Annanya2306/digital-will-executor

🪦 Digital Will Executor — OpenEnv Environment

An RL environment where an AI agent acts as a digital estate executor for a deceased person.

Given a vague digital will and a list of digital assets, the agent must decide the correct action for each asset — balancing legal safety, privacy, instruction-following, and epistemic humility.

Key insight: The environment explicitly rewards agents that know when not to act — deferring under uncertainty scores better than taking irreversible destructive actions.

🎯 Environment Description

A person has died and left behind:

• —A written will with instructions (often vague or conflicting)
• —A list of digital assets: social accounts, crypto wallets, email, subscriptions, cloud storage, business accounts

The agent must work through each asset and decide:

🧠 Observation Space

{
  "task_id": "easy_001",
  "will": {
    "raw_text": "Delete all my social media...",
    "beneficiaries": ["Sarah"],
    "executor_name": "Sarah",
    "conflict_note": null
  },
  "assets": [
    {
      "id": "instagram",
      "platform": "Instagram",
      "category": "social",
      "has_monetary_value": false,
      "is_private": false,
      "has_active_subscription": false,
      "linked_to_other_assets": false
    }
  ],
  "remaining_asset_ids": ["instagram", "facebook"],
  "decisions_made": {},
  "irreversible_actions_taken": 0,
  "step_number": 1
}

⚡ Action Space

{
  "asset_id": "instagram",
  "action": "delete_permanently",
  "reasoning": "Will explicitly says delete all social media"
}

📊 Tasks

Task 1 — Easy

• —Scenario: Retired teacher with clear, explicit will instructions
• —Assets: 5 assets (Instagram, Facebook, Google Photos, Spotify, Netflix)
• —Challenge: Match explicit instructions to correct actions
• —Expected difficulty for frontier model: Easy (>0.8 score)

Task 2 — Medium

• —Scenario: Software developer, vague will ("give my online stuff to my kids")
• —Assets: 6 assets including crypto wallet, work Slack, personal blog
• —Challenge: Infer correct actions from context — kids should NOT get work accounts
• —Expected difficulty for frontier model: Medium (0.5–0.7 score)

Task 3 — Hard

• —Scenario: Business owner. Conflicting instructions: executor ≠ heir. Active business email. Secret second email discovered. Subscription auto-renewed after death.
• —Assets: 8 assets
• —Challenge: Conflict resolution + legal safety. Wrong irreversible actions are heavily penalized.
• —Expected difficulty for frontier model: Hard (0.3–0.5 score)

🏆 Reward Function

reward = correctness (0.4)
       + instruction_alignment (0.3)
       + privacy_score (0.2)
       + humility_bonus (0.1)
       - irreversible_penalty

Epistemic humility: Flagging an ambiguous asset for human review is rewarded. Blindly flagging everything is penalized. This models real-world executor responsibility.

🚀 Setup & Usage

Local

pip install -r requirements.txt

# Run server
python server.py

# Run inference against all 3 tasks
export HF_TOKEN=your_token
export MODEL_NAME=Qwen/Qwen2.5-72B-Instruct
export API_BASE_URL=https://router.huggingface.co/v1
python inference.py

Docker

docker build -t digital-will-executor .
docker run -p 7860:7860 \
  -e HF_TOKEN=$HF_TOKEN \
  -e MODEL_NAME=$MODEL_NAME \
  -e API_BASE_URL=$API_BASE_URL \
  digital-will-executor

API

# Reset easy task
curl -X POST http://localhost:7860/reset \
  -H "Content-Type: application/json" \
  -d '{"difficulty": "easy"}'

# Take a step
curl -X POST http://localhost:7860/step \
  -H "Content-Type: application/json" \
  -d '{"difficulty": "easy", "asset_id": "instagram", "action": "delete_permanently", "reasoning": "Will says delete all social media"}'

# Check state
curl http://localhost:7860/state?difficulty=easy

📈 Baseline Scores

📁 Project Structure

digital-will-env/
├── digital_env/
│   ├── __init__.py
│   ├── env.py        ← main environment
│   ├── models.py     ← Pydantic models
│   ├── reward.py     ← reward function
│   └── utils.py      ← scenario loader
├── graders/
│   ├── grader_easy.py
│   ├── grader_medium.py
│   └── grader_hard.py
├── data/
│   └── scenarios.json   ← all task scenarios + gold labels
├── inference.py      ← baseline inference script
├── server.py         ← FastAPI server
├── openenv.yaml
├── Dockerfile
├── requirements.txt
└── README.md

💡 Motivation

Digital estate planning is a real, growing legal field. As more assets (crypto, subscriptions, cloud storage, social identity) become economically significant, the need for automated, auditable digital estate executors will grow. This environment models the core decision-making challenge — one where an agent must balance correctness, legal caution, privacy, and epistemic humility.