Deep-Dive

Pass@1 leaderboards say one thing, agentic end-to-end benchmarks say another, and the only reliable per-language agentic signal I found is a tiny indie repo measuring tokens-to-done.