Changes

See figure 4, p310 of the paper for a graphic intuitiion.

As in cheap talk models, the number of partitions that can be sustained is a function of the bias of the reporting agent. If <math>x_c \ge 3\sigma_{\omega}^2\,</math> only one partition can be supported and the model reverts to the unspecialized case. Likewise, as the bias decreases the number of partitions increases and the floor is able to make more refined inferences about <math>\omega\,</math>, and there is less loss due to uncertainty.

Note that it would be efficient for the committee to specialize if <math>k \ge 2K^U\,</math> as well, as this would lead to gains to the floor, but the committee doesn't do so and is therefore 'underspecialized'.

===Closed rule, no specialization===