Series

AI evaluations1 post

• A benchmark can measure its own metadata
  May 2026

Applied ML5 posts

• Who audits the auditor?
  Jun 2026
• A social-good protocol, built by an agent fleet
  Jun 2026
• When uncertainty makes active learning worse
  May 2026
• The model reads "not" — it just can't use it
  Apr 2026
• When it can't explain what it sees, it asks
  Apr 2026

Autonomous systems1 post

• Delete the science-fiction parts first
  Aug 2026

Bridge & DeFi security8 posts

• An arbitrage bot with no slippage is a sandwich
  Jun 2026
• Rebuilding a perps DEX from its docs
  Jun 2026
• The index fund that held the wrong asset
  Jun 2026
• Anatomy of a memecoin honeypot
  Jun 2026
• The bridge that paid twice
  Jun 2026
• How CoW Protocol settles a trade (and what my TWAP router got wrong)
  Jun 2026
• An invariant is not an end-to-end proof
  Jun 2026
• An AMM built to be attacked
  Apr 2026

Engineering practice3 posts

• Upstream is a different kind of test
  Aug 2026
• A research primitive should outlive the result that motivated it
  Aug 2026
• Correctness under hidden failure modes
  Aug 2026

On-chain randomness & verifiability7 posts

• Anatomy of a fake dice game
  Jun 2026
• Building ZK dark chess: real fog of war on a public chain
  Jun 2026
• What a ZK proof proves (and what it doesn't)
  Jun 2026
• Verifiable isn't trustless: a coin flip on Sui
  Jun 2026
• The other side of the wall: FHE where ZK stops
  Jun 2026
• The on-chain randomness landscape, or: how to pick a chess position fairly
  Jun 2026
• The post-quantum proof that Shor breaks anyway
  Jun 2026

Systems & infrastructure7 posts

• A new turbine inside a 1965 power plant
  Aug 2026
• Date and time bugs are systems bugs
  Aug 2026
• A default is executable policy
  Aug 2026
• Autograd is part of the API
  Aug 2026
• 197 passing tests, four real-hardware bugs
  Aug 2026
• Running an OP Stack L2 with reth
  Jun 2026
• Recursive types, finite values: an EIP-712 bug in alloy
  May 2026