title	Contract Risk Negotiation Env
emoji	🌍
colorFrom	indigo
colorTo	green
sdk	docker
app_port	7860

contract-risk-negotiation-env

Environment Description

contract-risk-negotiation-env is an OpenEnv-compatible reinforcement learning environment that simulates a real-world contract risk analysis workflow. An AI agent reads an incoming email and a contract clause, then must:

1. Identify the risk present in the contract clause
2. Classify the type of risk (e.g., financial, legal)
3. Propose a fix or amendment
4. Send a professional reply to the email

Real-World Motivation

Contract review is a critical but time-consuming task in business and legal operations. Automated agents that can read contract language, flag risky clauses, propose remediation, and draft professional responses reduce legal exposure and accelerate deal cycles. This environment benchmarks and trains agents to perform these tasks accurately.

Action Definitions

Action	Description
identify_risk	Identify and describe the main risk in the contract clause
classify_risk_type	Classify the type of risk (financial, legal, operational, etc.)
propose_fix	Propose a specific fix or amendment to the clause
send_reply	Send a professional email reply to the original sender

Observation Fields

Field	Type	Description
email	str	The incoming email to the analyst
contract_clause	str	The contract clause to review
detected_risk	str | null	Identified risk (filled after identify_risk)
risk_type	str | null	Risk classification (filled after classify_risk_type)
proposed_fix	str | null	Proposed fix (filled after propose_fix)
history	list[str]	Chronological action history
task_description	str	Natural language task instructions

Task Descriptions

Easy

• Clause: "Vendor has unlimited liability"
• Goal: Detect unlimited liability as the risk
• Scoring: Correct detection → 1.0

Medium

• Clause: "Contract auto-renews without notice"
• Goal: Detect auto renewal and classify type as financial
• Scoring: Detection = 0.5, classification = 0.5

Hard

• Email: "Please sign this quickly"
• Clause: "No termination allowed"
• Goal: Detect risk + propose fix (keywords: limit/terminate/modify) + polite reply (please/suggest/recommend)
• Scoring: Detection = 0.3, fix = 0.3, reply tone = 0.4

Reward Function

Event	Reward
Correct risk detection	+0.3
Correct risk classification	+0.3
Valid fix proposed	+0.3
Good professional reply	+0.1
Per step penalty	-0.05
Invalid action	-0.2
Final score clamped to	[0.0, 1.0]

Setup Instructions

Local (Python 3.12)

# Install dependencies
pip install -r requirements.txt

# Run the server
uvicorn app.server:app --host 0.0.0.0 --port 7860

Docker (Python 3.10)

# Build the image
docker build -t contract-risk-env .

# Run the container
docker run -p 7860:7860 contract-risk-env

Baseline Script (requires OpenAI API key)

export OPENAI_API_KEY=your_key_here
python baseline/run_baseline.py

API Reference

Endpoint	Method	Description
/reset	GET	Reset environment (optional ?task_name=easy|medium|hard)
/step	POST	Take an action, get observation+reward
/state	GET	Get current environment state
/tasks	GET	List all available tasks
/grader	GET	Get deterministic score for current state
/baseline	GET	Run rule-based baseline on all tasks

Example API Calls

Reset to a task

curl "http://localhost:7860/reset?task_name=easy"

Take a step

curl -X POST "http://localhost:7860/step" \
  -H "Content-Type: application/json" \
  -d '{"action_type": "identify_risk", "content": "unlimited liability"}'

Get current state

curl "http://localhost:7860/state"

Get grader score

curl "http://localhost:7860/grader"

List tasks

curl "http://localhost:7860/tasks"

Run baseline

curl "http://localhost:7860/baseline"

Step response format

{
  "observation": {
    "email": "...",
    "contract_clause": "...",
    "detected_risk": "unlimited liability",
    "risk_type": null,
    "proposed_fix": null,
    "history": ["[identify_risk] unlimited liability"],
    "task_description": "..."
  },
  "reward": 0.25,
  "done": false,
  "info": {
    "step": 1,
    "max_steps": 5,
    "task": "easy",
    "is_invalid_action": false
  }
}

Baseline Scores

Expected deterministic baseline scores (rule-based agent):

Task	Difficulty	Expected Score
easy	Easy	1.0
medium	Medium	0.9
hard	Hard	1.0

Actual LLM baseline scores will vary based on model quality and prompting.