Skip to main content

This Lab calls a model. Point it at your own endpoint (OpenCode Go / kimi-k3 works) — the key stays in your browser.

prompting game · pro

Pairwise Preference Arena

Run blind pairwise judging and watch judge bias appear in real time.

A or B — and would you bet your eval on it?

Engine 1.0.0 · game needs model key

Checking for a configured model key…

Method

  • Judges the same pair twice — A first, then B first — with a strict rubric judge, and reports whether the verdict flips with order.
  • Position bias is the lesson: production evals randomize order for exactly this reason.