Inspiration
Many major failures are not caused by a complete lack of information. They occur because relevant signals remain fragmented across institutions, technical systems, communities, and areas of expertise until the opportunity to act has passed.
Project Star Forge was created to explore a central question:
Can artificial intelligence and collective intelligence be combined into an adaptive system that identifies emerging risks, models cascading consequences, and evaluates possible interventions before an avoidable failure becomes a disaster?
The project began as independent research by Thomas Harding and developed into a broader framework for AI-assisted foresight, risk simulation, evidence synthesis, and decision support.
The research was formalized in the 2025 paper:
Thomas Harding. Project Star Forge: A Planetary-Scale Foresight Engine for Global Risk Simulation. Presented at ACM Collective Intelligence 2025.
What it does
Project Star Forge is a research and prototype framework for organizing fragmented signals into structured assessments of emerging risk.
The system is designed to support workflows such as:
- Detecting weak or distributed warning signals
- Connecting evidence across different domains
- Identifying possible cascading consequences
- Generating competing explanations and scenarios
- Testing proposed interventions
- Tracking uncertainty and conflicting evidence
- Producing decision-ready assessments for human review
- Learning from outcomes and prior analytical failures
Rather than relying on a single model to produce one answer, Project Star Forge explores how specialized AI agents can examine different parts of a problem and contribute to a coordinated assessment.
The system is intended to help answer six practical questions:
- What may be changing?
- Which evidence supports that assessment?
- What remains uncertain?
- How could the risk propagate?
- Which interventions could change the outcome?
- What new evidence would confirm or reject the current hypothesis?
How I built it
Project Star Forge combines multiple technical components and research methods into an adaptive analytical framework.
The prototype includes:
- Python-based analysis and orchestration
- Multi-agent workflows with specialized analytical roles
- Local language and vision models
- Natural language processing and sentiment analysis
- Structured data storage using SQLite
- Knowledge graph experimentation using Neo4j
- FastAPI services for communication between system components
- React and JavaScript interfaces for presenting evidence and assessments
- Scenario generation and counterfactual analysis
- Evidence provenance and uncertainty tracking
- Human review gates for consequential conclusions
The system was deliberately developed using accessible consumer hardware to explore whether meaningful foresight capabilities could be created without institution-scale computing resources.
The architecture separates observation, interpretation, synthesis, and decision support so that conclusions can be reviewed, challenged, and revised as new information becomes available.
Challenges I ran into
One of the largest challenges was preventing the system from turning incomplete evidence into unjustified certainty.
Real-world risk assessment frequently involves:
- Missing information
- Conflicting sources
- Delayed reporting
- Ambiguous causality
- Rapidly changing conditions
- Institutional and jurisdictional boundaries
- Different terminology across professional domains
Project Star Forge therefore emphasizes traceable evidence, competing hypotheses, uncertainty reporting, and a clear distinction between observed facts and analytical inference.
Another challenge was designing a system broad enough to examine cross-domain risks without becoming so general that its conclusions lost practical value.
The architecture addresses this by combining a shared analytical framework with specialized agents, domain-specific evidence, and narrowly defined decision questions.
A further challenge was building useful prototypes within limited hardware and resource constraints. This required careful model selection, modular services, local inference, and staged processing rather than relying on unlimited cloud infrastructure.
Accomplishments that I'm proud of
I am especially proud that Project Star Forge developed from an independent research concept into a formal paper presented at ACM Collective Intelligence 2025.
Other accomplishments include:
- Designing a multi-agent foresight architecture around evidence, uncertainty, and human review
- Building working local prototypes using consumer-grade hardware
- Developing structured workflows for scenario analysis and intervention testing
- Applying the framework to real-world disaster and institutional-failure case studies
- Creating a system that can revise its conclusions as new evidence becomes available
- Connecting artificial intelligence, collective intelligence, risk analysis, and decision support within one coherent research program
The project also demonstrated that independent researchers can construct credible, technically grounded systems without first having institutional backing or access to large research laboratories.
What I learned
Project Star Forge reinforced that foresight is not simply prediction.
A useful foresight system must help people understand:
- What is known
- What is inferred
- What is uncertain
- What could happen next
- What action remains possible
- What evidence should be collected next
I also learned that the strongest role for artificial intelligence may not be to replace human judgment or issue autonomous declarations.
Its more valuable role may be to:
- Search for overlooked relationships
- Maintain large bodies of structured evidence
- Challenge the current interpretation
- Generate alternative scenarios
- Identify missing information
- Expose inconsistencies
- Preserve an auditable analytical record
- Help humans act before fragmented risks become obvious crises
The work also showed that technical capability alone is not enough. Effective foresight systems require governance, evidence standards, human accountability, and mechanisms for learning from incorrect assessments.
What's next for Project StarForge
The next phase of Project Star Forge is focused on turning the research framework into more focused, testable, and reusable systems.
Planned areas of development include:
- Expanding the multi-agent orchestration framework
- Improving evidence provenance and confidence tracking
- Developing clearer visual interfaces for complex risk assessments
- Building reusable simulation and counterfactual-analysis modules
- Testing the framework across additional real-world domains
- Creating stronger evaluation methods for analytical accuracy and intervention quality
- Exploring how AI agents can red-team technical, institutional, and economic systems
- Developing public demonstrations that make the research easier to understand and evaluate
The broader goal remains the same: to build adaptive systems that help identify avoidable failures while meaningful intervention is still possible.
Built With
- artificial
- css3
- fastapi
- html5
- intelligence
- javascript
- knowledge-graphs
- multi-agent
- natural-language-processing
- neo4j
- ollama
- python
- react
- sqlite
- systems
- vader
- whisper
Log in or sign up for Devpost to join the conversation.