POC1: Core Workflow with Quality Gates

Last modified by Robert Schaub on 2025/12/24 21:53

POC1: Core Workflow with Quality Gates

Phase Goal: Prove AKEL can produce credible, quality outputs without manual intervention

Success Metric: <10% hallucination rate, quality gates prevent low-confidence publications

1. Overview

POC1 validates that the core AKEL workflow (Article → Claims → Verdicts) can produce trustworthy fact-checking analyses automatically. This phase implements 2 critical quality gates to prevent low-quality outputs from being published.

Key Innovation: Quality validation BEFORE publication, not after

What We're Proving:

  • AKEL can reliably extract factual claims from articles
  • AKEL can generate credible verdicts with proper evidence
  • AKEL can assess article credibility beyond simple claim averaging (context-aware analysis)
  • Quality gates prevent hallucinations and low-confidence outputs
  • Fully automated approach is viable

2. Scope

In Scope

  • Core AKEL workflow (claim extraction, verdict generation)
  • Gate 1: Claim Validation (factual vs. opinion/prediction)
  • Gate 4: Verdict Confidence Assessment (minimum 2 sources, quality thresholds)
  • Basic UI to display results
  • Manual quality tracking

Out of Scope (Deferred to POC2+)

  • User accounts / authentication
  • Corrections system
  • Search engine optimization (ClaimReview schema)
  • Image verification
  • API endpoints
  • Archive.org integration
  • Security hardening
  • A/B testing
  • Gates 2 & 3 (Evidence relevance, Scenario coherence)

Experimental Features (POC1)

Context-Aware Analysis (Approach 1: Single-Pass Holistic)

Goal: Test if AI can detect when an article's overall credibility differs from the average of its claim verdicts (e.g., accurate facts but misleading conclusion).

Implementation:

  • Enhanced AI prompt to evaluate logical structure
  • AI identifies article's main argument
  • AI assesses if conclusion follows from evidence
  • Article verdict may differ from claim average

Testing:

  • 30-article test set (10 straightforward, 10 misleading, 10 complex)
  • Success criteria: ≥70% accuracy on misleading articles
  • Marked as experimental - doesn't block POC1 success

See: Article Verdict Problem for complete analysis

Decision: 

  • If ≥70% accuracy → ship in POC2
  • If 50-70% → try weighted aggregation approach
  • If <50% → defer to POC2 with different approach

3. Requirements

3.1 NFR11: Quality Assurance Framework (POC1 Lite Version)

Importance: CRITICAL - Core POC1 Requirement
Fulfills: AI safety, credibility, prevents embarrassing failures

Specification:

AKEL must validate outputs before displaying to users. POC1 implements a 2-gate subset of the full NFR11 framework.

Gate 1: Claim Validation

Purpose: Ensure extracted claims are factual assertions, not opinions or predictions

Validation Checks:

  1. Factual Statement Test: Can this be verified with evidence?
    2. Opinion Detection: Contains hedging language? ("I think", "probably", "best", "worst")
    3. Specificity Score: Contains concrete details? (names, numbers, dates, locations)
    4. Future Prediction Test: Makes claims about future events?

Pass Criteria:
- isFactual: true
- opinionScore: 0.3
- specificityScore: 0.3
- claimType: FACTUAL

Action if Failed:

  • Flag as "Non-verifiable: Opinion/Prediction/Ambiguous"
  • Do NOT generate scenarios or verdicts
  • Display explanation to user

Target: 0% opinion statements processed as facts

Gate 4: Verdict Confidence Assessment

Purpose: Only publish verdicts with sufficient evidence and confidence

Validation Checks:

  1. Evidence Count: Minimum 2 independent sources
    2. Source Quality: Average reliability ≥ 0.6 (on 0-1 scale)
    3. Evidence Agreement: % supporting vs. contradicting ≥ 0.6
    4. Uncertainty Factors: Count of hedging statements in reasoning

Confidence Tiers:
HIGH (80-100%):
-3 sources
-0.7 average quality
-80% agreement

MEDIUM (50-79%):
-2 sources
-0.6 average quality
-60% agreement

LOW (0-49%):
-2 sources BUT low quality/agreement

INSUFFICIENT:
- <2 sourcesDO NOT PUBLISH

POC1 Publication Rule:

  • Minimum MEDIUM confidence required
  • Blocked verdicts show "Insufficient Evidence" message

Target: 0% verdicts published with <2 sources

3.2 Modified FR7: Automated Verdicts (Enhanced)

Enhancement for POC1:

After AKEL generates a verdict, it must pass through the quality validation pipeline:

AKEL Workflow (POC1):

1. Extract claims from article
 ↓
2. [GATE 1] Validate each claim is fact-checkable
 ↓ (pass claims only)
3. Generate verdicts for each claim
 ↓
4. [GATE 4] Validate verdict has sufficient evidence
 ↓ (pass verdicts only)
5. Display to user

Failed claims/verdicts:
- Store in database with failure reason
- Display explanatory message to user
- Log for quality metrics tracking

Updated Verdict States:

  • PUBLISHED - Passed all gates
  • INSUFFICIENT_EVIDENCE - Failed Gate 4
  • NON_FACTUAL_CLAIM - Failed Gate 1
  • PROCESSING - In progress
  • ERROR - System failure

3.3 Modified FR4: Analysis Summary (Enhanced)

Enhancement for POC1:

Analysis Summary must now display quality metadata:

Analysis Summary:
 Total Claims Found: 5
 Verifiable Claims: 3
 Non-verifiable (Opinion): 1
 Non-verifiable (Prediction): 1

 Verdicts Generated: 3
 High Confidence: 1
 Medium Confidence: 2
 Insufficient Evidence: 0

 Evidence Sources: 12 total
 Average Source Quality: 0.73

 Quality Score: 8.5/10

4. Success Criteria

POC1 is considered SUCCESSFUL if:

✅ Functional:

  • Processes diverse test articles without crashes
  • Generates verdicts for all factual claims
  • Blocks all non-factual claims (0% pass through)
  • Blocks all insufficient-evidence verdicts (0% with <2 sources)

✅ Quality:

  • Hallucination rate <10% (manual verification)
  • 0 verdicts with <2 sources published
  • 0 opinion statements published as facts
  • Average quality score ≥7.0/10

✅ Performance:

  • Processing time reasonable for POC demonstration
  • Quality gates execute efficiently
  • UI displays results clearly

✅ Learnings:

  • Identified prompt engineering improvements
  • Documented AKEL strengths/weaknesses
  • Validated threshold values
  • Clear path to POC2 defined

5. Decision Gates

POC1 → POC2 Decision:

  • IF hallucination rate >10% → Pause, improve prompts before POC2
  • IF majority of claims non-processable → Rethink claim extraction approach
  • IF quality gates too strict (excessive blocking) → Adjust thresholds
  • IF quality gates too loose (hallucinations pass) → Tighten criteria

Only proceed to POC2 if all success criteria met

6. Architecture Notes

POC1 Simplified Architecture:

User Input AKEL Processing Quality Gates Display
 (claim extraction (Gates 1 & 4)
+ verdicts)

vs. Full System (Future):

Input Claim Extractor Scenario Generator Evidence Linker
Verdict Generator All 4 Gates Review Queue Publication

POC1 Acceptable Simplifications:

  • Single AKEL call (not multi-component pipeline)
  • No scenarios (implicit in verdicts)
  • Basic evidence linking
  • 2 gates instead of 4
  • No review queue

See: Architecture for details

Related Pages

Document Status: ✅ POC1 Specification Complete - Ready for Implementation
Version: V0.9.70