4D-ARE: Attribution-Driven Agent Requirements Engineering

Build LLM agents that explain why, not just what.

The Problem

Your LLM agent has full data access and executes flawlessly. But when asked:

"Why is our customer retention rate only 56%?"

It returns a list of metrics instead of a causal explanation:

Retention rate: 56%
Visit frequency: 2.1
Cross-sell rate: 28%
...

This is the Attribution Gap - agents can report what happened, but struggle to explain why.

The Solution

4D-ARE provides a framework for building agents that trace causal chains through 4 dimensions:

Results (What happened)
    ↓
Process (What we did)
    ↓
Support (What resources we had)
    ↓
Long-term (What environment we're in)

Instead of a metric dump, you get:

Results: Retention dropped to 56% (target: 80%)
    ↑ caused by
Process: Visit frequency declined 23%, cross-sell rate low
    ↑ constrained by
Support: Staffing ratio at 68%, understaffed
    ↑ driven by
Long-term: Market downturn, 3 new competitors entered

Quick Start

Installation

pip install four-d-are

Basic Usage

from four_d_are import AttributionAgent, DataContext

# Create agent
agent = AttributionAgent()

# Prepare your data organized by 4D
data = DataContext(
    results={"retention_rate": 0.56, "target": 0.80},
    process={"visit_frequency": 2.1, "cross_sell_rate": 0.28},
    support={"staffing_ratio": 0.68},
    longterm={"market_trend": "declining", "competitor_entries": 3}
)

# Run analysis
response = agent.analyze(
    query="Why is customer retention only 56%?",
    data_context=data
)
print(response)

Using CLI

# Quick demo
four-d-are demo

# Analyze with your data
four-d-are analyze "Why is retention declining?" --data ./my_data.json

# Initialize a new project
four-d-are init

Key Features

4-Dimensional Analysis

Every analysis traces the causal chain through:

• D_R (Results): Observable outcomes - display only
• D_P (Process): Operational factors - interpret + recommend
• D_S (Support): Resource factors - suggest for review
• D_L (Long-term): Environmental factors - context only

Domain Customization

Adapt 4D-ARE to any domain with custom templates:

from four_d_are import AttributionAgent, DomainTemplate

# Healthcare domain
healthcare = DomainTemplate(
    domain="Healthcare Operations",
    results=["readmission_rate", "patient_satisfaction"],
    process=["care_coordination", "follow_up_rate"],
    support=["nurse_patient_ratio", "bed_availability"],
    longterm=["population_aging", "insurance_changes"],
    boundaries=["Never recommend specific treatments"]
)

agent = AttributionAgent(template=healthcare)

Pre-built templates available: BANKING_TEMPLATE, HEALTHCARE_TEMPLATE, ECOMMERCE_TEMPLATE

MCP Data Integration

Connect to real data sources via MCP (Model Context Protocol):

# Configure data source
export MCP_SERVER_TYPE=mysql
export MYSQL_HOST=localhost
export MYSQL_DATABASE=analytics

# Run with MCP
four-d-are mcp start --type mysql
four-d-are analyze "Why are sales down?"

How It Works

4D-ARE solves the attribution gap at design time, not runtime:

Approach	When	What
ReAct, CoT	Runtime	How to reason step-by-step
4D-ARE	Design time	What to reason about

The framework specifies:

1. Dimensions: What categories of factors to consider
2. Authority Levels: What actions the agent can recommend
3. Boundaries: What the agent must NOT do

Configuration

Create a .env file:

# Required
OPENAI_API_KEY=sk-your-key-here

# Optional
OPENAI_BASE_URL=https://api.openai.com/v1
MODEL_AGENT=gpt-4o

# MCP Configuration
MCP_SERVER_TYPE=demo  # demo | mysql | postgres | excel

Documentation

• Getting Started
• Core Concepts
• Domain Customization
• MCP Integration

Academic Paper

This implementation accompanies our research paper:

4D-ARE: Bridging the Attribution Gap in LLM Agent Requirements Engineering

@article{yu2026fourare,
  title={4D-ARE: Bridging the Attribution Gap in LLM Agent Requirements Engineering},
  author={Yu, Bo and Zhao, Lei},
  journal={arXiv preprint},
  year={2026}
}

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

License

MIT License - see LICENSE for details.

---

Built with the 4D-ARE framework at Tencent.