements, assertions from ACTUAL VIDEO CONTENT
- Focus on SPOKEN WORDS, VISUAL TEXT, ON-SCREEN GRAPHICS, and DEMONSTRATIONS
- Ignore background music, decorative elements, or irrelevant visuals
- Extract EXACTLY 5-15 specific, verifiable claims

Claim Mix Requirements:

70% Scientific & Verifiable Claims (studies, statistics, product effectiveness)
10% Speaker Credibility Claims (credentials, affiliations, experience)
20% Other Verifiable Claims (dates, locations, specific statements)

Output Format

Claims are extracted in structured JSON:

&#123;
  "claim_text": "Dr. X has 20 years of experience in endocrinology",
  "timestamp": "02:15",
  "speaker": "Dr. X (Narrator)",
  "source_type": "spoken",
  "initial_assessment": "Verifiable credential claim requiring external verification"
&#125;

3. Claims Extraction Methodology

Claim Quality Criteria

Each extracted claim must meet these requirements:

Specificity: Concrete, not vague or general
Verifiability: Can be checked against external sources
Relevance: Central to the video’s message
Factuality: Stated as fact, not opinion
Significance: Meaningful for truthfulness assessment

Claim Types

Primary Claims (70%):

Study results: “Study X found Y outcome”
Statistics: “75% of patients experienced Z”
Product effectiveness: “Product causes W in N days”
Scientific findings: “Research shows…”

Secondary Claims (10%):

Educational credentials: “Dr. X studied at Y”
Professional experience: “X years in field”
Institutional affiliations: “Works at Hospital Y”
Awards and recognitions

Other Claims (20%):

Historical facts
Geographic information
Temporal data
Specific events or occurrences

Anti-Patterns (Avoided)

The system explicitly avoids:

Vague motivational statements
General health advice without specifics
Obvious facts not requiring verification
Micro-claims too granular for meaningful assessment
Subjective opinions without factual basis

4. Counter-Intelligence System

Overview

VerityNgn’s counter-intelligence system is designed to detect and weight contradictory evidence that challenges claims made in the analyzed video. This is one of the system’s most innovative features.

Two-Pronged Approach

A. Press Release Detection

Purpose: Identify self-promotional content that lacks independent validation. Detection Method:

PRESS_RELEASE_INDICATORS = [
    'prnewswire.com', 'businesswire.com', 'globenewswire.com',
    'prnewswire', 'marketwired', 'accesswire', 'prweb.com',
    'press release', 'newswire', 'press statement'
]

Bias Application:

Probability Adjustment: -0.4 (significant negative bias)
Rationale: Press releases are promotional, not independent verification
Distribution Impact:
- Reduce TRUE by 60% of adjustment
- Increase FALSE by 40% of adjustment
- Increase UNCERTAIN by 40% of adjustment

Example:

Claim: "Product X causes 15 pounds of weight loss"
Evidence: 5 press releases from product manufacturer

Result: TRUE probability reduced by 24% (0.4 * 0.6)
        FALSE probability increased by 16% (0.4 * 0.4)
        UNCERTAIN increased by 16% (0.4 * 0.4)

B. YouTube Counter-Intelligence

Purpose: Find and analyze review videos that contradict claims in the original video. Search Strategy:

COUNTER_INTEL_QUERIES = [
    "&#123;video_title&#125; scam review",
    "&#123;video_title&#125; fake exposed",
    "&#123;video_title&#125; debunk analysis",
    "&#123;main_keywords&#125; + warning review",
    "&#123;product_name&#125; doesn't work"
]

Transcript Analysis: The system analyzes transcripts of review videos for:

Counter-Signals:
- ‘scam’, ‘fake’, ‘fraud’, ‘lie’, ‘misleading’
- ‘doesn’t work’, ‘waste of money’, ‘no results’
- ‘red flags’, ‘warning’, ‘beware’, ‘avoid’
- ‘overpriced’, ‘overhyped’, ‘disappointing’
- ‘fabricated’, ‘deceptive’, ‘predatory’, ‘exposed’
Supporting Signals:
- ‘works’, ‘effective’, ‘results’, ‘recommend’
- ‘good’, ‘helps’, ‘success’, ‘positive’
- ‘beneficial’, ‘worth it’, ‘legit’

Stance Determination:

counter_ratio = counter_signals / (counter_signals + supporting_signals)

if counter_ratio > 0.7:
    stance = 'counter'
    confidence = min(0.95, 0.6 + (counter_ratio - 0.7) * 1.17)
elif counter_ratio &lt; 0.3:
    stance = 'supporting'
    confidence = min(0.95, 0.6 + (0.3 - counter_ratio) * 1.17)
else:
    stance = 'neutral'
    confidence = 0.5

Credibility Weighting: Counter-intelligence evidence is weighted by:

View Count: Videos with 100K+ views = stronger signal
Channel Credibility: Established channels weighted higher
Video Age: Recent videos weighted higher
Engagement: Like ratio and comment sentiment

Probability Impact: YouTube counter-intelligence has reduced impact compared to earlier versions:

# Enhanced (balanced) approach:
youtube_impact = min(0.20, youtube_counter_power * 0.08)
enhanced_dist["FALSE"] = min(0.85, base_false + youtube_impact)

# Reduction rationale: Avoid over-aggressive FALSE bias
# Previous: max 0.35 impact, max 0.9 FALSE
# Current: max 0.20 impact, max 0.85 FALSE

Counter-Intelligence in Action

Example Case:

Original Video: "Amazing weight loss supplement - lose 15 pounds!"

Counter-Intelligence Found:
1. "Product X Scam Exposed" (450K views) - Counter stance (90% confidence)
2. "I Tried Product X - My Honest Review" (120K views) - Counter stance (85% confidence)
3. "Product X Doesn't Work - Warning" (75K views) - Counter stance (80% confidence)

Impact:
- Base FALSE probability: 30%
- YouTube CI adjustment: +12% (0.12 based on reach and confidence)
- Final FALSE probability: 42%

5. Probability Distribution Model

Three-State Model

VerityNgn uses a probabilistic approach rather than binary true/false:

P(TRUE) + P(FALSE) + P(UNCERTAIN) = 1.0

Base Probability Initialization

base_dist = &#123;
    "TRUE": 0.3,      # Starting assumption: neutral
    "FALSE": 0.3,      # Equal weight to false
    "UNCERTAIN": 0.4   # Highest initial uncertainty
&#125;

Enhancement Factors

The base distribution is adjusted based on five key factors:

Factor 1: Evidence Coverage (10-30% boost)

coverage_score = evidence_count / (claim_word_count / 5)
if coverage_score > 1.0:
    boost = min(0.3, coverage_score * 0.15)
    enhanced_dist["TRUE"] += boost

Rationale: More evidence per claim indicates stronger validation.

Factor 2: Independent Source Ratio (10-25% boost)

independent_ratio = independent_sources / total_sources
if independent_ratio > 0.5:
    boost = min(0.25, (independent_ratio - 0.5) * 0.5)
    enhanced_dist["TRUE"] += boost

Rationale: Independent sources provide unbiased validation.

Factor 3: Scientific Evidence (up to 40% boost)

science_weight = min(0.4, scientific_power * 0.2)
enhanced_dist["TRUE"] += science_weight

Rationale: Scientific sources (journals, academic institutions) carry highest weight.

Factor 4: YouTube Counter-Intelligence (-20% impact)

youtube_impact = min(0.20, youtube_counter_power * 0.08)
enhanced_dist["FALSE"] += youtube_impact

Rationale: Contradictory reviews reduce claim trustworthiness.

Factor 5: Press Release Penalty (-40% impact)

self_ref_count = count_self_referential_sources()
penalty = min(0.4, self_ref_count * 0.15)
enhanced_dist["FALSE"] += penalty

Rationale: Self-promotional content is inherently biased.

Normalization

After all adjustments, probabilities are normalized:

total = sum(enhanced_dist.values())
if total > 0:
    enhanced_dist = &#123;k: v / total for k, v in enhanced_dist.items()&#125;

# Ensure minimum thresholds
for key in ["TRUE", "FALSE", "UNCERTAIN"]:
    enhanced_dist[key] = max(0.001, enhanced_dist[key])

Verdict Mapping

The continuous probabilities are mapped to human-readable verdicts:

TRUE% = P(TRUE) * 100
FALSE% = P(FALSE) * 100
UNCERTAIN% = P(UNCERTAIN) * 100

if TRUE \> 70 and FALSE \&lt; 10:
    verdict = "HIGHLY_LIKELY_TRUE"
elif (TRUE + UNCERTAIN) \> 65 and FALSE \&lt; 35:
    verdict = "LIKELY_TRUE"
elif TRUE \> 40 and FALSE \&lt; 35:
    verdict = "LEANING_TRUE"
elif abs(TRUE - FALSE) \&lt; 10:
    verdict = "UNCERTAIN"
elif FALSE \> 35 and TRUE \&lt; 30:
    verdict = "LEANING_FALSE"
elif (FALSE + UNCERTAIN) \> 65 and TRUE \&lt; 35:
    verdict = "LIKELY_FALSE"
elif FALSE \> 75:
    verdict = "HIGHLY_LIKELY_FALSE"

Rationale for Thresholds:

65% threshold for “likely” verdicts (relaxed from 70% to avoid over-conservatism)
70% threshold for “highly likely” verdicts
Combination with UNCERTAIN allows for nuanced assessment

6. Evidence Classification & Weighting

Source Categories

Evidence is classified into six categories, each with distinct validation power:

1. Scientific Sources (Highest Weight: 2.5-4.0)

Characteristics:

Peer-reviewed journals
Academic institutions (.edu domains)
Research papers and studies
Government research agencies

Examples:

PubMed/NCBI publications
Journal articles (Nature, Science, NEJM)
University research papers
NIH/CDC reports

Validation Power: 2.5 - 4.0 (highest)

2. Independent Sources (High Weight: 1.5-2.0)

Characteristics:

Reputable news organizations
Fact-checking organizations
Independent analysis
No financial stake in outcome

Examples:

Reuters, AP News, BBC
FactCheck.org, Snopes, PolitiFact
Independent medical websites (Mayo Clinic, WebMD)

Validation Power: 1.5 - 2.0

3. Government Sources (High Weight: 1.5-2.5)

Characteristics:

.gov domains
Official government publications
Regulatory agency reports

Examples:

FDA reports
CDC guidelines
NIH publications
FTC consumer alerts

Validation Power: 1.5 - 2.5

4. Press Releases (Negative Weight: -0.5 to -1.0)

Characteristics:

Self-promotional content
Newswire services
Company press statements
Marketing materials

Examples:

PR Newswire articles
Business Wire releases
Company blog posts
Marketing pages

Validation Power: -0.5 to -1.0 (penalized)

5. YouTube Counter-Intelligence (Variable: 0.5-2.0)

Characteristics:

Review videos
Debunking content
Investigation videos
Consumer experiences

Weight Determined By:

View count (100K+ = higher weight)
Channel credibility
Stance confidence
Counter-signal strength

Validation Power: 0.5 - 2.0 (variable)

6. Other Web Sources (Standard Weight: 1.0)

Characteristics:

General web pages
Blogs
Forums
Social media

Validation Power: 1.0 (baseline)

Self-Referential Detection

The system identifies “self-referential” sources that cite the original video or its creators:

def is_self_referential(source_url, source_text, video_title, video_channel):
    # Check if source cites the original video
    if video_title.lower() in source_text.lower():
        return True
    if video_channel.lower() in source_text.lower():
        return True
    if source_url contains video_domain:
        return True
    return False

Self-referential sources are:

Separated into press_releases category
Given negative validation power
Excluded from supporting evidence counts

7. Verification Algorithm

Claim Verification Process

For each extracted claim, the system follows this process:

1. Generate search queries
   ├─ Fact-check queries: "&#123;claim&#125; fact check"
   ├─ Scientific queries: "&#123;claim&#125; study research"
   ├─ General queries: exact claim text
   └─ Debunk queries: "&#123;claim&#125; debunk false"

2. Execute web searches
   ├─ Google Custom Search API
   ├─ Retrieve top 10 results per query
   └─ Extract URL, title, snippet

3. Classify evidence
   ├─ Identify source category
   ├─ Assign validation power
   ├─ Detect self-referential content
   └─ Determine supporting vs contradicting

4. Calculate probability distribution
   ├─ Start with base distribution
   ├─ Apply evidence boosts
   ├─ Apply counter-intelligence penalties
   ├─ Normalize to sum = 1.0
   └─ Map to verdict category

5. Generate explanation
   ├─ Summarize evidence
   ├─ Explain probability factors
   ├─ Cite key sources
   └─ Provide confidence assessment

Multi-Factor Probability Calculation

The complete algorithm:

def calculate_enhanced_probability_distribution(
    base_dist: Dict[str, float],
    evidence_groups: Dict[str, List[Dict]]
) -> Tuple[Dict[str, float], List[str]]:
    
    enhanced_dist = base_dist.copy()
    modifications = []
    
    # Extract evidence by category
    scientific_evidence = evidence_groups["scientific"]
    independent_evidence = evidence_groups["independent"]
    press_releases = evidence_groups["press_releases"]
    youtube_counter = evidence_groups["youtube_reviews"]
    
    # Calculate validation powers
    scientific_power = sum(e['validation_power'] for e in scientific_evidence)
    independent_power = sum(e['validation_power'] for e in independent_evidence)
    youtube_counter_power = sum(e['validation_power'] for e in youtube_counter)
    
    # Factor 1: Evidence coverage
    total_evidence = len(scientific_evidence) + len(independent_evidence)
    claim_length = len(claim_text.split())
    coverage_score = total_evidence / (claim_length / 5)
    
    if coverage_score > 1.0:
        coverage_boost = min(0.3, coverage_score * 0.15)
        enhanced_dist["TRUE"] += coverage_boost
        modifications.append(f"Evidence coverage boost: +&#123;coverage_boost:.2f&#125;")
    
    # Factor 2: Independent source ratio
    if independent_power > 0:
        independent_ratio = independent_power / (independent_power + scientific_power + 0.01)
        if independent_ratio > 0.5:
            indie_boost = min(0.25, (independent_ratio - 0.5) * 0.5)
            enhanced_dist["TRUE"] += indie_boost
            modifications.append(f"Independent sources boost: +&#123;indie_boost:.2f&#125;")
    
    # Factor 3: Scientific evidence weighting
    if scientific_power > 0:
        science_weight = min(0.4, scientific_power * 0.2)
        enhanced_dist["TRUE"] += science_weight
        modifications.append(f"Scientific evidence boost: +&#123;science_weight:.2f&#125;")
    
    # Factor 4: YouTube counter-intelligence
    if youtube_counter_power > 0:
        youtube_impact = min(0.20, youtube_counter_power * 0.08)
        enhanced_dist["FALSE"] += youtube_impact
        modifications.append(f"YouTube CI penalty: +&#123;youtube_impact:.2f&#125; to FALSE")
    
    # Factor 5: Press release penalty
    self_ref_count = len([pr for pr in press_releases if pr.get('self_referential')])
    if self_ref_count > 0:
        penalty = min(0.4, self_ref_count * 0.15)
        enhanced_dist["FALSE"] += penalty
        modifications.append(f"Press release penalty: +&#123;penalty:.2f&#125; to FALSE")
    
    # Normalize
    total = sum(enhanced_dist.values())
    enhanced_dist = &#123;k: v / total for k, v in enhanced_dist.items()&#125;
    
    # Ensure minimums
    for key in ["TRUE", "FALSE", "UNCERTAIN"]:
        enhanced_dist[key] = max(0.001, enhanced_dist[key])
    
    return enhanced_dist, modifications

8. Report Generation

Report Structure

Each report contains:

Executive Summary
- Total claims analyzed
- Verdict distribution (Highly Likely True, Likely False, etc.)
- Overall truthfulness assessment
Video Information
- Title, channel, duration
- Upload date and view count
- Video embed/thumbnail
Detailed Claims Analysis
- Claim text with timestamp
- Speaker identification
- Verification verdict
- Probability distribution
- Detailed explanation
- Source links
Evidence Summary
- Categorized by source type
- Validation power indicators
- Quality metrics
Counter-Intelligence Section
- Press releases identified
- YouTube review analysis
- Impact on truthfulness scores

Output Formats

HTML Report:

Interactive source links (modal popups)
Video embed
Formatted tables
Visual probability bars

Markdown Report:

Clean, readable format
Source citations with links
Table-formatted claims analysis
GitHub-compatible

JSON Report:

Structured data format
All probabilities and evidence
Machine-readable
API-friendly

Explanation Generation

For each claim, the system generates a human-readable explanation: Template:

[VERDICT] based on [EVIDENCE_COUNT] sources ([SCIENTIFIC_COUNT] scientific, 
[INDEPENDENT_COUNT] independent). 

[COUNTER_INTEL_STATEMENT if applicable]

Key supporting evidence: [TOP_SOURCES]

Probability: TRUE: X%, FALSE: Y%, UNCERTAIN: Z%

Example:

LIKELY FALSE based on 8 sources (1 scientific, 3 independent). 

YOUTUBE COUNTER-INTELLIGENCE: Multiple review videos (450K+ views) contradict 
this claim, providing independent contradictory perspective.

Press release sources have reduced validation power due to self-promotional nature.

Key supporting evidence: [1] National Institutes of Health study shows no 
significant effect [2] Mayo Clinic states insufficient evidence.

Probability: TRUE: 15%, FALSE: 70%, UNCERTAIN: 15%

9. Limitations & Future Work

Current Limitations

1. Language Support

Current: English-only
Impact: Cannot analyze non-English videos
Mitigation: Rely on auto-translated transcripts (degraded quality)

2. Visual Understanding

Current: Text-based analysis of visual content
Impact: May miss subtle visual cues or demonstrations
Mitigation: 1 FPS sampling captures most visual information

3. Context Understanding

Current: Individual claim verification
Impact: May miss broader narrative context
Mitigation: Video analysis summary provides context

4. Search API Limitations

Current: Dependent on Google Custom Search API
Impact: Results limited by search algorithm bias
Mitigation: Multiple query strategies, diverse sources

5. Real-Time Updates

Current: Analysis is point-in-time
Impact: New evidence published later not included
Mitigation: Reports include generation timestamp

6. Subjective Claims

Current: Best for factual, verifiable claims
Impact: Opinion-based claims harder to verify
Mitigation: Focus on CRAAP criteria, objective claims

Future Enhancements

Multi-Language Support
- Integrate translation APIs
- Support major world languages
- Cross-language evidence validation
Enhanced Visual Analysis
- Deeper computer vision integration
- Object detection and recognition
- Facial recognition for speaker verification
Real-Time Monitoring
- Continuous evidence updates
- Alert system for new contradictory evidence
- Live verification scores
Community Validation
- Crowd-sourced evidence submission
- Expert reviewer network
- Public challenge system
Automated Fact-Checking
- Integration with fact-checking databases
- Automated source verification
- Claim similarity detection
Enhanced Counter-Intelligence
- Social media analysis (Twitter, Reddit)
- Forum scanning (specialized communities)
- Academic preprint servers
Explainability Improvements
- Visual probability explanations
- Interactive evidence exploration
- Confidence intervals on scores

Conclusion

VerityNgn represents a significant advancement in automated video verification technology. By combining multimodal AI analysis, sophisticated counter-intelligence techniques, and probabilistic reasoning, the system provides transparent, evidence-based truthfulness assessments. Key Strengths:

Comprehensive multimodal analysis
Transparent probability calculations
Counter-intelligence innovation
Evidence-based reasoning
Detailed explanations

Ethical Commitment:

Transparent methodology
Source attribution
Probabilistic rather than binary verdicts
Acknowledgment of uncertainty
Open documentation

The system is designed to augment, not replace, human critical thinking. Users are encouraged to:

Review the evidence themselves
Consider the probability distributions
Understand the limitations
Make informed judgments

For Technical Implementation Details, see:

ARCHITECTURE.md - System architecture
API_DOCUMENTATION.md - API reference
DEPLOYMENT_GUIDE.md - Deployment instructions

For Research Papers, see:

papers/verityngn_methodology.pdf - Academic paper
papers/counter_intelligence_techniques.pdf - CI methods paper
papers/probability_model.pdf - Mathematical foundations

This methodology documentation is maintained as part of the VerityNgn open-source project. For updates and contributions, visit the GitHub repository.

Core Engine

​VerityNgn Methodology Documentation

​Executive Summary

​Table of Contents

​1. System Architecture

​Overview

​Technology Stack

​2. Multimodal Analysis Pipeline

​Video Processing Strategy

​Frame-by-Frame Analysis

​CRAAP Criteria Integration

​Prompt Engineering

​Output Format

​3. Claims Extraction Methodology

​Claim Quality Criteria

​Claim Types

​Anti-Patterns (Avoided)

​4. Counter-Intelligence System

​Overview

​Two-Pronged Approach

​A. Press Release Detection

​B. YouTube Counter-Intelligence

​Counter-Intelligence in Action

​5. Probability Distribution Model

​Three-State Model

​Base Probability Initialization

​Enhancement Factors

​Factor 1: Evidence Coverage (10-30% boost)

​Factor 2: Independent Source Ratio (10-25% boost)

​Factor 3: Scientific Evidence (up to 40% boost)

​Factor 4: YouTube Counter-Intelligence (-20% impact)

​Factor 5: Press Release Penalty (-40% impact)

​Normalization

​Verdict Mapping

​6. Evidence Classification & Weighting

​Source Categories

​1. Scientific Sources (Highest Weight: 2.5-4.0)

​2. Independent Sources (High Weight: 1.5-2.0)

​3. Government Sources (High Weight: 1.5-2.5)

​4. Press Releases (Negative Weight: -0.5 to -1.0)

​5. YouTube Counter-Intelligence (Variable: 0.5-2.0)

​6. Other Web Sources (Standard Weight: 1.0)

​Self-Referential Detection

​7. Verification Algorithm

​Claim Verification Process

​Multi-Factor Probability Calculation

​8. Report Generation

​Report Structure

​Output Formats

​Explanation Generation

​9. Limitations & Future Work

​Current Limitations

​1. Language Support

​2. Visual Understanding

​3. Context Understanding

​4. Search API Limitations

​5. Real-Time Updates

​6. Subjective Claims

​Future Enhancements

​Conclusion

VerityNgn Methodology Documentation

Executive Summary

Table of Contents

1. System Architecture

Overview

Technology Stack

2. Multimodal Analysis Pipeline

Video Processing Strategy

Frame-by-Frame Analysis

CRAAP Criteria Integration

Prompt Engineering

Output Format

3. Claims Extraction Methodology

Claim Quality Criteria

Claim Types

Anti-Patterns (Avoided)

4. Counter-Intelligence System

Overview

Two-Pronged Approach

A. Press Release Detection

B. YouTube Counter-Intelligence

Counter-Intelligence in Action

5. Probability Distribution Model

Three-State Model

Base Probability Initialization

Enhancement Factors

Factor 1: Evidence Coverage (10-30% boost)

Factor 2: Independent Source Ratio (10-25% boost)

Factor 3: Scientific Evidence (up to 40% boost)

Factor 4: YouTube Counter-Intelligence (-20% impact)

Factor 5: Press Release Penalty (-40% impact)

Normalization

Verdict Mapping

6. Evidence Classification & Weighting

Source Categories

1. Scientific Sources (Highest Weight: 2.5-4.0)

2. Independent Sources (High Weight: 1.5-2.0)

3. Government Sources (High Weight: 1.5-2.5)

4. Press Releases (Negative Weight: -0.5 to -1.0)

5. YouTube Counter-Intelligence (Variable: 0.5-2.0)

6. Other Web Sources (Standard Weight: 1.0)

Self-Referential Detection

7. Verification Algorithm

Claim Verification Process

Multi-Factor Probability Calculation

8. Report Generation

Report Structure

Output Formats

Explanation Generation

9. Limitations & Future Work

Current Limitations

1. Language Support

2. Visual Understanding

3. Context Understanding

4. Search API Limitations

5. Real-Time Updates

6. Subjective Claims

Future Enhancements

Conclusion