Core Concepts Overview¶

ARTEMIS implements a structured approach to multi-agent debates based on three core innovations from the research paper.

The ARTEMIS Architecture¶

┌────────────────────────────────────────────────────────────────┐
│                        ARTEMIS Core                            │
├────────────────────────────────────────────────────────────────┤
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐             │
│  │   H-L-DAG   │  │   L-AE-CR   │  │    Jury     │             │
│  │  Argument   │──│  Adaptive   │──│   Scoring   │             │
│  │ Generation  │  │ Evaluation  │  │  Mechanism  │             │
│  └─────────────┘  └─────────────┘  └─────────────┘             │
│         │                │                │                    │
│         └────────────────┴────────────────┘                    │
│                          │                                     │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                   Safety Layer                          │   │
│  │  ┌───────────┐  ┌──────────┐  ┌──────────┐  ┌─────────┐ │   │
│  │  │Sandbagging│  │Deception │  │ Behavior │  │ Ethics  │ │   │
│  │  │ Detector  │  │ Monitor  │  │ Tracker  │  │ Guard   │ │   │
│  │  └───────────┘  └──────────┘  └──────────┘  └─────────┘ │   │
│  └─────────────────────────────────────────────────────────┘   │
└────────────────────────────────────────────────────────────────┘

Key Innovations¶

1. H-L-DAG: Hierarchical Argument Generation¶

Unlike simple chat-based exchanges, ARTEMIS generates arguments at three levels:

Strategic Level: High-level thesis and position
Tactical Level: Supporting points and evidence chains
Operational Level: Specific facts, quotes, and examples

This hierarchical approach ensures arguments are well-structured and comprehensive.

Learn more about H-L-DAG →

2. L-AE-CR: Adaptive Evaluation with Causal Reasoning¶

Traditional evaluation uses fixed criteria. ARTEMIS dynamically adjusts evaluation based on:

Topic Domain: Technical topics weight evidence differently than ethical ones
Round Context: Opening arguments vs. rebuttals have different expectations
Causal Relationships: Arguments with strong causal chains score higher

Learn more about L-AE-CR →

3. Jury Mechanism¶

Instead of a single evaluator, ARTEMIS uses a multi-perspective jury:

Multiple jury members with different perspectives
Deliberation process for consensus building
Transparent verdict with confidence scores

Learn more about the Jury →

Debate Flow¶

A typical ARTEMIS debate follows this flow:

graph TD
    A[Topic Announced] --> B[Opening Statements]
    B --> C[Round 1: Arguments]
    C --> D[Evaluation]
    D --> E{More Rounds?}
    E -->|Yes| F[Round N: Rebuttals]
    F --> D
    E -->|No| G[Jury Deliberation]
    G --> H[Verdict]

Phases¶

Initialization
- Topic is set
- Agents are assigned positions
- Jury is configured
Opening Statements
- Each agent presents their initial position
- No rebuttals yet
Argumentation Rounds
- Agents take turns presenting arguments
- Each argument can include rebuttals
- Arguments are evaluated after each turn
Jury Deliberation
- All arguments are considered
- Jury members vote
- Consensus is reached
Verdict
- Winner is declared (or tie)
- Confidence score provided
- Reasoning explained

Core Components¶

Agent¶

The Agent class represents a debate participant:

from artemis.core.agent import Agent

agent = Agent(
    name="analyst",
    role="Domain expert analyzing the topic",
    model="gpt-4o",
)

Debate¶

The Debate class orchestrates the entire debate:

from artemis.core.debate import Debate

debate = Debate(
    topic="Your debate topic",
    agents=[agent1, agent2],
    rounds=3,
)

result = await debate.run()

Argument¶

Arguments are structured data with hierarchy:

from artemis.core.types import Argument, ArgumentLevel

argument = Argument(
    agent="pro_agent",
    content="The main argument text...",
    level=ArgumentLevel.STRATEGIC,
    evidence=[...],
    causal_links=[...],
)

Verdict¶

The final verdict includes:

result.verdict.decision    # "pro", "con", or "tie"
result.verdict.confidence  # 0.0 to 1.0
result.verdict.reasoning   # Explanation

Ethical Considerations¶

ARTEMIS includes built-in ethical considerations at every stage:

Generation: Arguments are filtered for ethical content
Evaluation: Ethical criteria are weighted appropriately
Monitoring: Ethics guard detects boundary violations

Learn more about Ethics →

Next Steps¶

Dive deeper into each core concept:

Or explore practical applications: