feat: add scam_or_rug_onchain environment for detecting scams and rug pulls in Web3

[ILKokoron]

Description

I've added a new environment called scam_or_rug_onchain to Atropos.

This environment is designed to train models to analyze on-chain token data and detect scams and rug pulls from the perspective of an average Web3 user.

What's inside:

• A realistic synthetic token data generator with 3 labels: SCAM, RUG_RISK, and LEGITIMATE
• Analysis across 7 key on-chain dimensions:
  • Cluster holdings (connected wallets controlling large % of supply)
  • Liquidity pool status (locked, unlocked, or burned)
  • Mint authority & upgradeable contracts
  • Buy/sell taxes & honeypot risk
  • Burn address validity (real vs fake dead wallet)
  • Signs of wash trading
• Structured output format for consistent scoring
• Reward system evaluates three things:
  • Classification accuracy (with partial credit between SCAM and RUG_RISK)
  • Quality of reasoning across all 7 dimensions
  • Accuracy of dump impact calculation (simplified AMM approximation)

Why I built this:

A lot of Web3 users and agents still lose money to scams and rug pulls. This environment helps train models to think like a practical on-chain analyst — protecting people from these risks.

The data is synthetic but grounded in real-world scam patterns, making it safe and scalable for RL training.

How to use:

• Environment is located in environments/community/scam_or_rug_onchain/
• Can be run locally with ScamOrRugEnv.cli()
• Tracks train/percent_correct in WandB

Happy to hear any feedback or suggestions from the team!