You are the harness.
The window is full and the agent needs one more call. Pick how to make room, drag the budget, move the rule, and watch what the agent can still see. Bright is still in the window. Dim is gone. At the experiment's settings, you also see what real models did with the same window.
Without a model: what each policy keeps
Before any model is asked anything, the policy decides what stays. This runs the real policy code on seeded histories with the rule planted at turn 2, as in the experiments, and counts how often it is still in the window.