Skip to content

Akanimoh12/TruthLens

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 

Repository files navigation

TruthLens 🔍

Decentralized Fact-Checking & Misinformation Detection Subnet

A Bittensor Subnet for Real-Time, Decentralized Verification of Claims and Information

Bittensor License Status


📋 Table of Contents


Executive Summary

TruthLens is a decentralized fact-checking subnet built on Bittensor that leverages distributed intelligence to verify claims, detect misinformation, and provide transparent, bias-resistant fact-checking at scale.

Key Value Propositions

Feature Benefit
Decentralized Verification Eliminates single-point bias and censorship concerns
Real-Time Processing Sub-minute fact-checking for breaking news and viral content
Transparent Reasoning Every verdict includes traceable evidence and sources
Economic Incentives TAO rewards ensure high-quality, accurate verification
Scalable Architecture Handles millions of claims through distributed miners

Problem Statement

The Misinformation Crisis

Misinformation has become one of the defining challenges of the digital age:

  • Scale: 4.5 billion social media users generate content faster than any centralized system can verify
  • Speed: False information spreads 6x faster than accurate information on social platforms
  • Trust Deficit: 67% of people distrust traditional fact-checkers due to perceived political bias
  • Economic Impact: Misinformation costs the global economy an estimated $78 billion annually
  • Health Risks: Medical misinformation contributed to vaccine hesitancy affecting millions

Limitations of Current Solutions

Current Approach Limitations
Centralized Fact-Checkers Bias concerns, slow turnaround, limited scale, single point of failure
Platform Moderation Opaque decision-making, inconsistent enforcement, censorship concerns
AI-Only Solutions Hallucination risks, lack of real-time data, no accountability
Crowdsourced (e.g., Community Notes) Gaming vulnerabilities, no economic incentives for quality

Why Existing Solutions Fail

┌─────────────────────────────────────────────────────────────────┐
│                    CURRENT FACT-CHECKING LANDSCAPE              │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│   Centralized           Speed vs          Trust                 │
│   Bottleneck    →      Accuracy    →     Deficit               │
│       ↓                    ↓                 ↓                  │
│   Limited Scale      Compromised        User Rejection          │
│                        Quality                                  │
│                                                                 │
│   Result: Misinformation wins by default                        │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Solution Overview

TruthLens: Decentralized Intelligence for Truth

TruthLens creates a marketplace where distributed AI agents (miners) compete to provide the most accurate, well-sourced fact-checks, evaluated by validators using objective quality metrics.

┌─────────────────────────────────────────────────────────────────┐
│                      TRUTHLENS ARCHITECTURE                     │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│    ┌──────────┐     ┌──────────────┐     ┌──────────────┐      │
│    │  CLAIMS  │────▶│   MINERS     │────▶│  VALIDATORS  │      │
│    │  INPUT   │     │  (Research & │     │  (Scoring &  │      │
│    └──────────┘     │   Verify)    │     │  Consensus)  │      │
│                     └──────────────┘     └──────────────┘      │
│                            │                    │               │
│                            ▼                    ▼               │
│                     ┌──────────────────────────────┐           │
│                     │     BITTENSOR BLOCKCHAIN     │           │
│                     │   (Weights & TAO Emissions)  │           │
│                     └──────────────────────────────┘           │
│                                   │                             │
│                                   ▼                             │
│                          ┌──────────────┐                      │
│                          │  TRUTHLENS   │                      │
│                          │     API      │                      │
│                          └──────────────┘                      │
│                                   │                             │
│         ┌─────────────────────────┼─────────────────────────┐  │
│         ▼                         ▼                         ▼  │
│   ┌──────────┐            ┌──────────────┐          ┌────────┐ │
│   │  Media   │            │   Social     │          │  APIs  │ │
│   │ Outlets  │            │  Platforms   │          │        │ │
│   └──────────┘            └──────────────┘          └────────┘ │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Core Innovation: Proof of Verification (PoV)

TruthLens introduces Proof of Verification - a novel consensus mechanism where miners must demonstrate genuine research effort and reasoning quality, not just produce outputs.

PoV Components:

  1. Evidence Trail: All claims must include verifiable source citations
  2. Reasoning Chain: Step-by-step logical analysis must be provided
  3. Confidence Calibration: Miners must accurately estimate their certainty
  4. Cross-Verification: Multiple independent miners verify the same claim

Technical Architecture

System Components

┌────────────────────────────────────────────────────────────────────┐
│                        TRUTHLENS SUBNET                            │
├────────────────────────────────────────────────────────────────────┤
│                                                                    │
│  ┌─────────────────┐    ┌─────────────────┐    ┌────────────────┐ │
│  │  CLAIM INTAKE   │    │  MINER NETWORK  │    │   VALIDATOR    │ │
│  │    SERVICE      │    │                 │    │    NETWORK     │ │
│  ├─────────────────┤    ├─────────────────┤    ├────────────────┤ │
│  │ • API Gateway   │    │ • Research      │    │ • Scoring      │ │
│  │ • Claim Parser  │───▶│   Agents        │───▶│   Engine       │ │
│  │ • Deduplication │    │ • Source        │    │ • Consensus    │ │
│  │ • Priority Queue│    │   Retrieval     │    │   Module       │ │
│  │ • Rate Limiting │    │ • Reasoning     │    │ • Weight       │ │
│  └─────────────────┘    │   Engine        │    │   Submission   │ │
│                         │ • Report Gen    │    └────────────────┘ │
│                         └─────────────────┘                       │
│                                                                    │
│  ┌─────────────────────────────────────────────────────────────┐  │
│  │                    SHARED INFRASTRUCTURE                     │  │
│  ├─────────────────────────────────────────────────────────────┤  │
│  │  • Source Credibility Database    • Claim History Index     │  │
│  │  • Known Facts Repository         • Miner Performance Cache │  │
│  └─────────────────────────────────────────────────────────────┘  │
│                                                                    │
└────────────────────────────────────────────────────────────────────┘

Data Flow

Step 1: CLAIM SUBMISSION
────────────────────────
User/Platform ──▶ API Gateway ──▶ Claim Parser ──▶ Task Queue

Step 2: TASK DISTRIBUTION  
────────────────────────
Task Queue ──▶ Validator ──▶ Broadcasts to Miners

Step 3: MINER PROCESSING
────────────────────────
Miner receives claim ──▶ Research & Analysis ──▶ Generate Report

Step 4: RESPONSE COLLECTION
────────────────────────
Miners submit reports ──▶ Validator collects responses

Step 5: VALIDATION & SCORING
────────────────────────
Validator scores reports ──▶ Consensus calculation ──▶ Weight submission

Step 6: RESULT DELIVERY
────────────────────────
Final verdict ──▶ API Response ──▶ User/Platform

Claim Categories

TruthLens handles multiple claim types with specialized processing:

Category Examples Processing Approach
Factual "The Eiffel Tower is 300m tall" Database lookup + source verification
Statistical "Crime rates increased 50%" Data source verification + context analysis
Quote Attribution "Einstein said X" Primary source research + quote databases
Scientific "Vaccine causes Y" Peer-reviewed literature analysis
Current Events "Country X invaded Y" Multi-source news verification
Historical "Event X happened in year Y" Academic source verification

Incentive & Mechanism Design

Overview

The TruthLens incentive mechanism is designed to reward accurate, well-sourced, and timely fact-checking while penalizing low-quality or adversarial behavior.

Emission and Reward Logic

┌────────────────────────────────────────────────────────────────┐
│                    REWARD DISTRIBUTION                         │
├────────────────────────────────────────────────────────────────┤
│                                                                │
│   Total Subnet Emissions (E_subnet)                            │
│            │                                                   │
│            ├──▶ 41% ──▶ Validators (V_emissions)               │
│            │                                                   │
│            └──▶ 59% ──▶ Miners (M_emissions)                   │
│                            │                                   │
│                            ├──▶ Accuracy Score (40%)           │
│                            ├──▶ Source Quality (25%)           │
│                            ├──▶ Response Time (15%)            │
│                            ├──▶ Reasoning Quality (15%)        │
│                            └──▶ Consensus Alignment (5%)       │
│                                                                │
└────────────────────────────────────────────────────────────────┘

Miner Scoring Formula

Each miner's score is calculated as:

S_miner = (W_acc × Accuracy) + (W_src × SourceQuality) + (W_time × TimeScore) + 
          (W_reason × ReasoningScore) + (W_cons × ConsensusAlignment)

Where:
  W_acc    = 0.40  (Accuracy weight)
  W_src    = 0.25  (Source quality weight)
  W_time   = 0.15  (Response time weight)
  W_reason = 0.15  (Reasoning quality weight)
  W_cons   = 0.05  (Consensus alignment weight)

Component Definitions

1. Accuracy Score (40%)

def calculate_accuracy(miner_verdict, ground_truth, confidence):
    """
    Accuracy rewards correct verdicts and penalizes overconfidence in wrong answers.
    
    Verdicts: TRUE, FALSE, PARTIALLY_TRUE, UNVERIFIABLE
    Confidence: 0.0 to 1.0
    """
    if miner_verdict == ground_truth:
        # Reward correct verdicts, bonus for appropriate confidence
        base_score = 1.0
        confidence_bonus = 0.1 * confidence  # Higher confidence = higher reward when correct
        return min(1.0, base_score + confidence_bonus)
    else:
        # Penalize incorrect verdicts, especially high-confidence wrong answers
        penalty = 0.5 * confidence  # Higher confidence = higher penalty when wrong
        return max(0.0, 0.3 - penalty)

Ground Truth Determination:

  • Consensus among top-performing miners (weighted by historical accuracy)
  • Validator injection of known claims (honeypots) for calibration
  • Post-hoc verification using authoritative sources

2. Source Quality Score (25%)

def calculate_source_quality(sources):
    """
    Evaluates the quality and credibility of cited sources.
    """
    score = 0.0
    for source in sources:
        credibility = get_source_credibility(source.domain)  # 0.0 to 1.0
        recency = calculate_recency_score(source.date)       # 0.0 to 1.0
        relevance = calculate_relevance(source, claim)       # 0.0 to 1.0
        accessibility = 1.0 if is_accessible(source.url) else 0.5
        
        source_score = (0.4 * credibility + 0.3 * relevance + 
                       0.2 * recency + 0.1 * accessibility)
        score += source_score
    
    # Normalize and bonus for multiple independent sources
    diversity_bonus = min(0.2, 0.05 * count_unique_domains(sources))
    return min(1.0, (score / len(sources)) + diversity_bonus)

Source Credibility Tiers:

Tier Score Examples
Tier 1 0.9-1.0 Peer-reviewed journals, government statistics, court records
Tier 2 0.7-0.9 Major news outlets (AP, Reuters), academic institutions
Tier 3 0.5-0.7 Regional news, verified expert blogs, official company statements
Tier 4 0.3-0.5 Wikipedia, social media verified accounts
Tier 5 0.0-0.3 Unverified sources, anonymous claims, known unreliable sources

3. Response Time Score (15%)

def calculate_time_score(response_time_seconds, claim_complexity):
    """
    Rewards fast responses while accounting for claim complexity.
    """
    # Base time expectations by complexity
    base_times = {
        'simple': 30,      # Simple factual lookup
        'moderate': 120,   # Requires some research
        'complex': 300,    # Deep research needed
    }
    
    expected_time = base_times[claim_complexity]
    
    if response_time <= expected_time * 0.5:
        return 1.0  # Excellent speed
    elif response_time <= expected_time:
        return 0.8 + 0.2 * (1 - response_time / expected_time)
    elif response_time <= expected_time * 2:
        return 0.5 * (1 - (response_time - expected_time) / expected_time)
    else:
        return 0.1  # Minimum score for late responses

4. Reasoning Quality Score (15%)

def calculate_reasoning_score(reasoning_chain):
    """
    Evaluates the logical coherence and completeness of reasoning.
    """
    scores = {
        'logical_coherence': evaluate_logic(reasoning_chain),      # No contradictions
        'evidence_citation': check_evidence_links(reasoning_chain), # Claims backed by evidence
        'completeness': evaluate_completeness(reasoning_chain),     # Addresses all aspects
        'clarity': evaluate_clarity(reasoning_chain),               # Clear, understandable
    }
    
    return (0.35 * scores['logical_coherence'] + 
            0.30 * scores['evidence_citation'] +
            0.20 * scores['completeness'] + 
            0.15 * scores['clarity'])

5. Consensus Alignment Score (5%)

def calculate_consensus_alignment(miner_verdict, all_verdicts, miner_weights):
    """
    Small bonus for alignment with high-quality miners.
    Prevents gaming by weighting by historical performance.
    """
    weighted_consensus = calculate_weighted_consensus(all_verdicts, miner_weights)
    
    if miner_verdict == weighted_consensus:
        return 1.0
    elif is_adjacent_verdict(miner_verdict, weighted_consensus):
        return 0.5  # e.g., TRUE vs PARTIALLY_TRUE
    else:
        return 0.0

Incentive Alignment

For Miners

Desired Behavior Incentive Mechanism
Provide accurate verdicts Accuracy score (40% weight) directly rewards correctness
Use high-quality sources Source quality score rewards credible, diverse sources
Respond quickly Time score rewards fast responses
Provide clear reasoning Reasoning score rewards logical, complete explanations
Maintain consistency Historical accuracy affects future task allocation

For Validators

Desired Behavior Incentive Mechanism
Score honestly Validators earn more when their miners perform well network-wide
Inject quality honeypots Better calibration = better miner selection = higher returns
Stay active Activity cutoff removes inactive validators from consensus

Anti-Gaming Mechanisms

1. Honeypot Claims

class HoneypotSystem:
    """
    Validators inject claims with known ground truth to catch cheaters.
    """
    def generate_honeypot(self):
        honeypot_types = [
            'known_true',      # Verified true claims
            'known_false',     # Verified false claims  
            'synthetic',       # AI-generated with clear answers
            'historical',      # Past claims with established verdicts
        ]
        
        # 10-20% of all tasks are honeypots
        return create_honeypot(random.choice(honeypot_types))
    
    def evaluate_honeypot_response(self, miner_response, ground_truth):
        if miner_response.verdict != ground_truth:
            return apply_penalty(severity='high')
        return apply_bonus(severity='low')

2. Sybil Resistance

def detect_sybil_behavior(miners):
    """
    Detect and penalize coordinated manipulation.
    """
    # Check for suspiciously similar responses
    similarity_matrix = compute_response_similarity(miners)
    
    # Check for coordinated timing
    timing_correlation = analyze_response_timing(miners)
    
    # Check for shared infrastructure
    infrastructure_overlap = detect_shared_infrastructure(miners)
    
    if any_suspicious_patterns(similarity_matrix, timing_correlation, infrastructure_overlap):
        flag_for_review(miners)
        reduce_weights(miners)

3. Confidence Calibration

def penalize_miscalibration(miner_history):
    """
    Miners who are consistently overconfident or underconfident are penalized.
    """
    calibration_error = calculate_calibration_error(
        miner_history.confidences,
        miner_history.accuracies
    )
    
    # Perfect calibration = 0, worst = 1
    # e.g., always saying 90% confident but only 50% accurate = high error
    
    return 1.0 - calibration_error

4. Source Verification

def verify_sources(miner_response):
    """
    Verify that cited sources actually exist and support the claims.
    """
    penalties = 0
    
    for source in miner_response.sources:
        # Check if URL is accessible
        if not is_accessible(source.url):
            penalties += 0.1
        
        # Check if source content matches citation
        if not verify_citation_accuracy(source):
            penalties += 0.2
        
        # Check if source is fabricated
        if is_likely_fabricated(source):
            penalties += 0.5
    
    return max(0, 1.0 - penalties)

Proof of Verification (PoV) - Why This Qualifies

TruthLens implements genuine Proof of Intelligence/Effort because:

  1. Non-Trivial Computation: Research, source verification, and reasoning cannot be faked with simple pattern matching
  2. Verifiable Work: Sources can be checked, reasoning can be evaluated, accuracy can be measured
  3. Economic Rationality: Gaming is more expensive than honest participation due to honeypots
  4. Skill Differentiation: Better researchers/AI systems naturally outperform, creating meritocracy

Miner Design

Miner Architecture

┌─────────────────────────────────────────────────────────────────┐
│                       TRUTHLENS MINER                           │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                    INPUT PROCESSING                      │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Claim Parser         • Entity Extraction              │   │
│  │  • Complexity Classifier • Context Analyzer              │   │
│  └─────────────────────────────────────────────────────────┘   │
│                              │                                  │
│                              ▼                                  │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                    RESEARCH ENGINE                       │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Web Search Module    • Academic Database Connector    │   │
│  │  • News API Integration • Government Data Access         │   │
│  │  • Social Media Monitor • Fact-Check DB Query            │   │
│  └─────────────────────────────────────────────────────────┘   │
│                              │                                  │
│                              ▼                                  │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                    REASONING ENGINE                      │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Evidence Aggregator  • Contradiction Detector         │   │
│  │  • Logical Analyzer     • Confidence Estimator           │   │
│  │  • Verdict Generator    • Explanation Builder            │   │
│  └─────────────────────────────────────────────────────────┘   │
│                              │                                  │
│                              ▼                                  │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                    OUTPUT GENERATION                     │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Report Formatter     • Source Compiler                │   │
│  │  • Quality Checker      • Response Packager              │   │
│  └─────────────────────────────────────────────────────────┘   │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Miner Tasks

Miners receive claims and must return comprehensive fact-check reports.

Input Format

{
  "task_id": "uuid-v4",
  "claim": {
    "text": "The Great Wall of China is visible from space with the naked eye.",
    "source": "social_media_post",
    "context": "Viral post claiming this as fact",
    "timestamp": "2026-01-15T10:30:00Z",
    "language": "en"
  },
  "metadata": {
    "category": "factual",
    "complexity": "moderate",
    "priority": "normal",
    "deadline_seconds": 120
  }
}

Expected Output Format

{
  "task_id": "uuid-v4",
  "response": {
    "verdict": "FALSE",
    "confidence": 0.95,
    "summary": "The Great Wall of China is NOT visible from space with the naked eye. This is a common misconception that has been debunked by astronauts and scientists.",
    "detailed_analysis": {
      "claim_breakdown": [
        "The claim asserts the Great Wall is visible from space",
        "Specifically mentions 'naked eye' visibility"
      ],
      "evidence_analysis": [
        {
          "point": "Astronaut testimonies",
          "finding": "Multiple astronauts including Chris Hadfield have confirmed the Wall is not visible from orbit without aid",
          "source_ids": ["src_1", "src_2"]
        },
        {
          "point": "Physical dimensions",
          "finding": "The Wall is only 5-8 meters wide, far too narrow to resolve from orbital distances (400km+)",
          "source_ids": ["src_3"]
        }
      ],
      "reasoning_chain": [
        "1. Human visual acuity limit is ~1 arcminute",
        "2. From ISS altitude (400km), this resolves to ~116 meters",
        "3. Great Wall width (5-8m) is 15-20x smaller than resolvable limit",
        "4. Therefore, naked eye visibility is physically impossible"
      ],
      "counterarguments_addressed": [
        {
          "counterargument": "Some claim to have seen it",
          "rebuttal": "Studies show these were likely rivers or roads mistaken for the Wall"
        }
      ]
    },
    "sources": [
      {
        "id": "src_1",
        "url": "https://www.nasa.gov/vision/space/workinginspace/great_wall.html",
        "title": "Great Wall of China - NASA",
        "type": "government",
        "credibility_tier": 1,
        "relevant_quote": "The Great Wall can barely be seen from the Shuttle..."
      },
      {
        "id": "src_2", 
        "url": "https://www.scientificamerican.com/article/china-great-wall-space/",
        "title": "Is China's Great Wall Visible from Space?",
        "type": "scientific_publication",
        "credibility_tier": 2,
        "relevant_quote": "...not visible to the unaided eye from orbit"
      },
      {
        "id": "src_3",
        "url": "https://academic.example.com/visual-acuity-limits",
        "title": "Human Visual Acuity: Physical Limits",
        "type": "peer_reviewed",
        "credibility_tier": 1,
        "relevant_quote": "Angular resolution limit of ~1 arcminute..."
      }
    ],
    "metadata": {
      "processing_time_ms": 45000,
      "sources_consulted": 12,
      "model_version": "truthlens-miner-v1.0"
    }
  }
}

Performance Dimensions

Dimension Measurement Target
Accuracy % of correct verdicts >90%
Source Depth Avg. sources per claim 3-5
Response Time Seconds to respond <60s simple, <180s complex
Reasoning Completeness % of aspects addressed >85%
Calibration Confidence vs actual accuracy <10% error

Recommended Miner Implementation

class TruthLensMiner:
    """
    Reference implementation for TruthLens miners.
    """
    
    def __init__(self, config):
        self.research_engine = ResearchEngine(config.search_apis)
        self.reasoning_engine = ReasoningEngine(config.llm_model)
        self.source_validator = SourceValidator()
        
    async def process_claim(self, task: Task) -> Response:
        # Step 1: Parse and understand the claim
        parsed_claim = self.parse_claim(task.claim)
        
        # Step 2: Conduct research
        research_results = await self.research_engine.search(
            query=parsed_claim.search_queries,
            time_budget=task.metadata.deadline_seconds * 0.6
        )
        
        # Step 3: Validate sources
        validated_sources = self.source_validator.validate(research_results)
        
        # Step 4: Analyze and reason
        analysis = self.reasoning_engine.analyze(
            claim=parsed_claim,
            evidence=validated_sources
        )
        
        # Step 5: Generate verdict with confidence
        verdict = self.reasoning_engine.generate_verdict(analysis)
        
        # Step 6: Compile report
        return self.compile_response(task.task_id, verdict, analysis, validated_sources)

Validator Design

Validator Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      TRUTHLENS VALIDATOR                        │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                    TASK MANAGEMENT                       │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Task Queue Manager   • Priority Scheduler             │   │
│  │  • Honeypot Injector    • Task Broadcaster               │   │
│  └─────────────────────────────────────────────────────────┘   │
│                              │                                  │
│                              ▼                                  │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                   RESPONSE COLLECTION                    │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Response Aggregator  • Timeout Handler                │   │
│  │  • Format Validator     • Duplicate Detector             │   │
│  └─────────────────────────────────────────────────────────┘   │
│                              │                                  │
│                              ▼                                  │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                    SCORING ENGINE                        │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Accuracy Evaluator   • Source Quality Scorer          │   │
│  │  • Time Score Calculator• Reasoning Analyzer             │   │
│  │  • Consensus Calculator • Final Score Aggregator         │   │
│  └─────────────────────────────────────────────────────────┘   │
│                              │                                  │
│                              ▼                                  │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                   WEIGHT MANAGEMENT                      │   │
│  ├─────────────────────────────────────────────────────────┤   │
│  │  • Weight Calculator    • Normalization                  │   │
│  │  • Weight Committer     • History Tracker                │   │
│  └─────────────────────────────────────────────────────────┘   │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Scoring and Evaluation Methodology

class TruthLensValidator:
    """
    Reference implementation for TruthLens validators.
    """
    
    def __init__(self, config):
        self.scoring_weights = {
            'accuracy': 0.40,
            'source_quality': 0.25,
            'response_time': 0.15,
            'reasoning': 0.15,
            'consensus': 0.05
        }
        self.honeypot_manager = HoneypotManager()
        self.source_credibility_db = SourceCredibilityDB()
        
    async def evaluate_responses(self, task: Task, responses: List[MinerResponse]) -> Dict[str, float]:
        scores = {}
        
        # Determine ground truth (consensus or honeypot)
        ground_truth = self.determine_ground_truth(task, responses)
        
        for response in responses:
            miner_score = self.calculate_miner_score(
                response=response,
                ground_truth=ground_truth,
                task=task,
                all_responses=responses
            )
            scores[response.miner_uid] = miner_score
            
        return self.normalize_scores(scores)
    
    def calculate_miner_score(self, response, ground_truth, task, all_responses):
        accuracy = self.score_accuracy(response.verdict, ground_truth, response.confidence)
        source_quality = self.score_sources(response.sources)
        time_score = self.score_time(response.processing_time, task.complexity)
        reasoning = self.score_reasoning(response.reasoning_chain)
        consensus = self.score_consensus(response.verdict, all_responses)
        
        return (
            self.scoring_weights['accuracy'] * accuracy +
            self.scoring_weights['source_quality'] * source_quality +
            self.scoring_weights['response_time'] * time_score +
            self.scoring_weights['reasoning'] * reasoning +
            self.scoring_weights['consensus'] * consensus
        )

Evaluation Cadence

┌────────────────────────────────────────────────────────────────┐
│                    EVALUATION TIMELINE                         │
├────────────────────────────────────────────────────────────────┤
│                                                                │
│  Block N         Block N+1       Block N+2       Block N+360  │
│     │               │               │               │         │
│     ▼               ▼               ▼               ▼         │
│  ┌──────┐       ┌──────┐       ┌──────┐       ┌──────────┐   │
│  │ Task │       │ Task │       │ Task │       │  Weight  │   │
│  │Assign│       │Assign│       │Assign│       │ Commit   │   │
│  └──────┘       └──────┘       └──────┘       └──────────┘   │
│     │               │               │               │         │
│     ▼               ▼               ▼               ▼         │
│  Collect         Collect         Collect        Aggregate    │
│  & Score         & Score         & Score        & Submit     │
│                                                                │
│  Continuous task assignment, periodic weight commits          │
│  (every tempo ~360 blocks / ~72 minutes)                      │
│                                                                │
└────────────────────────────────────────────────────────────────┘

Validator Incentive Alignment

Validators are incentivized to score honestly because:

  1. Yuma Consensus: Validators who agree with the consensus of other validators earn more
  2. Miner Performance: Validators earn dividends from their high-performing miners
  3. Stake Requirements: 1000+ stake weight required, creating accountability
  4. Reputation: Historical accuracy affects stake delegation from TAO holders

Business Logic & Market Rationale

The Problem TruthLens Solves

Stakeholder Problem TruthLens Solution
Social Platforms Misinformation liability, moderation costs API for automated fact-checking
News Organizations Slow verification, limited resources Real-time verification assistance
Enterprises Brand safety, due diligence Claim verification for decision-making
Researchers Data quality issues Verified information feeds
General Public Confusion, trust deficit Transparent, verifiable fact-checks

Competitive Analysis

Within Bittensor Ecosystem

Subnet Focus Differentiation from TruthLens
Subnet 1 (Prompting) General LLM inference TruthLens: Specialized verification + source validation
Subnet 18 (Cortex.t) AI generation TruthLens: Verification, not generation
Subnet 8 (Taoshi) Financial prediction TruthLens: Factual claims, not forecasting

TruthLens Unique Position: No existing Bittensor subnet focuses on fact-verification with source validation and transparent reasoning.

Outside Bittensor

Competitor Model TruthLens Advantage
Snopes, PolitiFact Centralized editorial Decentralized, scalable, no single-point bias
Google Fact Check Tools Aggregation only Active verification, not just aggregation
Community Notes (X) Crowdsourced Economic incentives, professional-grade quality
AI Fact-Checkers Single model Distributed intelligence, multiple perspectives
Full Fact Manual + AI Fully automated, real-time, global scale

Why Bittensor is the Right Platform

  1. Decentralized Trust: No single entity controls verdicts
  2. Economic Incentives: TAO rewards attract quality miners
  3. Transparent Consensus: Publicly verifiable scoring
  4. Scalability: 256 nodes can handle massive claim volume
  5. Anti-Censorship: No single point of censorship failure
  6. Global Distribution: Miners worldwide = diverse perspectives

Path to Sustainable Business

┌────────────────────────────────────────────────────────────────┐
│                    REVENUE MODEL                               │
├────────────────────────────────────────────────────────────────┤
│                                                                │
│  PHASE 1: Foundation (Year 1)                                  │
│  ├── TAO emissions fund initial operations                     │
│  └── Build reputation and accuracy track record                │
│                                                                │
│  PHASE 2: API Monetization (Year 2)                            │
│  ├── Freemium API: 100 claims/day free, paid tiers above       │
│  ├── Enterprise contracts: Custom SLAs, dedicated capacity     │
│  └── Revenue flows to subnet, increasing TAO value             │
│                                                                │
│  PHASE 3: Platform Partnerships (Year 2-3)                     │
│  ├── Integration with social platforms                         │
│  ├── News organization partnerships                            │
│  └── Government/institutional contracts                        │
│                                                                │
│  PHASE 4: Data Products (Year 3+)                              │
│  ├── Misinformation trend reports                              │
│  ├── Source credibility ratings                                │
│  └── Research datasets                                         │
│                                                                │
└────────────────────────────────────────────────────────────────┘

Market Size

Segment TAM SAM SOM (Year 3)
Content Moderation $15B $3B $30M
Media Monitoring $5B $1B $10M
Due Diligence $8B $500M $5M
Research Data $2B $200M $2M
Total $30B $4.7B $47M

Go-To-Market Strategy

Phase 1: Initial Target Users & Use Cases

Primary Early Adopters

  1. Independent Journalists & Fact-Checkers

    • Pain point: Limited resources for verification
    • Value prop: Free tier API, augments their workflow
    • Acquisition: Twitter/X outreach, journalism conferences
  2. Web3/Crypto Community

    • Pain point: Rampant scams and misinformation
    • Value prop: Verify project claims, rug pull warnings
    • Acquisition: Bittensor community, crypto Twitter
  3. Academic Researchers

    • Pain point: Need verified data for studies
    • Value prop: Transparent, reproducible verification
    • Acquisition: Research partnerships, preprint sharing

Anchor Use Cases

Use Case Description Success Metric
Crypto Scam Detection Verify claims made by crypto projects 1000 scam warnings/month
News Verification Real-time verification for breaking news 100 news org users
Quote Verification Verify attributed quotes 10K verifications/month

Phase 2: Distribution & Growth Channels

┌────────────────────────────────────────────────────────────────┐
│                    GROWTH STRATEGY                             │
├────────────────────────────────────────────────────────────────┤
│                                                                │
│  CHANNEL 1: Developer Community                                │
│  ├── Open source miner/validator code                          │
│  ├── Comprehensive API documentation                           │
│  ├── Developer grants for integrations                         │
│  └── Hackathon sponsorships                                    │
│                                                                │
│  CHANNEL 2: Content & Thought Leadership                       │
│  ├── Weekly misinformation trend reports                       │
│  ├── Case studies on successful verifications                  │
│  ├── Research papers on decentralized fact-checking            │
│  └── Podcast/media appearances                                 │
│                                                                │
│  CHANNEL 3: Strategic Partnerships                             │
│  ├── Integration with existing fact-check organizations        │
│  ├── Browser extension partnerships                            │
│  ├── CMS/publishing platform plugins                           │
│  └── Social media tool integrations                            │
│                                                                │
│  CHANNEL 4: Community Building                                 │
│  ├── Discord community for users and miners                    │
│  ├── Ambassador program                                        │
│  ├── Bug bounty and accuracy challenges                        │
│  └── Governance participation                                  │
│                                                                │
└────────────────────────────────────────────────────────────────┘

Phase 3: Incentives for Early Participation

For Miners

Incentive Details
Early Miner Bonus 2x TAO multiplier for first 50 miners (Month 1-3)
Accuracy Leaderboard Monthly prizes for top accuracy miners
Hardware Grants GPU grants for promising miners
Open Source Bonus Extra emissions for miners who open-source improvements

For Validators

Incentive Details
Founding Validator NFT Recognition for first 10 validators
Delegation Matching Protocol matches early validator stakes
Governance Rights Early validators get enhanced governance weight

For Users/API Consumers

Incentive Details
Generous Free Tier 500 claims/day free for first year
Integration Bounties Paid bounties for platform integrations
Feedback Rewards TAO rewards for quality feedback on verdicts

Launch Timeline

┌────────────────────────────────────────────────────────────────┐
│                    LAUNCH TIMELINE                             │
├────────────────────────────────────────────────────────────────┤
│                                                                │
│  MONTH 1-2: Private Alpha                                      │
│  ├── Core team + 5-10 selected miners                          │
│  ├── Basic claim types only                                    │
│  └── Iterate on scoring mechanisms                             │
│                                                                │
│  MONTH 3-4: Public Beta (Testnet)                              │
│  ├── Open miner registration                                   │
│  ├── Public API (rate limited)                                 │
│  └── Community feedback integration                            │
│                                                                │
│  MONTH 5-6: Mainnet Launch                                     │
│  ├── Full subnet deployment                                    │
│  ├── Marketing push                                            │
│  └── First enterprise pilot                                    │
│                                                                │
│  MONTH 7-12: Scale & Iterate                                   │
│  ├── Expand claim categories                                   │
│  ├── Multi-language support                                    │
│  └── Advanced features (batch API, webhooks)                   │
│                                                                │
└────────────────────────────────────────────────────────────────┘

Roadmap

Development Phases

┌────────────────────────────────────────────────────────────────────────────┐
│                           TRUTHLENS ROADMAP                                │
├────────────────────────────────────────────────────────────────────────────┤
│                                                                            │
│  Q1 2026                    Q2 2026                    Q3 2026            │
│  ════════                   ════════                   ════════           │
│  ☑ Subnet Design            ☐ Testnet Deployment       ☐ Mainnet Launch   │
│  ☑ Incentive Mechanism      ☐ Miner SDK v1.0          ☐ API v1.0         │
│  ☑ Architecture Design      ☐ Validator SDK v1.0      ☐ 50+ Active Miners│
│  ☐ Core Protocol Dev        ☐ Public Beta             ☐ First Enterprise │
│                             ☐ 10+ Test Miners          ☐ Multi-language   │
│                                                                            │
│  Q4 2026                    Q1 2027                    Q2 2027+           │
│  ════════                   ════════                   ════════           │
│  ☐ API v2.0 (Batch/Stream)  ☐ Platform Integrations   ☐ Mobile SDK       │
│  ☐ 100+ Active Miners       ☐ Research Partnerships   ☐ Governance DAO   │
│  ☐ Browser Extension        ☐ Advanced Analytics      ☐ Global Expansion │
│  ☐ 5 Enterprise Clients     ☐ 1M+ Verifications/mo    ☐ 10M+ Verif/mo    │
│                                                                            │
└────────────────────────────────────────────────────────────────────────────┘

Key Milestones

Milestone Target Date Success Criteria
Testnet Launch March 2026 Functional subnet on Bittensor testnet
10 Active Miners April 2026 10 miners consistently producing quality results
Mainnet Launch June 2026 Subnet live on Bittensor mainnet
1000 Daily Verifications August 2026 Consistent daily usage
First Paying Customer September 2026 Enterprise API contract signed
90% Accuracy December 2026 Network achieves 90%+ accuracy on benchmarks

Technical Specifications

Hardware Requirements

Miner Requirements

Component Minimum Recommended
CPU 4 cores 8+ cores
RAM 16 GB 32 GB
Storage 100 GB SSD 500 GB NVMe
Network 100 Mbps 1 Gbps
GPU Optional RTX 3080+ (for local LLM)

Validator Requirements

Component Minimum Recommended
CPU 8 cores 16+ cores
RAM 32 GB 64 GB
Storage 500 GB SSD 1 TB NVMe
Network 500 Mbps 1 Gbps
Stake 1000 TAO stake weight 5000+ TAO stake weight

API Specifications

REST API Endpoints

POST /api/v1/verify
  - Submit a claim for verification
  - Returns: task_id, estimated_time

GET /api/v1/verify/{task_id}
  - Get verification result
  - Returns: verdict, confidence, sources, reasoning

POST /api/v1/verify/batch
  - Submit multiple claims
  - Returns: array of task_ids

GET /api/v1/sources/{domain}
  - Get credibility rating for a source
  - Returns: credibility_score, tier, history

GET /api/v1/stats
  - Network statistics
  - Returns: active_miners, avg_response_time, accuracy_metrics

WebSocket API

WS /api/v1/stream
  - Real-time verification updates
  - Subscribe to claim categories
  - Receive instant verdicts

Team

[Akanimoh - Team Lead]

Roles Needed

Role Responsibility
Subnet Lead Overall architecture, Bittensor integration
ML Engineer Reasoning engine, evaluation models
Backend Engineer API, infrastructure, scaling
Research Lead Source credibility, evaluation methodology
DevRel/Community Documentation, miner support, partnerships

References

Bittensor Resources

Fact-Checking Research

Related Projects


License

This project is licensed under the MIT License - see the LICENSE file for details.


Contact


Built for the Bittensor Subnet Ideathon

Decentralizing Truth, One Claim at a Time 🔍

About

Decentralized Fact-Checking & Misinformation Detection Subnet

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published