FactTrace Hackathon · University of Cambridge · 31 Jan 2026

Team: Julia D, Iuliia V, Paulina C, Raj.

Award: Winning team.

We built a multi-agent fact-verification system where AI agents disagree, argue, and negotiate to determine whether an external claim is a faithful representation of a source fact or a mutation.

Instead of a single black-box verdict, our system exposes the reasoning process through an adversarial, courtroom-style debate.

🔗 Project visualization demo:
👉 https://v0.app/t/KEPLER_FACTTRACE_DEMO
(interactive UI showing agent debates, verdicts, and confidence)

Based on a dataset of claims and truthful affirmations (link), this system addresses the challenge: "Is an external claim a faithful representation of an internal fact, or is it a mutation?"

Our system employs 4 specialized agents in an adversarial debate architecture:

1. Prosecutor - Aggressively hunts for mutations, distortions, and misrepresentations
2. Defense - Argues for faithful interpretation and semantic equivalence
3. Epistemologist - Quantifies uncertainty and identifies ambiguous cases
4. Jury Foreman - Synthesizes arguments and delivers the final verdict

• Adversarial Design: Forces consideration of multiple perspectives
• Transparent Reasoning: Full debate transcripts show the decision-making process
• Uncertainty Quantification: Explicitly identifies ambiguous cases
• Multi-Round Debates: Agents can challenge and respond to each other's arguments
• Mutation Detection: Identifies 8 types of claim mutations (numerical distortion, missing context, causal confusion, etc.)

• Python: 3.11 or higher
• OpenAI API Key: Required for running the agents

# Create environment from environment.yml
conda env create -f environment.yml

# Activate environment
conda activate hackathon

# Set your OpenAI API key as an environment variable
export OPENAI_API_KEY='your-api-key-here'

Or create a .env file in the project root:

OPENAI_API_KEY=your-api-key-here

Navigate to the kepler directory:

cd kepler

Run with pre-selected strategic cases that showcase different mutation types:

python main.py

This will analyze 5 carefully selected cases demonstrating:

• Numerical boundary manipulation
• Added information
• Negation framing
• Borderline rounding
• Faithful representation

# Run specific cases by index
python main.py --cases 0,1,2

# Interactive case selection
python main.py --interactive

# Run all cases (expensive!)
python main.py --all

Compare the multi-agent system against a single-agent baseline:

cd kepler
python compare_systems.py

This will:

• Run both single-agent and multi-agent systems on the same cases
• Generate a detailed comparison report (comparison_report.md)
• Export results to JSON files for further analysis
• Show verdict agreements/disagreements and confidence differences

cambridge-dis-hackathon/
├── kepler/                          # Main source code
│   ├── agents.py                    # Multi-agent debate system
│   ├── main.py                      # Primary entry point
│   ├── compare_systems.py           # Single vs multi-agent comparison
│   ├── single_agent_baseline.py     # Simple baseline for comparison
│   ├── visualize.py                 # Visualization and export utilities
│   ├── demo.py                      # Demo script
│   ├── export_comparison_data.py    # Data export utilities
│   ├── export_debates.py            # Debate transcript export
│   ├── view_raw_responses.py        # View raw agent responses
│   ├── Kepler.csv                   # Dataset (claim-truth pairs)
│   └── requirements.txt             # Python dependencies
├── requirements.txt                 # Root dependencies
├── environment.yml                  # Conda environment specification
├── README.md                        # This file
├── Instructions.md                  # Hackathon instructions
├── LICENSE                          # License file
└── *.json                           # Output files (results, debates, etc.)

• FAITHFUL: The external claim accurately represents the internal fact
• MUTATED: The claim distorts, exaggerates, or misrepresents the fact
• AMBIGUOUS: Genuine uncertainty exists; reasonable interpretations differ

1. Numerical Distortion: Changed numbers or statistical boundaries
2. Missing Context: Omitted crucial contextual information
3. Causal Confusion: Misrepresented cause-effect relationships
4. Exaggeration: Amplified or dramatized claims
5. Scope Change: Altered the scope or generality of the claim
6. Temporal Mismatch: Changed time references or periods
7. Added Information: Introduced details not in the source
8. Negation Framing: Reframed using negation (e.g., "failed to" vs "did not")

FINAL VERDICT: AMBIGUOUS (80% confidence)

REASONING: The external claim closely approximates the original death toll 
figure with minor inequality inversion and omission of additional 
epidemiological data...

PROSECUTOR ARGUMENTS:
- Inverts inequality direction from lower bound to upper bound
- Removes broader epidemiological context

DEFENSE ARGUMENTS:
- Uses close numerical figure within narrow range
- Focusing on death toll is common journalistic practice

EPISTEMOLOGIST ANALYSIS:
- Core uncertainty: Whether inequality inversion constitutes meaningful 
  distortion or acceptable paraphrasing

Running the system generates several output files:

• debate_results.json - Full debate results with all agent responses
• multi_agent_results.json - Multi-agent system results
• single_agent_results.json - Single-agent baseline results
• visualization_data_*.json - Data for visualizations

1. Adversarial Testing: Prosecutor and Defense challenge each other
2. Bias Reduction: Multiple perspectives prevent single-viewpoint bias
3. Calibrated Confidence: Debate leads to more realistic confidence scores
4. Transparent Process: Full debate transcript enables human oversight
5. Nuanced Analysis: Multi-round exchanges capture subtle distinctions

# Export debate transcripts
python export_debates.py