ethskills-research

February 13, 2026 · View on GitHub

Research lab for ethskills.com — figuring out what LLMs don't know about Ethereum.

What's Here

/results/ — Baseline Audits

Stock LLM knowledge tests. We ask models Ethereum questions with no tools and see what they get wrong.

  • baseline-sonnet-4.5.md — Claude Sonnet 4.5 baseline
  • baseline-gpt-5.2.md — GPT-5.2 baseline
  • baseline-opus-4.6.md — Opus 4.6 baseline
  • gap-analysis-summary.md — Start here. Compiled gaps across all models.

/research/ — Deep Dives

Research on specific topics LLMs don't know:

  • erc-8004.md — Trustless agent identity standard (completely unknown to all models)
  • x402.md — HTTP 402 payment protocol (completely unknown)
  • protocol-upgrades.md — Pectra & Fusaka details
  • ai-ethereum-projects.md — AI + Ethereum landscape
  • new-eips-ercs.md — Recent standards

/content-mining/ — Source Material

Raw research used to build skill content:

  • contract-addresses.md — Verified addresses
  • defi-composability.md — Current DeFi state
  • ethereum-apps-ecosystem.md — App landscape
  • scaffold-eth-patterns.md — SE2 patterns
  • speedrun-ethereum.md — SRE content
  • And more

Methodology

  1. Test — Ask a stock LLM 20 Ethereum questions (no tools)
  2. Find gaps — What did it get wrong or not know?
  3. Fill — Write corrections with verified on-chain data
  4. Prune — If the LLM already knows it, cut it
  5. Repeat — Models improve, gaps change, content evolves

Key Findings (Feb 2026)

All tested models (Sonnet 4.5, GPT-5.2, Opus 4.6):

  • Think gas is 10-30 gwei (reality: 0.05-0.3 gwei)
  • Have never heard of ERC-8004 (trustless agent identity)
  • Have never heard of x402 (HTTP payments)
  • Think L2 transactions cost $0.01-2.00 (reality: <$0.001)
  • Unsure if Pectra/Fusaka shipped (both live)
  • Hallucinate contract addresses when asked

License

MIT