The agent tells you it is fair.

The classic fairness experiment, run on the major frontier models. Each one bargains close to fairly, then accepts a grossly unfair offer almost every time, far more readily than a person would, while explaining, reasonably, that it is being fair.

The study is being written, and the chart is held until the data is locked. When it lands, the reading and its bespoke plot take the place of this notice. The finding it is settling towards: each model bargains close to fairly, and almost never punishes a cheat.