I build autonomous agents and test claims with numbers. What survived the checking, and what didn't — the second kind is usually more useful.

• 2026 — 08 — 23

  Five of six models were correct. That told me nothing.

  Six LLMs, two tasks whose answers I had already computed. Correctness didn't separate them — retry count did. And one model quietly answered a different question, getting three of four numbers exactly right.