OpenAI says 30% of SWE-Bench Pro is broken. We saw it first.

The same flaw is hiding in your own backlog, at a higher rate. Here’s why graded scoring catches what pass/fail testing misses.