Build quiet tests for real intelligence.

We care about the moments where an agent almost understands the job, then slips. That is where a better dataset begins.

Each sample is a small, verifiable environment designed to reveal what the next model still needs to learn.