Skip to Content

FAQ

Last modified by Robert Schaub on 2025/12/24 20:30

Frequently Asked Questions (FAQ)

Common questions about FactHarbor's design, functionality, and approach.

1. How do facts get input into the system?

FactHarbor uses a hybrid model combining three complementary approaches:

1.1 1. AI-Generated Content (Scalable)

What: System dynamically researches claims using AKEL (AI Knowledge Extraction Layer)

Process:

  • Extracts claims from submitted text
  • Generates structured sub-queries
  • Performs mandatory contradiction search (actively seeks counter-evidence, not just confirmations)
  • Runs automated quality gates
  • Publishes with clear "AI-Generated" labels

Publication: Mode 2 (public, AI-labeled) when quality gates pass

Purpose: Handles scale — emerging claims get immediate responses with transparent reasoning

1.2 2. Expert-Authored Content (Authoritative)

What: Domain experts directly author, edit, and validate content

Focus: High-risk domains (medical, legal, safety-critical)

Publication: Mode 3 ("Human-Reviewed" status) with expert attribution

Authority: Tier A content requires expert approval

Purpose: Provides authoritative grounding for critical domains where errors have serious consequences

1.3 3. Audit-Improved Quality (Continuous)

What: Sampling audits where experts review AI-generated content

Rates:

  • High-risk (Tier A): 30-50% sampling
  • Medium-risk (Tier B): 10-20% sampling
  • Low-risk (Tier C): 5-10% sampling

Impact: Expert feedback systematically improves AI research quality

Purpose: Ensures AI quality evolves based on expert validation patterns

1.4 Why All Three Matter

Complementary Strengths:

  • AI research: Scale and speed for emerging claims
  • Expert authoring: Authority and precision for critical domains
  • Audit feedback: Continuous quality improvement

Expert Time Optimization:

Experts can choose where to focus their time:

  • Author high-priority content directly
  • Validate and edit AI-generated outputs
  • Audit samples to improve system-wide AI performance

This focuses expert time where domain expertise matters most while leveraging AI for scale.

1.5 Current Status

POC v1: Demonstrates the AI research pipeline (fully automated with transparent reasoning and quality gates)

Full System: Will support all three pathways with integrated workflow

2. What prevents FactHarbor from becoming another echo chamber?

FactHarbor includes multiple safeguards against echo chambers and filter bubbles:

Mandatory Contradiction Search:

  • AI must actively search for counter-evidence, not just confirmations
  • System checks for echo chamber patterns in source clusters
  • Flags tribal or ideological source clustering
  • Requires diverse perspectives across political/ideological spectrum

Multiple Scenarios:

  • Claims are evaluated under different interpretations
  • Reveals how assumptions change conclusions
  • Makes disagreements understandable, not divisive

Transparent Reasoning:

  • All assumptions, definitions, and boundaries are explicit
  • Evidence chains are traceable
  • Uncertainty is quantified, not hidden

Audit System:

  • Human auditors check for bubble patterns
  • Feedback loop improves AI search diversity
  • Community can flag missing perspectives

Federation:

  • Multiple independent nodes with different perspectives
  • No single entity controls "the truth"
  • Cross-node contradiction detection

3. How does FactHarbor handle claims that are "true in one context but false in another"?

This is exactly what FactHarbor is designed for:

Scenarios capture contexts:

  • Each scenario defines specific boundaries, definitions, and assumptions
  • The same claim can have different verdicts in different scenarios
  • Example: "Coffee is healthy" depends on:
    • Definition of "healthy" (reduces disease risk? improves mood? affects specific conditions?)
    • Population (adults? pregnant women? people with heart conditions?)
    • Consumption level (1 cup/day? 5 cups/day?)
    • Time horizon (short-term? long-term?)

Truth Landscape:

  • Shows all scenarios and their verdicts side-by-side
  • Users see *why* interpretations differ
  • No forced consensus when legitimate disagreement exists

Explicit Assumptions:

  • Every scenario states its assumptions clearly
  • Users can compare how changing assumptions changes conclusions
  • Makes context-dependence visible, not hidden

4. What makes FactHarbor different from traditional fact-checking sites?

Traditional Fact-Checking:

  • Binary verdicts: True / Mostly True / False
  • Single interpretation chosen by fact-checker
  • Often hides legitimate contextual differences
  • Limited ability to show *why* people disagree

FactHarbor:

  • Multi-scenario: Shows multiple valid interpretations
  • Likelihood-based: Ranges with uncertainty, not binary labels
  • Transparent assumptions: Makes boundaries and definitions explicit
  • Version history: Shows how understanding evolves
  • Contradiction search: Actively seeks opposing evidence
  • Federated: No single authority controls truth

5. How do you prevent manipulation or coordinated misinformation campaigns?

Quality Gates:

  • Automated checks before AI-generated content publishes
  • Source quality verification
  • Mandatory contradiction search
  • Bubble detection for coordinated campaigns

Audit System:

  • Stratified sampling catches manipulation patterns
  • Expert auditors validate AI research quality
  • Failed audits trigger immediate review

Transparency:

  • All reasoning chains are visible
  • Evidence sources are traceable
  • AKEL involvement clearly labeled
  • Version history preserved

Moderation:

  • Moderators handle abuse, spam, coordinated manipulation
  • Content can be flagged by community
  • Audit trail maintained even if content hidden

Federation:

  • Multiple nodes with independent governance
  • No single point of control
  • Cross-node contradiction detection
  • Trust model prevents malicious node influence

6. What happens when new evidence contradicts an existing verdict?

FactHarbor is designed for evolving knowledge:

Automatic Re-evaluation:

  1. New evidence arrives
    2. System detects affected scenarios and verdicts
    3. AKEL proposes updated verdicts
    4. Reviewers/experts validate
    5. New verdict version published
    6. Old versions remain accessible

Version History:

  • Every verdict has complete history
  • Users can see "as of date X, what did we know?"
  • Timeline shows how understanding evolved

Transparent Updates:

  • Reason for re-evaluation documented
  • New evidence clearly linked
  • Changes explained, not hidden

User Notifications:

  • Users following claims are notified of updates
  • Can compare old vs new verdicts
  • Can see which evidence changed conclusions

7. Who can submit claims to FactHarbor?

Anyone - even without login:

Readers (no login required):

  • Browse and search all published content
  • Submit text for analysis
  • New claims added automatically unless duplicates exist
  • System deduplicates and normalizes

Contributors (logged in):

  • Everything Readers can do
  • Submit evidence sources
  • Suggest scenarios
  • Participate in discussions

Workflow:

  1. User submits text (as Reader or Contributor)
    2. AKEL extracts claims
    3. Checks for existing duplicates
    4. Normalizes claim text
    5. Assigns risk tier
    6. Generates scenarios (draft)
    7. Runs quality gates
    8. Publishes as AI-Generated (Mode 2) if passes

8. What are "risk tiers" and why do they matter?

Risk tiers determine review requirements and publication workflow:

Tier A (High Risk):

  • Domains: Medical, legal, elections, safety, security, major financial
  • Publication: AI can publish with warnings, expert review required for "Human-Reviewed" status
  • Audit rate: Recommendation 30-50%
  • Why: Potential for significant harm if wrong

Tier B (Medium Risk):

  • Domains: Complex policy, science causality, contested issues
  • Publication: AI can publish immediately with clear labeling
  • Audit rate: Recommendation 10-20%
  • Why: Nuanced but lower immediate harm risk

Tier C (Low Risk):

  • Domains: Definitions, established facts, historical data
  • Publication: AI publication default
  • Audit rate: Recommendation 5-10%
  • Why: Well-established, low controversy

Assignment:

  • AKEL suggests tier based on domain, keywords, impact
  • Moderators and Experts can override
  • Risk tiers reviewed based on audit outcomes

9. How does federation work and why is it important?

Federation Model:

  • Multiple independent FactHarbor nodes
  • Each node has own database, AKEL, governance
  • Nodes exchange claims, scenarios, evidence, verdicts
  • No central authority

Why Federation Matters:

  • Resilience: No single point of failure or censorship
  • Autonomy: Communities govern themselves
  • Scalability: Add nodes to handle more users
  • Specialization: Domain-focused nodes (health, energy, etc.)
  • Trust diversity: Multiple perspectives, not single truth source

How Nodes Exchange Data:

  1. Local node creates versions
    2. Builds signed bundle
    3. Pushes to trusted neighbor nodes
    4. Remote nodes validate signatures and lineage
    5. Accept or branch versions
    6. Local re-evaluation if needed

Trust Model:

  • Trusted nodes → auto-import
  • Neutral nodes → import with review
  • Untrusted nodes → manual only

10. Can experts disagree in FactHarbor?

Yes - and that's a feature, not a bug:

Multiple Scenarios:

  • Experts can create different scenarios with different assumptions
  • Each scenario gets its own verdict
  • Users see *why* experts disagree (different definitions, boundaries, evidence weighting)

Parallel Verdicts:

  • Same scenario, different expert interpretations
  • Both verdicts visible with expert attribution
  • No forced consensus

Transparency:

  • Expert reasoning documented
  • Assumptions stated explicitly
  • Evidence chains traceable
  • Users can evaluate competing expert opinions

Federation:

  • Different nodes can have different expert conclusions
  • Cross-node branching allowed
  • Users can see how conclusions vary across nodes

11. What prevents AI from hallucinating or making up facts?

Multiple Safeguards:

Quality Gate 4: Structural Integrity:

  • Fact-checking against sources
  • No hallucinations allowed
  • Logic chain must be valid and traceable
  • References must be accessible and verifiable

Evidence Requirements:

  • Primary sources required
  • Citations must be complete
  • Sources must be accessible
  • Reliability scored

Audit System:

  • Human auditors check AI-generated content
  • Hallucinations caught and fed back into training
  • Patterns of errors trigger system improvements

Transparency:

  • All reasoning chains visible
  • Sources linked
  • Users can verify claims against sources
  • AKEL outputs clearly labeled

Human Oversight:

  • Tier A requires expert review for "Human-Reviewed" status
  • Audit sampling catches errors
  • Community can flag issues

12. How does FactHarbor make money / is it sustainable?

[ToDo: Business model and sustainability to be defined]

Potential models under consideration:

  • Non-profit foundation with grants and donations
  • Institutional subscriptions (universities, research organizations, media)
  • API access for third-party integrations
  • Premium features for power users
  • Federated node hosting services

Core principle: Public benefit mission takes priority over profit.

13. Related Pages

20. Glossary / Key Terms

Phase 0 vs POC v1

These terms refer to the same stage of FactHarbor's development:

  • Phase 0 - Organisational perspective: Pre-alpha stage with founder-led governance
  • POC v1 - Technical perspective: Proof of Concept demonstrating AI-generated publication

Both describe the current development stage where the platform is being built and initially validated.

Beta 0

The next development stage after POC, featuring:

  • External testers
  • Basic federation experiments
  • Enhanced automation

Release 1.0

The first public release featuring:

  • Full federation support
  • 2000+ concurrent users
  • Production-grade infrastructure