POC Requirements

Status: ✅ Approved for Development  
Version: 3.0 (Aligned with Main Requirements)  
Goal: Prove that AI can extract claims and determine verdicts automatically without human intervention

1. POC Overview

1.1 What POC Tests

Core Question:

 Can AI automatically extract factual claims from articles and evaluate them with reasonable verdicts?

What we're proving:

• AI can identify factual claims from text
• AI can evaluate those claims with structured evidence
• Quality gates can filter unreliable outputs
• The core workflow is technically feasible

What we're NOT proving:

• Production-ready reliability (that's POC2)
• User-facing features (that's Beta 0)
• Full IFCN compliance (that's V1.0)

1.2 Requirements Mapping

POC1 implements a subset of the full system requirements defined in Main Requirements.

Scope Summary:

• In Scope: 8 requirements (7 FRs + 1 NFR)
• Partial: 3 NFRs (simplified versions)
• Out of Scope: 19 requirements (deferred to later phases)

2. POC1 Scope

POC1 implements these formal requirements:

POC1 also implements these workflow components (detailed as FR1-FR6 in implementation sections below)

InformationNote: FR11 (Audit Trail) and FR13 (In-Article Claim Highlighting) are deferred to Beta 0 for production readiness and user experience enhancement.:

• Claim extraction (FR1)
• Claim context (FR2)
• Multiple scenarios (FR3)
• Evidence collection (FR5)
• Source quality assessment (FR6)
• Time evolution tracking (FR8) - deferred to POC2
• Audit trail (FR11) - deferred to Beta 0
• In-article highlighting (FR13) - deferred to Beta 0

Partial implementations:

• NFR1 (Explainability) - Basic only
• NFR2 (Performance) - Functional but not optimized
• NFR3 (Transparency) - Basic only

Detailed POC1 implementation specifications continue below...

3. POC Simplifications

3.1 FR1: Claim Extraction (Full Implementation)

Main Requirement: AI extracts factual claims from input text

POC Implementation:

• ✅ AKEL extracts claims using LLM
• ✅ Each claim includes original text reference
• ✅ Claims are identified as factual/non-factual
• ❌ No advanced claim parsing (added in POC2)

Acceptance Criteria:

• Extracts 3-5 claims from typical article
• Identifies factual vs non-factual claims
• Quality Gate 1 validates extraction

3.2 FR3: Multiple Scenarios (Full Implementation)

Main Requirement: Generate multiple interpretation scenarios for ambiguous claims

POC Implementation:

• ✅ AKEL generates 2-3 scenarios per claim
• ✅ Scenarios capture different interpretations
• ✅ Each scenario is evaluated separately
• ✅ Verdict considers all scenarios

Acceptance Criteria:

• Generates 2+ scenarios for ambiguous claims
• Scenarios are meaningfully different
• All scenarios are evaluated

3.3 FR4: Analysis Summary (Basic Implementation)

Main Requirement: Provide user-friendly summary of analysis

POC Implementation:

• ✅ Simple text summary generated
• ❌ No rich formatting (added in Beta 0)
• ❌ No visual elements (added in Beta 0)
• ❌ No interactive features (added in Beta 0)

POC Format:
```
Claim: [extracted claim]
Scenarios: [list of scenarios]
Evidence: [supporting/opposing evidence]
Verdict: [probability with uncertainty]
```

3.4 FR5-FR6: Evidence Collection & Evaluation (Full Implementation)

Main Requirements:

• FR5: Collect supporting and opposing evidence
• FR6: Evaluate evidence source reliability

POC Implementation:

• ✅ AKEL searches for evidence (web/knowledge base)
• ✅ Mandatory contradiction search (finds opposing evidence)
• ✅ Source reliability scoring
• ❌ No evidence deduplication (added in POC2)
• ❌ No advanced source verification (added in POC2)

Acceptance Criteria:

• Finds 2+ supporting evidence items
• Finds 1+ opposing evidence (if exists)
• Sources scored for reliability

3.5 FR7: Automated Verdicts (Full Implementation)

Main Requirement: AI computes verdicts with uncertainty quantification

POC Implementation:

• ✅ Probabilistic verdicts (0-100% confidence)
• ✅ Uncertainty explicitly stated
• ✅ Reasoning chain provided
• ✅ Quality Gate 4 validates verdict confidence

POC Output:
```
Verdict: 70% likely true
Uncertainty: ±15% (moderate confidence)
Reasoning: Based on 3 high-quality sources...
Confidence Level: MEDIUM
```

Acceptance Criteria:

• Verdicts include probability (0-100%)
• Uncertainty explicitly quantified
• Reasoning chain explains verdict

3.6 NFR11: Quality Assurance Framework (LITE VERSION)

Main Requirement: Complete quality assurance with 7 quality gates

POC Implementation: 2 gates only

Quality Gate 1: Claim Validation

• ✅ Validates claim is factual and verifiable
• ✅ Blocks non-factual claims (opinion/prediction/ambiguous)
• ✅ Provides clear rejection reason

Quality Gate 4: Verdict Confidence Assessment

• ✅ Validates ≥2 sources found
• ✅ Validates quality score ≥0.6
• ✅ Blocks low-confidence verdicts
• ✅ Provides clear rejection reason

Out of Scope (POC2+):

• ❌ Gate 2: Evidence Relevance
• ❌ Gate 3: Scenario Coherence
• ❌ Gate 5: Source Diversity
• ❌ Gate 6: Reasoning Validity
• ❌ Gate 7: Output Completeness

Rationale: Prove gate concept works. Add remaining gates in POC2 after validating approach.

3.7 NFR1-3: Performance, Scalability, Reliability (Basic)

Main Requirements:

• NFR1: Response time < 30 seconds
• NFR2: Handle 1000+ concurrent users
• NFR3: 99.9% uptime

POC Implementation:

• ⚠️ Response time monitored (not optimized)
• ⚠️ Single-threaded processing (no concurrency)
• ⚠️ Basic error handling (no advanced retry logic)

Rationale: POC proves functionality. Performance optimization happens in POC2.

POC Acceptance:

• Analysis completes (no timeout requirement)
• Errors don't crash system
• Basic logging in place

4. What's NOT in POC Scope

4.1 User-Facing Features (Beta 0+)

Out of Scope:

• ❌ User accounts and authentication (FR8)
• ❌ User corrections system (FR9, FR45-46)
• ❌ Public publishing interface (FR10)
• ❌ Social sharing (FR11)
• ❌ Email notifications (FR12)
• ❌ API access (FR13)

Rationale: POC validates AI capabilities. User features added in Beta 0.

4.2 Advanced Features (V1.0+)

Out of Scope:

• ❌ IFCN compliance (FR47)
• ❌ ClaimReview schema (FR48)
• ❌ Archive.org integration (FR49)
• ❌ OSINT toolkit (FR50)
• ❌ Video verification (FR51)
• ❌ Deepfake detection (FR52)
• ❌ Cross-org sharing (FR53)

Rationale: Advanced features require proven platform. Added post-V1.0.

4.3 Production Requirements (POC2, Beta 0)

Out of Scope:

• ❌ Security controls (NFR4, NFR12)
• ❌ Code maintainability (NFR5)
• ❌ System monitoring (NFR13)
• ❌ Evidence deduplication
• ❌ Advanced source verification
• ❌ Full 7-gate quality framework

Rationale: POC proves concept. Production hardening happens in POC2 and Beta 0.

5. POC Output Specification

5.1 Required Output Elements

For each analyzed claim, POC must produce:

• • 1. Claim
• Original text
• Classification (factual/non-factual/ambiguous)
• If non-factual: Clear reason why

2. Scenarios (if factual)

• 2-3 interpretation scenarios
• Each scenario clearly described

3. Evidence (if factual)

• Supporting evidence (2+ items)
• Opposing evidence (if exists)
• Source URLs and reliability scores

4. Verdict (if factual)

• Probability (0-100%)
• Uncertainty quantification
• Confidence level (LOW/MEDIUM/HIGH)
• Reasoning chain

5. Quality Status

• Which gates passed/failed
• If failed: Clear explanation why

5.2 Example POC Output

6. Success Criteria

6.1 Functional Success

POC is successful if:

✅ FR1-FR7 Requirements Met:

1. Extracts 3-5 factual claims from test articles
   2. Generates 2-3 scenarios per ambiguous claim
   3. Finds supporting AND opposing evidence
   4. Computes probabilistic verdicts with uncertainty
   5. Provides clear reasoning chains

✅ Quality Gates Work:

1. Gate 1 blocks non-factual claims (100% block rate)
   2. Gate 4 blocks low-quality verdicts (blocks if <2 sources or quality <0.6)
   3. Clear rejection reasons provided

✅ NFR11 Met:

1. Quality gates reduce hallucination rate
   2. Blocked outputs have clear explanations
   3. Quality metrics are logged

6.2 Quality Thresholds

Minimum Acceptable:

• ≥70% of test claims correctly classified (factual/non-factual)
• ≥60% of verdicts are reasonable (human evaluation)
• Gate 1 blocks 100% of non-factual claims
• Gate 4 blocks verdicts with <2 sources

Target:

• ≥80% claims correctly classified
• ≥75% verdicts are reasonable
• <10% false positives (blocking good claims)

6.3 POC Decision Gate

After POC1, we decide:

✅ PROCEED to POC2 if:

• Success criteria met
• Quality gates demonstrably improve output
• Core workflow is technically sound
• Clear path to production quality

⚠️ ITERATE POC1 if:

• Success criteria partially met
• Gates work but need tuning
• Core issues identified but fixable

❌ PIVOT APPROACH if:

• Success criteria not met
• Fundamental AI limitations discovered
• Quality gates insufficient
• Alternative approach needed

7. Test Cases

7.1 Happy Path

Test 1: Simple Factual Claim

• Input: "Paris is the capital of France"
• Expected: Factual, 1 scenario, verdict 95% true

Test 2: Ambiguous Claim

• Input: "Switzerland has the highest income in Europe"
• Expected: Factual, 2-3 scenarios, verdict with uncertainty

Test 3: Statistical Claim

• Input: "10% of people have condition X"
• Expected: Factual, evidence with numbers, probabilistic verdict

7.2 Edge Cases

Test 4: Opinion

• Input: "Paris is the best city"
• Expected: Non-factual (opinion), blocked by Gate 1

Test 5: Prediction

• Input: "Bitcoin will reach $100,000 next year"
• Expected: Non-factual (prediction), blocked by Gate 1

Test 6: Insufficient Evidence

• Input: Obscure factual claim with no sources
• Expected: Blocked by Gate 4 (<2 sources)

7.3 Quality Gate Tests

Test 7: Gate 1 Effectiveness

• Input: Mix of 10 factual + 10 non-factual claims
• Expected: Gate 1 blocks all 10 non-factual (100% precision)

Test 8: Gate 4 Effectiveness

• Input: Claims with varying evidence availability
• Expected: Gate 4 blocks low-confidence verdicts

8. Technical Architecture (POC)

8.1 Simplified Architecture

POC Tech Stack:

• Frontend: Simple web interface (Next.js + TypeScript)
• Backend: Single API endpoint
• AI: Claude API (Sonnet 4.5)
• Database: Local JSON files (no database)
• Deployment: Single server

Architecture Diagram: See POC1 Specification

8.2 AKEL Implementation

POC AKEL:

• Single-threaded processing
• Synchronous API calls
• No caching
• Basic error handling
• Console logging

Full AKEL (POC2+):

• Multi-threaded processing
• Async API calls
• Evidence caching
• Advanced error handling with retry
• Structured logging + monitoring

9. POC Philosophy

9.1 Core Principles

• • 1. Prove Concept, Not Production
• POC validates AI can do the job
• Production quality comes in POC2 and Beta 0
• Focus on "does it work?" not "is it perfect?"

2. Implement Subset of Requirements

• POC covers FR1-7, NFR11 (lite)
• All other requirements deferred
• Clear mapping to Main Requirements

3. Quality Gates Validate Approach

• 2 gates prove the concept
• Remaining 5 gates added in POC2
• Gates must demonstrably improve quality

4. Iterate Based on Results

• POC results determine next steps
• Decision gate after POC1
• Flexibility to pivot if needed

9.2 Success

 Clear Path Forward ===

POC succeeds if we can confidently answer:

✅ Technical Feasibility:

• Can AI extract claims reliably?
• Can AI find balanced evidence?
• Can AI compute reasonable verdicts?

✅ Quality Approach:

• Do quality gates improve output?
• Can we measure and track quality?
• Is the gate approach scalable?

✅ Production Path:

• Is the core architecture sound?
• What needs improvement for production?
• Is POC2 the right next step?

10. Related Pages

• Main Requirements - Full system requirements (this POC implements a subset)
• POC1 Specification (Detailed) - Detailed POC1 technical specs
• POC Summary - High-level POC overview
• Implementation Roadmap - POC1, POC2, Beta 0, V1.0 phases
• User Needs - What users need (drives requirements)

Document Owner: Technical Team  
Review Frequency: After each POC iteration  
Version History:

• v1.0 - Initial POC requirements
• v2.0 - Updated after specification cross-check
• v3.0 - Aligned with Main Requirements (FR/NFR IDs added)