MUNvare Platform
Agentic MUN preparation system: RAG-based researcher builds position papers from live sources, an LLM judge scores speeches against delegate criteria, and a debate simulator runs adversarial argument rounds.
Skills involved
What This Is
A full agentic stack for Model United Nations preparation. Three core capabilities: deep research, speech drafting and critique, and adversarial debate simulation. The goal is to compress weeks of delegate prep into hours and to make the output better than what most delegates produce manually.
MUN is a high-signal testbed for AI argumentation: the task requires multi-source synthesis, positional consistency, rhetorical adaptation, and real-time counter-argumentation. All of these are frontier LLM capabilities.
The Research Agent
The researcher doesn't just retrieve documents it runs a multi-hop RAG pipeline. Given a topic and country assignment, it:
- Identifies the key sub-questions a delegate needs to answer
- Searches across UN documents, news archives, and academic sources
- Synthesises a position paper with citations, detecting contradictions between sources
- Flags areas of genuine uncertainty vs. well-established policy
The hard part: policy questions are inherently ambiguous. The system has to distinguish between a source being outdated, a source being contradicted by another, and a genuine policy grey area and handle each differently.
The Speech Evaluator
MUN speech quality is multi-dimensional: factual accuracy, rhetorical effectiveness, positional consistency, procedural correctness. We use an LLM judge with explicit rubrics per dimension similar to constitutional AI evaluation but adapted to parliamentary debate norms.
The evaluator outputs per-dimension scores with explanations, allowing delegates to target specific weaknesses rather than getting a single opaque grade.
The Debate Simulator
Given a resolution and a set of countries, the simulator runs multi-round adversarial debate: agents propose amendments, raise points of information, and attempt to shift the draft resolution. The interesting engineering is in the negotiation protocol how agents update their positions based on coalition pressure while maintaining core national interests.
What's Next
- Live web scraping for current events integration
- Speaker style adaptation (formal UN register vs. crisis committee tone)
- Simulation quality metrics: does practising against the bot improve real-conference performance?
Last updated Mar 12, 2026
← Back to all projects