Leaderboards are a camera. Your workload is the room. Before you rewrite a vendor shortlist because a rank moved overnight, run a small blind sample on the tasks you actually pay for.

Loop from leaderboard jump through blind sample to keep-or-walk decision — Ten tasks beat one headline. Soft-lander eval literacy.

A boring sample that works

Pick 8–12 prompts from real tickets (redact secrets).
Hide model names from the scorer.
Score with a written rubric: correctness, citations, tool discipline, latency feel.
Repeat once after a prompt-template change so you are not measuring luck.

Counsel tone: if the board and your sample disagree, your sample wins for purchasing. Boards still help for orientation — they are not a purchase order.

Pair with the leaderboard-dimensions and buyer-checklist essays already on this lander. No architecture claims, no overclaim.

Sources: Public benchmark methodology caveats; NIST AI RMF measurement language; common pairwise/arena blind evaluation practice. Educational soft-lander essay.