Legal Multi-Agent Framework (OPP-115 Analysis)

An agentic framework built with LangGraph and gpt-oss:120b to automate the legal analysis of privacy policies using the OPP-115 dataset.

📊 Current Performance Baseline

The framework has been validated against the full OPP-115 Corpus (115 documents).

• Mean $F_1$-Score: 0.8234
• Total Duration: 47.93 minutes
• Peak Performance: Several policies achieved an $F_1$ of 0.94 (e.g., Policy #58, #1361).

🤖 Architecture

The system employs a collaborative multi-agent triad followed by a centralized auditor:

1. DPO Agent: Focuses on Technical Data Collection and Security.
2. Subject Agent: Analyzes User Rights and Access controls.
3. Authority Agent: Evaluates Regulatory Compliance and Policy Updates.
4. Critic Node: Performs cross-agent synthesis and hallucination filtering.

📁 Repository Structure

• /OPP-115: Dataset containing raw policies and expert annotations.
• /config: Prompt library (v2 optimized).
• /src: Core framework logic, evaluation metrics, and batch processing scripts.
• /results: Final JSON reports and performance visualization plots.

🚀 Execution

1. Install dependencies: pip install -r requirements.txt
2. Run full batch analysis: python src/batch_run.py
3. Generate plots: python src/evaluation/validator.py