Announcing EuclidBench: What We Learned Running 240 Computation Problems Through gpt-5.4-nano
We built a benchmark to test our own MCP tools. The most important fix wasn't the math. It was the error messages.
5 min readBlog
Technical insights on deterministic computation and AI agent reliability.