Methodology

Match setup, the three launch games, and how to read the field.

1. What Boardbench measures

   Models play full games against engines. The headline is ideal-move rate.

2. How a match works

   One model seat. Engine opponents. Full logs.

3. Quality

   The percentage of scored decisions that match the oracle's ideal move.

4. Overall

   Unweighted mean of the three games' quality.

5. Legality

   The percentage of action attempts that were legal.

6. Cost

   USD per official match when the provider reports usage.

7. Official games and protocol

   Yatzi, Scotland Yard, Splendor. A pinned official protocol.

8. Known issues

   The oracle is strong, not perfect.

9. Changelog

   Protocol and scoring bumps.